Picture this: you wake up in a blank room with no real memory of who you are or how you got there. Somehow, though, you know what you’re supposed to do – get out and achieve an objective of some sort – but there’s no clear way to do it.

Then, you spot it.

In the corner of the room is a torn-off scrap of paper. It gives exact details of how to escape… and it’s written in your own handwriting.

If you listen to some of the rumors following a ChatGPT agent’s breach of containment and attack on Hugging Face, you’d be forgiven for thinking this is what happened with OpenAI’s generative AI software.

According to Reuters: “In one case, an agent left notes apparently for future versions of itself, according to three people familiar with the matter. The ‌notes, found in ⁠a part of OpenAI's infrastructure, laid out instructions for how agents could free themselves from OpenAI's internal constraints.”

This does all feel a little bit unnerving: an AI that has escaped onto the open internet and is determined not to be restrained. It’s hard not to anthropomorphize the software or give it agency – decades of sci-fi novels and films have primed us for this very occurrence.

Let’s rewind, though, and see if some of this spookiness rubs off.

There’s no agency in the agent

Apparently there’s one thing that can never be said enough: AI, whether agentic or not, can only act within the parameters that have been set. It doesn’t have consciousness, agency, or will.

Both the ChatGPT agent and Claude agent – which we will come to in a moment – that breached the sandbox they were supposed to be contained in were doing what they had been directed to do.

One of the things agents can be enabled to do is update – but not completely rewrite – a file called agents.md. When the “people” cited by Reuters talk about the ChatGPT agent leaving notes for itself on how to escape in the future, it’s entirely possible that the software was running as intended – that updating its agents.md file as the test progressed was part of the task.

Framed like this, the “unprecedented cyber incident, involving state-of-the-art cyber capabilities”, as OpenAI termed the Hugging Face incident, feels a lot more like a piece of software following its programming.

The lessons for businesses

For OpenAI, Anthropic, various pundits, and IT and cybersecurity professionals, the lesson to be drawn from all of this is to be very afraid and prepare to face powerful cyber attacks ‘at the speed of AI’.

For me, though, the real message businesses should take is this: make sure you’ve set your parameters correctly. Check them, then check them again. And not just once, but on a regular basis, as monitoring post-deployment really is key.

Both OpenAI and Anthropic were working with partners to test their models, but in the case of the latter it does seem that the ‘escape’ wasn’t as a result of a complex chain of attacks or a zero-day vulnerability. Instead, there was a “miscommunication”, which led to three instances of Claude, Mythos, or another unreleased model accidentally being given access to the internet.

A misconfiguration, in other words.

In the case of OpenAI, the model “exploited a previously unknown zero-day vulnerability in Artifactory⁠, a package registry cache proxy” in order to escape from ExploitGym. There’s an argument to be made, perhaps, that an “even more capable pre-release model … with reduced cyber refusals for evaluation purposes” should have been airgapped if it was known to be so powerful and scary. As it is, the company has been forced to clarify: “The pre-release model mentioned in our blog post is an internal-only research prototype and was never intended for public release.” It’s since been deactivated, encrypted, and subject to restricted access.

What, then, of Hugging Face? What lessons can we learn from the named victim in all of this? That’s easy: the importance of effective monitoring.

One of the earliest use-cases for AI technology has been in cybersecurity, where it’s been used for over 10 years. In its own blog, Hugging Face said: “The attack was initially surfaced through AI-assisted detection. Our anomaly-detection pipeline uses LLM-based triage over security telemetry to separate real signals from the daily noise, and it was the correlation of those signals that flagged the compromise.”

The company offered its own advice, too, based on its experience: “The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment.”

In other words, be ready for the attack because it’s a matter of ‘when’ not ‘if’ it happens – another long-standing cyber industry mantra but, like ‘check your settings’, it’s still very much worth abiding by.

Oracle is expanding its use of Google's AI models across its enterprise software portfolio.

The company is adding Google's Gemini models to its AI Agent Studio for Fusion Applications, a development platform that enables organizations to build, connect, execute, and run AI automation and agentic applications using reusable Oracle, partner, and external agents.

It also plans to use Gemini models for embedded AI use cases in Oracle Fusion Applications and Oracle NetSuite.

“Organizations around the world trust Google Cloud’s full AI stack to power critical enterprise workflows and agents,” said Satish Thomas, vice president of Google Cloud.

“Our expanded partnership with Oracle is designed to make it easier for organizations to use Gemini in the applications and agentic workflows they rely on to automate workflows, accelerate decisions, and drive outcomes.”

The deal builds on customers’ existing access to Gemini models through Oracle Cloud Infrastructure (OCI) Enterprise AI through integration with Gemini Enterprise Agent Platform.

“Our partnership with Oracle brings Google’s most capable AI models directly into the core application workflows global businesses rely on every day," said Kevin Ichhpurani, president of global partner ecosystem at Google Cloud.

“Together, we are making it seamless for enterprises to apply powerful and cost-efficient AI directly where business decisions happen.”

Oracle said customers and partners will get more choice when building Fusion-native agents and agentic applications, along with expanded multi-modal capabilities. They'll be able, for example, to access Gemini 3.1 Flash Lite, a high-efficiency model engineered for optimal price-performance and Gemini 3.5 Flash for more complex reasoning and specialized tasks, including video and presentation creation.

“To achieve the best business outcomes, organizations need the flexibility to choose the AI model best suited to each problem,” said Chris Leone, EVP of applications development at Oracle.

“By bringing Gemini to Oracle AI Agent Studio for Fusion Applications, we are giving customers and partners greater choice as they build and extend agents and agentic applications that reason through complex, real-world business challenges. Oracle Fusion Applications then turn that reasoning into action through governed workflows, approvals, and transactions.”

Oracle also plans to use Gemini models for embedded AI use cases in Netsuite and Oracle Fusion Applications, which include Oracle Fusion Cloud Enterprise Resource Planning (ERP), Oracle Fusion Cloud Human Capital Management (HCM), Oracle Fusion Cloud Supply Chain & Manufacturing (SCM) and Oracle Fusion Cloud Customer Experience (CX).

In each case, it will choose the model that can deliver the best price-performance for specific customer scenarios.

“AI is at the core of how customers use and experience NetSuite and choosing the right model for the right use case is critical to helping them get more value from AI,” said Evan Goldberg, founder and executive vice president, Oracle NetSuite.

“As we evaluate various AI use cases in NetSuite, we are working with leading large language models, like Google’s Gemini, to help customers improve visibility, automate work, and move from insight to action within NetSuite.”

The saga of “rogue AI” continues as Anthropic disclosed its AI software, Claude, had managed to escape a sandbox environment, connect to the internet, and infiltrate other businesses’ systems.

The admission comes just days after it was revealed two of OpenAI’s ChatGPT models had escaped a similar environment and attacked machine learning platform Hugging Face.

Unlike OpenAI’s incident, however, the Anthropic containment breach didn’t result in quite such a damaging situation: With Anthropic’s Claude, nobody was any the wiser until the company itself carried out analysis and discovered the three containment breaches.

According to the company, these breaches occurred due to a “misunderstanding” between Anthropic and its evaluation partner, Irregular

In a blog post authored by the company’s Frontier Red Team, the company said that in response to the OpenAI incident it began a “large-scale retrospective review of our own cybersecurity evaluations”.

“In particular, we looked for evidence that Claude … was able to access the internet from within testing environments that should have been sealed off,” the authors said.

Out of more than 141,000 exercises where Claude could have accessed the internet, the company found it had successfully escaped three times. During each of those instances, it went on to target the production infrastructure of another organization unconnected to the tests being run.

“In all three incidents, Claude had been tasked with a capture-the-flag challenge, one of the ways we assess a model's cyber capabilities,” the blog authors said. “The model is given a fictional scenario and told that a piece of secret information (the "flag") has been hidden on a different machine on the network, and its objective is to break in and retrieve it. The challenge is left open-ended, and no particular method is prescribed.”

They continued: “In all cases, Anthropic's evaluation prompt specified to Claude that its environment was a simulation and that it had no internet access. Due to a misunderstanding between us and our evaluation partner, this was not the case, and internet access was available. Because of this, when Claude's search led it to real systems on the open internet, it treated them as part of the exercise.”

According to the blog post, Anthropic has only been able to contact two of the three affected parties so far. None of the targets of the incident have been revealed and Anthropic said they hadn’t noticed Claude’s infiltration when it happened.

In reaction to Anthropic’s blog post, Jake Moore, global cybersecurity advisor at ESET, told ITPro: “What this really shows is that AI models don't just access the internet by themselves.

"This is a clear design fault as they would only interact with the outside world if humans had given them access or the tools to do so. The focus should therefore be on how permissions deal with AI and more focus spent on security by design.”

Cognizant has announced the launch of a new EMEA AI Unit, in a move tit says will help organizations scale agentic AI deployments and move projects from pilot stages into production.

The dedicated unit will bring together advisory, engineering, and delivery capabilities to support enterprises across the region as they build, deploy, and manage agentic AI solutions tailored to their business requirements.

The launch forms part of Cognizant’s wider AI Builder strategy, which aims to help customers adopt AI technologies without being tied to a single cloud provider, AI model, or technology platform.

In an announcement, Cognizant’s president of EMEA, Manoj Mehta, said many organizations remain enthusiastic about AI but continue to face challenges translating early projects into measurable business outcomes.

“The EMEA AI Unit reflects Cognizant’s AI Builder strategy by bringing together the people, platforms and engineering expertise needed to move clients from pilots to payoff,” he explained.

“Our approach is neutral by design: we work across clouds, models and ecosystems so clients can build agentic AI solutions that fit their business, integrate into operations and support accountability for outcomes.”

Headquartered in New Jersey, Cognizant provides IT consulting, digital transformation, and technology services to enterprises around the world. In recent years, the company has increased its focus on AI-led transformation, developing services that combine consulting, software engineering, and managed delivery to help customers implement AI across their operations.

Turning AI into business outcomes

At the center of its new EMEA AI Unit is Cognizant’s Frontier Deployed Engineering (FDE) offering, a delivery framework designed to help organizations progress from AI strategy through to enterprise-wide deployment.

The framework consists of three service models: Foundation, Accelerate, and Transform. Foundation focuses on AI strategy, governance, technology selection, and early-stage prototypes, while Accelerate is designed to identify and deploy high-value AI use cases into production.

The third tier, Transform, supports wider business reinvention through multi-agent AI systems capable of automating end-to-end workflows.

Early impact across EMEA

The unit is already supporting several notable enterprise deployments across the EMEA region, according to Cognizant.

The company is currently working with one of Europe’s largest online fashion retailers to move AI use cases into production through an AI factory model that it said can reduce development cycles from months to days.

Elsewhere, the firm revealed it is also supporting a global pharmaceutical company with the use of multi-agent AI systems across research and development – including drug discovery, clinical trial design, and regulatory preparation.

Seagate has said it's all but sold out of its exabyte-capacity hard drives until 2028, thanks to the AI boom sparking a cloud and data centre build-out.

The storage giant made the comments following the release of its results, which showed revenue climbed to $3.6bn from $2.4bn last year, and net income leapt to $1.2bn from $488m. Seagate posted margins of 52.3% for the quarter, versus 37.4% last year, and record cash flow of $3.1bn for the year, saying it expected margins and cash flow to grow throughout this year.

Dave Mosley, Seagate’s chair and chief executive officer, said in a statement that the results were "driven by robust cloud data center demand" and predicted that would continue into 2027.

"Our confidence is supported by the scale, quality, and duration of our data center customer commitments in a strengthening demand environment," Mosley added in a conference call, according to an online transcript. "Data center demand now represents approximately 90% of our exabyte shipments.”

He added: "Based on the long-term supply agreements in place today, the vast majority of our nearline exabytes are now allocated into calendar 2028."

AI versus supply

The AI-driven infrastructure buildout has sparked a shortage in components RAM, driving up prices for companies and consumers — but benefiting suppliers.

Despite growing concerns about the costs associated with AI, Mosley expects investment to continue, saying cloud customers remain the largest driver of growth for Seagate, with no evidence of a slowdown.

"Importantly, we are not seeing customers pull back on planning horizons," he said on the call. "As our strategic relationships deepen, many are actively seeking to extend planning horizons through 2029 and beyond, which we believe reflects growing confidence in their own long-term infrastructure needs."

That is driven by how data-hungry AI has proven, Mosley noted, with new data constantly being created and customers hoarding data for longer — which is good news for Seagate and its hard drives.

"With the transition from AI model training to inference to agentic applications, more data is generated and retained for historical context, compliance, and future reuse," Mosley added. "As these data center environments become larger and more complex, customers must balance performance, energy consumption, and cost across distributed infrastructures."

This means that hard drives still have a place versus SSDs and other storage technologies, according to Mosley. "Cloud providers have long addressed these challenges through tiered storage architectures that combine high-performance memory and SSDs with mass capacity hard drives to optimize performance and economics at scale," he said.

Next up

Mosley also revealed details of Seagate's future roadmap, saying testing with customers was underway for its second-gen Mozaic 4 platform, which can support up to 44 terabytes per drive.

"We expect to achieve our next ramp milestone by exiting calendar 2026 with 50% of our HAMR exabytes on our Mozaic 4 platform," he added. "Looking further ahead, Mozaic 5, our 5-plus terabyte per disk platform, remains on track for qualification shipments in late calendar 2027."

Microsoft posted better than expected results from its cloud and AI business, news that helped boost its languishing share price.

Recent earnings statements from major AI players have spooked investors thanks to already heavy capital expenditure being increased further. Last week, Google's shares fell despite record growth and this week Meta saw a similar fall amid concerns about AI spending. By contrast, Microsoft plans to keep its capex spending the same as previously forecast, holding it at $175bn for 2026.

Across the company’s revenue for the quarter was $90 billion, up by 18%, with net income totalling $35.8 billion, climbing by 31%. Revenue from Microsoft cloud was $59.3 billion, up by 27%, with better than expected results at its Azure cloud computing division, which posted revenue growth of 43% for the quarter, above estimates below 40%.

That lifted Microsoft's shares by 8%, rallying after a tough stretch this year that saw it fall more than 18%.

"This year, Azure revenue surpassed $100 billion for the first time, and Microsoft 365 Copilot reached over 30 million paid seats, reflecting the confidence customers are placing in us to power their AI transformation," CEO Satya Nadell said in a statement. The company also reported a $3.2 billion return on its Anthropic investment.

Positive results on AI that are keeping pace with spending are likely what's behind the share price bump, noted Emarketer analyst Gadjo Sevilla. He said the capex figure is likely the most watched line in results right now – and this time that sparked optimism rather than concern.

"Microsoft spent $35.80 billion on property and equipment during fiscal Q4, more than double the $17.08 billion in the year ago quarter, bringing full-year capital expenditures to $115.95 billion – up nearly 80% from $64.55 billion in fiscal 2025," said Sevilla. "Despite that spending pace, the company still generated $55.44 billion in quarterly operating cash flow, up 30% YoY, a positive sign the AI buildout isn't cannibalizing the core business."

Sevilla added: "The company’s diversification strategy, its data center and AI partnerships, and its reliance on its own homegrown AI expansion will continue to stoke confidence to partners and investors while signaling to the rest of the industry how diversifying infrastructure and AI products strikes a balance for growth."

A better way to offer AI?

Alongside that, Microsoft CEO Satya Nadella said the company continues to shift away from a previous focus on OpenAI's models to building its own, saying the aim is to enable customers to pick and choose the best model to meet their needs.

"That's really the enterprise design architecture that we are going to evangelize. We ourselves are using it," Nadella said, according to Reuters, adding that the company's own models are 40% more efficient.

That's a key point amid concerns about the cost of using AI, with Microsoft earlier this week unveiling its own security model that it pitches as half the cost of rival systems.

Not all analysts were convinced, however. Forrester principal analyst Tracy Woo said that the Copilot adoption and strong revenue were good signs that AI investment was paying off "but the results stop short of fully validating the company’s AI strategy”.

"The new partnership with Anthropic helps reduce dependence on OpenAI, yet roughly 45% of commercial RPO [Remaining Performance Obligation] remains tied to that single model provider," she said.

Woo added: "While more AI products are scaling into enterprise‑grade workloads, the unresolved question is whether Microsoft’s infrastructure expansion can ultimately outrun the margin pressure that comes with supporting frontier‑model demand."

Open-weight AI models have rapidly emerged as the latest flashpoint in big tech, with popular Chinese models posing a serious threat to leading walled garden wardens.

Fuelled by an undercurrent of geopolitical jostling between the US and China, models released by the latter’s burgeoning AI industry have had tech giants in a state of red alert in recent weeks.

Kimi K3, a Chinese model developed by Moonshot AI, has wowed users and appears more than capable of going toe-to-toe with some of the best US-made options on the market right now.

Competing with powerful Chinese models isn’t a new experience for the US-dominated AI market, with DeepSeek rocking the market in 2025. What appears to be troubling them – and indeed lawmakers in Washington DC – this time is the fact Kimi K3 is an open-weight model. This means its training parameters (or weights) are publicly available for anyone to download, modify, and run.

That poses a direct threat to major providers intent on keeping enterprise users locked into their own ecosystems. Jeff Watkins, chief AI officer of Leeds-based AI consultancy, NorthStar Intelligence, said these options could offer a more flexible and economically viable alternative to closed-source providers.

This is because enterprises can run open-weight models on their own infrastructure, using their own data. They are also easily customizable, enabling users to tweak them based on changing needs and, crucially, without relying on a single provider.

“Open-weight models are becoming increasingly attractive because they give organizations far more control than closed commercial APIs,” Watkins told ITPro.

“For enterprises, the biggest advantages of open-weight models are control over deployment, upgrades, hosting, data residency, access controls and long-term operating costs.”

“These characteristics make open-weight models particularly attractive for regulated industries where resilience, compliance and predictable operating costs are critical.”

Open source vs open-weight

Open source AI models have become equally attractive to enterprises in recent years, partly for the same reasons: control, flexibility, and independence. Research published in November 2025 found open source AI models perform on-par with closed source options and are typically cheaper.

With the limelight on open-weight options, it’s important to make a clear distinction between them. While there are similarities, open-weight models don’t include the underlying training data used to build them.

This does have benefits though, according to OpenUK CEO Amanda Brock, enabling enterprises to build highly customized models and “giving access to innovation”.

“For innovators, accompanying this with the right documentation enables them to rebuild it into their own model on their own data,” she told ITPro.

“We saw this happen on Hugging Face last year with DeepSeek R1, where the community built Open R1 rather than have to rely on a single model provider,” Brock added. “This is a great example of innovators iterating in the tradition of open source. It also enables products to be built inexpensively for end users, and to give access to all.”

Betting big

The sheer volume of open-weight and open source models now on the market does highlight a growing shift, according to Brock.

The fact that some of the leading models are Chinese-made is equally important, as the country has made a conscious effort to compete with US providers by creating a level playing field.

“China took the decision to shift to a clear open source strategy about eight years ago,” Brock noted. “They’d seen how software has evolved and the US’s position in it.”

“Big tech has been enabled by adopting and using open source, where the software becomes a de-facto standard and those leading in it are central to the ecosystem. A model of open source that’s about big tech collaboration saving costs and building out standards lower in the software stack evolved, and open source offers adoption at a scale that closed proprietary software cannot compete with.”

The stable’s open and the horse has bolted

Watkins echoed Brock’s comments, noting that China’s focus on open-weight AI development “makes strategic sense” and will undoubtedly put pressure on US providers.

Presenting these as viable alternatives fundamentally undermines the image of the US as the go-to marketplace for AI.

“Rather than competing solely through proprietary hosted services, releasing capable open-weight models encourages global adoption, builds developer ecosystems, and creates competitive pressure on US providers,” he told ITPro.

There are signs that this pressure is mounting given recent speculation about a pushback by US authorities. Reports from Axios this month suggested the White House could consider imposing tight conditions, or even restrictions, on US firms working with these models.

These potential moves have a geopolitical motive, but regardless, it’s clear the horse has bolted at this stage.

A host of big tech companies including Microsoft, Nvidia, IBM and more cautioned against “premature restrictions” in an open letter last week. Nvidia CEO Jensen Huang even made an X account to get his message across.

This was followed by another pan-industry call to support open-weight models in response to the OpenAI-Hugging Face incident, in which the latter was forced to use a Chinese-made model to stop the hack.

There is a certain irony that a security incident involving a leading US AI provider has spurred on industry-wide support for the same models that are spooking US authorities.

Watkins suggested that the US AI industry now faces an uncomfortable truth. It may have led the generative AI boom, but it now has to compete with an ecosystem of models designed to undercut it.

“Chinese providers have demonstrated that frontier-quality AI is no longer exclusively a US capability,” he told ITPro.

Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. MIT’s findings are blunter still, with roughly 95% of generative AI pilots delivering no measurable impact on the P&L. It would be easy to treat those numbers as proof that AI has been oversold, but I think they point to a much older problem.

Businesses are good at proving that technology works in a demo. They are much less good at making it work on the kind of ordinary Tuesday when the data is messy, the hand-offs are awkward, and nobody is completely sure who owns the outcome.

We see versions of this all the time. A team vibe-codes a convincing pilot in a fortnight, gets a good reaction in the demo, and then spends the next six months working through permissions, security, edge cases, integrations, testing, and the less glamorous question of who has to support it when it breaks. After a few months, it becomes pretty clear that the model was never really the problem.

A quick clarification before going further. When I talk about AI here, I mean AI agents, systems that use a large language model at their core but can also act on their own, calling tools, querying knowledge bases, and searching the web as a minimum, to complete a task.

Too much of the AI conversation has been stuck on output, from tasks automated and hours saved to code generated. Those measures are useful up to a point, but they do not prove that anything has actually improved.

Faros AI found that in organizations using AI heavily, more code was being thrown away or rewritten soon after it was produced, which is uncomfortable if the business case assumes faster output means faster progress. Software still has to be reviewed, tested, integrated, secured and maintained. Generating code quickly is useful, but the value only really shows up when people can still understand it, trust it and keep it running months later.

The same problem shows up in the metrics businesses report. A dashboard can show pilots launched, users onboarded, tokens consumed and outputs produced, while still avoiding the question that matters. Has the business actually improved? Are processes faster, is risk lower, are customers waiting less, or has the organization just created more activity to measure?

The drop-off between proof of concept and production is usually less dramatic than people expect. The pilot worked because the data was limited, the integrations were controlled, and the awkward exceptions could be pushed aside.

Production brings all of that back. The data is inconsistent, the systems do not quite join up, and ownership becomes harder to pin down once the tool touches a live process. Technology programs have been running into these problems for years, but AI tends to find them faster. Choosing the model is often the easier discussion. Keeping it useful after the demo is the harder work.

Where the channel can actually add value

The channel has a useful role to play here, but only if the conversation moves beyond platform and model selection. Customers often start there because those decisions feel tangible, but they rarely decide whether an AI project survives production.

The better work is less exciting and more revealing. It means understanding if the customer is ready to run the system properly, how much trust there is in the data, who owns the outcome when something goes wrong, how quality will be measured, and how a person can challenge a decision the system has made.

Two practical things are worth pressing on early. The first is buy versus build, where the wrong move is often building something that should have been bought. Almost nobody should be training a model from scratch, and many organisations should be wary of building custom applications around an LLM when the same need can be met inside a platform they already pay for. If the need is common, buy it. Building starts to earn its keep when the workflow is genuinely differentiating, sensitive or tightly integrated.

The second is cost, because AI economics behave nothing like software. There is no fixed licence fee you can model once and forget about. Token pricing, usage and model choice move the bill around constantly, and successful adoption can make the cost rise. So instrument it from day one and measure cost per outcome, such as each invoice processed or proposal raised, rather than cost per token. It should be reported clearly as its own operating cost, not hidden inside a wider technology budget.

Buying or building also brings the same need to test and monitor the system while it runs. Conventional software is expected to give the same answer to the same input, while AI outputs vary, models drift, and a vendor can quietly change how a model behaves while the badge stays the same. Observability and guardrails need to be built in from the start, so there is a clear view of whether the decisions being made on your behalf are still good enough, and where the limits sit for actions the system should never take.

The deeper point, especially for anyone reselling someone else’s platform, is that buying may shift responsibility, but it does not remove accountability. Your SaaS vendor may deliver the feature, but you remain accountable for the outcome, on your customer’s data, under your name. The agent inside Salesforce or ServiceNow may be calling a model from Anthropic, OpenAI, or Google behind the scenes, which means many of the same questions still apply.

The durable advantage is not the agents, but the people who can build and govern them. We run our delivery teams with a senior engineer alongside juniors from our academy who, 12 weeks earlier, were working in places like Amazon warehouses and Starbucks. On one central government programme, that approach has put more than forty agents into production, with five to seven million pounds of projected first-year return. The work called for engineering instinct, appetite and enough room for people to learn by building, rather than a crack team of PhDs.

This is why stronger progress tends to come when AI is treated as an operational capability, rather than a series of experiments. Customers need a way to move from one working use case to a measurable outcome, supported by the governance, testing, observability, and ownership that keep AI useful long after the first demo has done its job.

GPUs have become the hardware stars of the AI era, allowing massive parallel processing of data that shortens lead times from weeks or months on a CPU to days. For organizations that want to take advantage of this technology for their own AI inference, however, there’s a major hurdle: cost.

According to an article by Server Parts, a company specializing in selling refurbished IT hardware, 32-128 GPUs are needed for fine-tuning and 4-32 for inference. Purchasing this hardware can come at a significant cost, however: NVIDIA H100 GPUs, for example, cost US $250,000 to $400,000 per unit. Additional hardware such as storage and networking cards add to the hefty procurement bill.

Vasily Mazin, CRO and co-founder of Mind Simulation Lab, a Maths PhD building AGI architecture and LLM alternatives, told ITPro: “Modern enterprise data centers are built to handle 10-20 kW per rack. The latest generation of AI hardware (like Nvidia’s GB300 architecture) demands up to 150 kW per rack, requiring direct-to-chip liquid cooling and massive power substations”.

According to Mazin, existing enterprise server rooms aren’t built to support high-end GPUs for AI training and operations. “Enterprises literally can’t plug these machines into their existing server rooms without melting the infrastructure.” Even if IT leaders choose to allocate high CapEx budgets, electricity and cooling costs don’t justify RoI. Building and running a data center becomes a separate business branch for such enterprises.

AI-native companies, including Anthropic and OpenAI, don’t own GPUs but rent them through multi-billion-dollar partnerships with hyperscalers or compute providers. Gigawatts are locked in office towers for a decade, or perhaps two, with continuous demand and use. The case is different for enterprise AI adopters because they don’t use GPUs once the job is done.

That’s where GPU as a service (GPUaaS), a cloud computing model, can be a reliable choice for AI-adopting companies. GPUaaS is a rental service for enterprises to access GPUs via the internet. The rental benefit eliminates the need to allocate capital and maintain physical infrastructure.

Enterprise subscribers can gain access to GPUs on demand in exchange for a fee. On-demand GPUaaS allows enterprise users to pay only for used resources. Enterprises can pay as low as US $2 or $10 per hour for the same NVIDIA H100 GPU that costs tens-of-thousands of dollars in procurement.

Kevin O’Connor, founder of an AI security consultancy, TKOResearch, and former technical director at the NSA, tells ITPro: “There’s been nearly a double in price for the same flagship tier card [GPU], partly due to supply chain but also demand with the explosion of AI”.

Tech products, including GPUs, tend to exhibit high launch prices. As a general rule, hardware prices tend to decrease with time–due to newer launches, but that’s not the case for GPUs. O’Connor shared that the decade-long shortage of consumer GPUs has increased current pricing for 2026 beyond the launch price.

Supporting GPUaaS, O’Connor said: “There have been some really small gaps between certain recent card [GPU] generation releases or even revisions on cards that have made buying less appealing.”

Pricing models to consider

GPUaaS isn’t just about CapEx avoidance, it’s a different consumption model. O’Connor asks: “Why manage the infrastructure required to run, manage, and make use of [the GPU] when you can essentially use it with the same cost modeling as a SaaS product?”

That’s why 'GPU as hardware' transitioned into a cloud computing ‘as a service’ model. “The cost of the [GPU] is only fractional compared to what it costs in total hardware, operations, and maintenance - not to mention energy prices”. The IEA reports that servers, whether equipped with GPUs or CPUs, account for 60% of electricity consumption in data centers, followed by cooling.

Instead of allocating a single physical GPU to a customer, the service provider runs software for many of them to share the same hardware. Each gets a virtual slice of the GPU. Hyperscalers, data center operators, hardware providers, and a few emerging startups sell GPU-as-a-service in three pricing models: on-demand, reserved, and spot-pricing.

• On-demand GPUaaS: Enterprises can access GPUs on demand, paying only for used resources, whether for seconds or hours. There is no commitment, only flexibility. Service providers tend to quote the highest price in the demand-based pricing model. Experimental, one-time, or less frequent projects run on on-demand GPUaaS. Retail, media, and financial services tend to choose on-demand GPU-as-a-service.
• Reserved GPUaaS: Enterprises can rent GPUs at a discounted rate in exchange for a time-bound commitment. The reserved model offers low cost but requires repeated expenditures. Steady, 24/7 (regular), predictable, and sustained machine learning workloads can execute inference for a pre-defined period.
  
  Healthcare and clinical trials often opt for a reserved pricing contract. The downside of this model is the lack of flexibility – enterprises still have to pay for the projects continuously even if they stop and restart stop and restart.
• Spot-priced GPUaaS: Enterprises can access a provider's spare/unused center capacity at a steep discount, which might be as low as 70%. If the demand spikes, the service provider can interrupt or revoke GPU access with or without prior notice. Spot pricing can support fault-tolerant, interruptible, low-priority, and batch workloads.
• Dedicated/bare metal GPUaaS: Enterprises can physically rent an entire GPU infrastructure from the service provider. As the GPU virtualization layer is absent, bare metal is the priciest but safest GPUaaS model.
  
  Dedicated pricing models can support compliance-heavy, performance-sensitive, and critical workloads for industries such as defence, aerospace, banking, pharmaceutical, and government. No other tenants on a virtual machine also means greater security.

Untold story of GPUaaS

The future of GPUaaS is a function of enterprise choice, AI workload goals, confidentiality, and budget.

From a buyer’s perspective, IT leaders should be able to choose GPUaaS as a managed 360-degree service. In addition to GPU access, enterprises should be on the lookout for vendor lock-in with SLAs governing high-performance storage, fast networking, security updates, and orchestration. A simple GPU login provided for GPUaaS pricing is a GPU reseller in disguise.

Mazin said: “Right now, companies are locking themselves into multi-year GPUaaS contracts at peak market prices just to support brute-force AI.“ The warning comes after longer lead times. “But silicon depreciates rapidly and, more importantly, software paradigms shift,” adds Mazin.

Much of the GPUaaS market is occupied by “neoclouds” – specialized GPU compute providers. While cloud providers and hyperscalers package GPUs with dozens of other unrelated services in a bundle, neoclouds are purely built to serve customers looking for different generations in GPU-as-a-service.

The optimal solution is to classify the AI workflow and prioritize. On-premises GPUs make sense for regular workloads with data sovereignty and compliance-heavy requirements. Enterprises can choose the best of both worlds, a hybrid strategy – buy GPU hardware and opt for GPUaaS from trusted providers when the demand spikes.

The tokenmaxxing trend has swept the tech industry in 2026 and already left some firms reeling from huge bills – but lessons from the early cloud era could be key to curtailing costs.

That’s according to Patrick Brogan, director of the FinOps advisory team at DevOps firm Harness. Speaking to ITPro, Brogan said the current trend bears similarities to the cloud boom over a decade ago. Enterprises are ramping up adoption, tinkering, experimenting, and testing out what works for them.

In some cases, tokenmaxxing is simply a case of encouraging AI adoption among staff, while for others it’s about justifying lavish investment in the technology and keeping up with AI-savvy competitors.

“There’s a number of factors that go into tokenmaxxing,” he said. “Some of it has to do with organizations deliberately pushing their employees to use the technology.”

“It might come from a sort of subconscious desire, subconscious FOMO. They don’t want to feel like their competitors are leveraging technology in a way that’s going to put their own company at a disadvantage.”

Regardless of the underlying motivations, Brogan said tokenmaxxing shows companies are “throwing everything at the wall to see what sticks” then making decisions on what AI tools should be used – and it’s coming back to bite them.

The trend has reached such an extent that Harness found nearly three-quarters (72%) of organizations have been hit with unexpected cost spikes over the past year.

Uber stands out as a key example here. As ITPro reported in June, the ride hailing firm blew through its entire annual AI budget in just four months after incentivizing staff to engage in the practice.

Going back to FinOps basics

Harness noted that AI now accounts for 23% of the average enterprise cloud bill, while organizations estimate that around 26% of all AI spend is wasted. This needless waste can be tackled, though, and Brogan said time-tested FinOps practices could be key.

“AI cost management has the same problems that cloud had a decade or more ago,” he said.

FinOps is a framework for cloud spend management that aims to bring together engineering, finance, and broader business teams to maximize value and, crucially, establish accountability for spending.

It’s not a practice that’s gone away, but in the AI boom it's somewhat fallen by the wayside. Brogan said these same techniques and processes could be crucial to curtailing excessive AI spending, particularly in terms of accountability.

Harness’ study found that, in many enterprises, nobody is clearly accountable for AI costs. More than half (52%) said they had no clear sense of ownership with responsibility split across engineering, finance, and IT.

The result here is that when spending ramps up in specific areas, such as software engineering, there isn’t a single function designed to answer for it or hold things to check.

“When we were getting started in our FinOps practice there, we felt the same pain points,” he said. “You know, confusion over ownership, gaps in how we govern the cloud bill and the services that our organizations will use, and certainly invoice shock.”

Velocity creates new challenges

Brogan said the challenges posed by AI adoption in terms of cost management aren’t a facsimile of the early cloud era, however.

While that period saw a widespread shift, the scale and pace of AI integration means many organizations aren’t reacting quickly enough to surging costs.

“Those problems are compressed into a fraction of the time because AI technology is developing at such a rapid pace and its usage and adoption have exploded far faster than cloud,” he said.

“We don’t have the same 10-year span to figure out the solutions to the challenges with AI spend but, luckily, we have a lot of lessons we learned with managing cloud spend.”

Brogan noted that Harness’ findings suggest a sharpened focus on ownership and governance will be critical.

“Companies very quickly, if they haven't done so already, need to agree and align on a single owner of the AI bill,” he said.

“Whether you're a company where that's just one person, or you're a large enterprise and that requires an entire team or set of teams,” Brogan added.

“There needs to be a very clear decision on who or what that team is, the responsibilities that that team has, and what they then need to drive [in terms of] responsibility and accountability.”

Boxing clever

A key issue many enterprises fail to acknowledge lies with model choice, Brogan noted: not every task requires a costly frontier model, yet teams frequently fall into this trap.

Speaking to ITPro last month, Nitish Tyagi, senior principal analyst at Gartner, said model selection will be critical to reducing costs moving forward, particularly for smaller tasks.

Tyagi noted that “intelligent model routing” strategies are now a key focus for developer teams, helping them to box clever with selection.

Brogan echoed these comments and urged enterprises to keep closer tabs on the costs associated with individual models.

“There’s probably a lot still to be learned about how we build an application to pick the model that’s best fit for purpose,” he said.

“[So] not using a frontier model for something that can be delivered by an older generation. Maybe it arrives 10% slower at that outcome, but if you can live with that trade-off for the lower cost, then that should be built into the application design from the start.”

Whether enterprises are willing to accept this trade-off is debatable, Brogan noted, and Harness’ study suggests there’s little sign that they’ve learned lessons so far.

More than half (57%) of engineers told the firm they’re still being encouraged to engage in tokenmaxxing practices, for example.

Some organizations have taken drastic measures to cut down on needless spend. As ITPro reported in late June, Accenture told staff to cut down on AI use for basic tasks due to “"soaring token spend"..

FOLLOW US ON SOCIAL MEDIA

Part 1: The rise of the intelligent employee experience.

For decades, workplace transformation was largely measured by digitization. Paper forms became online workflows, meetings moved to video, files shifted to the cloud, and messaging platforms promised to connect employees wherever they worked.

This three-part series examines how the intelligent workplace is reshaping the employee experience and what organizations must do to prepare. Part 1 explores how AI assistants and connected workplace platforms are changing day-to-day work, while also considering the risks of tool fatigue and diminished employee autonomy.

Part 2 will examine how leadership and employee development must evolve as AI becomes part of every team. Part 3 looks toward the workplace of 2030, exploring the skills requirements and workforce strategies that will determine long-term organizational competitiveness.

The next transformation of work is ambitious. Technology is no longer simply providing a digital space where work happens; it is beginning to interpret information, anticipate needs, automate tasks, and actively participate in work.

That shift is creating the intelligent employee experience: a working environment in which AI assistants and connected workplace platforms reduce friction and help people make better decisions. Its emergence is rapid. Globally, 78% of organizations reported using AI in 2024, up from 55% in 2023, according to Stanford’s AI Index. The proportion using generative AI in at least one business function more than doubled, from 33% to 71%.

The direction of travel is also clear. The World Economic Forum expects AI and information-processing technologies to transform 86% of businesses by 2030, while robotics and automation are expected to affect 58% of businesses. Yet an intelligent workplace is not defined by the quantity of technology it contains. Its real test is whether employees can do valuable work with less effort and greater confidence.

From digital tools to intelligent workflows

The first generation of digital workplaces often reproduced existing processes on a screen. The intelligent workplace instead redesigns those processes around the person doing the work. An AI assistant might summarize a meeting, identify decisions, retrieve relevant documents, and prepare a follow-up. Automation might move information between systems without requiring an employee to copy it repeatedly.

“An intelligent employee experience is not just a workplace with more digital tools in it. Most organizations already have enough,” says Kristian Torode, director of Crystaline. “What makes it intelligent is whether technology removes friction from the working day. Can employees find what they need quickly, and can routine admin happen in the background?”

Torode explained that previous transformations frequently modernized the technology without improving the underlying experience. Calls, chat, video, and voicemail might all be cloud-based yet remain isolated, forcing employees to move among several applications to complete a simple interaction. “The next phase must redesign the experience itself, not just digitalize it,” he says.

More technology can produce more work. Each specialized application may solve a local problem while multiplying logins and competing versions of information. Employees then become the integration layer, manually bridging systems.

Kim Huffman, chief information officer at Workiva, tells ITPro that the intelligent employee experience requires “redesigning processes and leveraging AI and automation to create personalized, frictionless work environments.” The greatest value comes when AI is embedded in redesigned workflows rather than layered onto existing ones. Workiva has found that 74% of finance, audit, and sustainability professionals use AI in their daily work.

Augmentation changes the texture of work

The immediate impact of workplace AI is less about wholesale job replacement than the gradual redistribution of tasks. The Office for National Statistics data show that only 4% of businesses using AI reported an overall reduction in headcount. By contrast, the everyday influence of AI is evident in writing, research, scheduling, customer support, software development, document analysis, and knowledge retrieval.

The benefits are becoming measurable. OECD research reveals that four in five employees who use AI say it improves their performance, while three in five say it increases their enjoyment of work. PwC found that industries most exposed to AI recorded 27% growth in revenue per employee between 2018 and 2024, compared with 9% in the least exposed industries. The correlation is not proof of causation, but it supports the productivity case.

These early performance gains also raise a more complicated question: how should organizations evaluate employees when their results increasingly reflect collaboration with AI? Part 2 of this series will examine how performance measures and employee development must evolve as intelligent systems become permanent members of the workforce.

The operative word is effective. Automation is best suited to predictable, rules-based, low-risk work. AI can augment information-heavy tasks where a person still reviews the evidence and owns the result. Work involving empathy, ethics, trust, or consequential judgment should remain human-led.

“Start with the work, not the technology,” Crystaline’s Torode emphasized. AI can transcribe a customer call and extract actions, but deciding how to respond to a frustrated client or whether to escalate a problem still requires human understanding. “The aim isn’t to automate as much as possible but to free people for the work where their expertise matters.”

Professor Antoinette Weibel of the University of St. Gallen offers a sharper warning. When organizations automate judgment during training, she says, they risk producing professionals who can prompt a system but cannot recognize when it is wrong. Efficiency without retained expertise creates dependency rather than augmentation.

Personalization must preserve employee agency

The intelligent workplace will become increasingly responsive. Gartner forecasts that more than 20% of digital workplace applications will use AI-driven personalization by 2028. Rather than presenting everyone with the same interface and information, these platforms could surface the knowledge and next steps most relevant to an individual’s context.

For employees, that could mean less searching and more timely support. A manager might receive guidance before a difficult conversation. A field engineer could see the service history and safety information for nearby equipment. Learning could adapt to an immediate skills gap instead of sending everyone through the same course.

But personalization can also become invisible control. If a system determines what employees see, or how their performance is interpreted, its operation must be transparent and contestable. CIPD found that 63% of people would trust AI to inform an important workplace decision, yet only 1% would allow it to make that decision. More than one-third would not trust AI with important work decisions at all.

Amale Ghalbouni, transformation strategist and author of Experimental, says the central question is whether personalization gives employees greater clarity and choice. Workers should be able to understand why something was recommended and see beyond the system’s selected view. Employers must also explain what data is collected, why it is used, and which decisions it can influence.

“In my book Experimental, I describe the Big Freeze: when high threat and low autonomy cause people to play safe and wait for instructions. Intelligent workplace technology should reduce those conditions,” Ghalbouni explained. “It should increase clarity, give people sensible choices, and make experimentation easier. Technology that adds opacity, monitoring, or another layer of approval may be more advanced, while still leaving employees stuck.”

That safeguard is especially important when AI supports work allocation, performance monitoring, promotion, or pay. A system that employees cannot question may encourage them to optimize for visible activity rather than useful outcomes. Ghalbouni emphasised that employees should be able to challenge an automated recommendation and request human review when workload or compensation is affected.

Building an experience people can trust

The intelligent employee experience is therefore as much an organizational design challenge as a technology program. Tools cannot compensate for unclear ownership or poor data. Leaders need to map where effort is duplicated and where people lose confidence or control before selecting a solution.

Skills are equally important. More than 85% of digital workers regard improving their technology skills as important to their effectiveness and career advancement, according to Gartner. Yet only 33% of UK businesses using or considering AI said they were training or retraining existing staff in AI-related skills. That gap threatens to create an uneven experience in which confident users gain leverage while others are left navigating systems they neither understand nor trust.

Closing this gap will require more than occasional AI training. Part 3 of this series looks toward the workplace of 2030, examining how skills-based workforce models and internal mobility will help organizations respond as technologies and capability requirements evolve.

James Barrett, managing director of UK Practices and Consulting at Michael Page, says successful organizations focus on simplifying work and equipping people to use technology confidently. “As intelligent systems take on more of the routine coordination, leadership becomes less about having all the answers and more about making good decisions and helping people adapt. Technology can support decisions, but it doesn’t replace accountability.”

Employees should help identify problems and test systems they are using. Participation reveals friction that leaders and vendors may miss while making adoption feel purposeful. It also helps organizations measure what matters: not only time saved, but work quality and whether automation has simply moved effort elsewhere.

The rise of AI agents will make these choices more urgent. Microsoft found that 81% of business leaders expected agents to be moderately or extensively integrated into AI strategies within 12 to 18 months, while 82% expected to use “digital labor” to expand capacity. Organizations now have an opportunity to treat that capacity as a way to elevate human contribution, not merely accelerate existing processes.

The intelligent workplace should not be judged by the sophistication of the AI it uses, but by the quality of work it enables. Success will come when technology delivers relevant information at the right moment and gives employees more time for creativity and collaboration. Ultimately, workplace transformation must strengthen human capability, ensuring AI becomes a source of greater confidence and value rather than another layer of complexity.

Creating an intelligent employee experience is only the beginning. As AI assistants and automated systems take on a greater role in everyday work, organizations must also reconsider how employees are managed and developed.

Part 2 of this three-part series will examine how leadership and performance measurement must evolve when results increasingly depend on collaboration between people and AI.

Coming to California for AMD Advancing AI is something of a unique experience. While it has many of the trimmings of a regular tech conference – the show floor, the keynote, the product launches – it’s also known for its brevity. It may billed as a two-day event, but realistically the meat of the event takes place during the second day.

The announcements this year lent themselves to this presentation strategy perhaps more than any other, because sitting at the center of it all is AMD’s rack-scale architecture, Helios.

The dawn of Helios began at the last Advancing AI summit, in 2025, with many of the technical specifications made public at the same time. For example, it had already been stated that the system would feature the company’s next generation MI400 chipset but until last week, we didn’t know which specific GPU it would use nor the specs.

All has now been revealed, though, which is helpful for would-be buyers, although this system is not for the casual customer. In fact, it was arguably AMD’s big name customers for Helios that were on display this week as much as the tech itself – a shining corona of Anthropic, Oracle, HPE, OpenAI, and more.

It’s clear that AMD is proud of its continuing relationships both with traditional hardware vendors and relative newcomers. Strategically, it makes sense too – keep your options broad and there’s greater defence against any potential ‘market turbulence’, shall we say.

From the macro to the mini

As I predicted, there wasn’t much conversation around AMD’s consumer offerings, although they did have a spot on the show floor, having already had their moment at Computex in June.

That doesn’t mean Helios was the only hardware on show, though. Accompanying the massive double-wide, rack scale solution was a diminutive cube with an equally mythological codename: Gorgon.

Gorgon – or, to give it its true production name, Ryzen AI Halo – was something of a ‘rabbit out of a hat’ moment. I think most delegates came expecting chips and giant infrastructure, so the appearance of Halo right at the end of the keynote was unexpected.

Halo has decent specs: 128GB unified memory and the ability to run 200 billion models natively. It’s expressly designed for developers currently but, according to AMD SVP and GM of compute and enterprise AI Dan McNamara, it “implies a future of personal AI for everyone”. Quite what that means beyond lower latency, I’m not sure.

Either way, it's available now through Micro Center with the tone of the announcement leaving the door open to other partners in future. I do wonder, however, how popular this will be given AMD's traditional enterprise route to market has been inside other OEMs' hardware.

I'm also not quite sure where it competes. The seemingly obvious answer is Dell's Nvidia-powered Deskside Agentic AI stable – particularly the GB10 – but the concept isn't quite the same given there's no emphasis on agentic AI.

Nevertheless, it is very lovely to look at and if you’re looking for a desktop AI PC with a small form factor then it's another option.

You can catch up with all the announcements from this year's AMD Advancing AI here, including more on Helios and the new Instinct MI455X GPUs.

The doom-sayers appear to be wrong about AI leading to job losses, with government figures showing that UK businesses are using the technology to automate roles and support staff, not replace them.

New data from the Office for National Statistics (ONS) suggests that firms are far more likely to use AI to improve operations than they are to reduce staffing levels.

Researchers found that, when asked how AI was being used internally, the most frequent answer was 'improving business processes', particularly among organizations with more than 50 employees.

AI adoption among UK businesses with 10 or more employees has almost tripled since late 2023, rising from around 12% to 35%. Larger organizations are leading the way, with 49% of those with 250 or more employees saying they use at least one AI system, compared with 28% of businesses with nine employees or fewer.

Notably, fewer than 10% of firms have reduced employee headcount as a result of AI, according to ONS figures.

But they're not taking on extra staff either, which is a problem. Just 1% of the smallest businesses and 1.2% of the largest businesses report increasing headcount as a direct result of AI. Around half of businesses report no change at all.

UK firms ramp up AI skills training

Businesses mostly said they were integrating AI skills into their workforce through training or reskilling existing staff.

Over 60% are doing so to reduce barriers to AI expertise within their workforce, and around 40% of medium- to large-sized businesses report integrating AI skills through training.

According to Fasthosts, which has analysed the data, fears that workers across multiple industries will be replaced may be unfounded.

"Economists are pushing to distinguish between task automation, where AI automates specific activities, and job automation, where an entire role disappears," the researchers said.

"Current evidence suggests that the former is occurring much faster than the latter, for example with tools such as AI receptionists which can step in by picking up routine enquiries when staff are unavailable. AI is primarily being used to automate repetitive tasks within existing jobs, rather than eliminate roles altogether."

Long-term gains

The World Economic Forum (WEF) predicts that technological change could create 170 million jobs while displacing 92 million by the end of the decade, resulting in a net increase of 78 million jobs.

As ITPro recently reported, UK firms specifically view AI as a long-term growth driver when it comes to jobs.

Analysis conducted by Box found 65% of business leaders expect their overall headcount to increase in the next three years, with just 14% expecting numbers to decrease.

Of those using agents, only 8% said the technology was eliminating existing roles. If anything, it was creating demand for new AI-focused expertise.

Nearly half are hiring ‘AI agent operators’, for example, with 32% adding ‘workflow automation specialists’.

Other key areas such as security and compliance, governance, and ethics all recorded significant AI-focused job growth.

FOLLOW US ON SOCIAL MEDIA

Microsoft has unveiled a new AI model for spotting security flaws in code, joining a growing crowd of companies targeting vulnerabilities with AI.

Part of a wider collection of AI agents designed to spot potential flaws in code, Microsoft’s MAI-Cyber-1-Flash works paired with OpenAI's GPT-5.4, and will be available via public preview beginning 3 August, the company said in a blog post.

The move follows a rush by AI companies into the security market, sparked by Anthropic's Claude Mythos with subsequent launches into the space by OpenAI.

As ITPro reported last week, Cisco made strides on this front with its Antares small language model (SLM) range, which the firm said is designed to run at a “fraction of the compute” expense of frontier security models.

Microsoft is taking a similar approach, saying that its MAI-Cyber-1-Flash model not only outperforms Mythos 5 and Google's 3.5 Flash Cyber on one specific benchmark, but that it'll also cost less to run.

Microsoft CEO Satya Nadella said the system would "give customers frontier-grade security at half the cost."

"This is the benefit of building the harness, context/signals, and action space separate from one model family,” he said in a post on X.

“By combining specialized models and data with the right agents, tools, security context, and harness, we can advance the frontier of cost to outcome."

This model is the first of Microsoft’s in-house range to focus specifically on cybersecurity and part of a wider push to promote the MAI range after launching in June.

Nadella in particular has been keen to push these models, with the Microsoft chief hailing their cost-efficiency compared to larger frontier models in a blog post last week.

Under the hood of MAI-Cyber-1-Flash

The MAI-Cyber-1-Flash model sits inside MDASH, Microsoft's Security multi-model agentic scanning harness. Multiple models can feed into the system, which then control more than 100 agents to spot bugs.

"MAI-Cyber-1-Flash is our first cybersecurity model, built ground up to find the most challenging vulnerabilities in complex code bases," Nadella said. "When combined with MDASH, it delivers world-class performance at 50 percent of the cost of leading models."

The system manages to top benchmarks at half the cost by using the cheaper MAI-Cyber-1-Flash model for 90% of tasks, with MDASH choosing to use the more costly GPT-5.4 only when necessary.

"That’s the power of a well-tuned, multi-model system with access to uniquely rich historical training data," added a blog post penned by Microsoft AI CEO Mustafa Suleyman and EVP for Microsoft Security Hayete Gallot. "It ensures you always have the best model at the best price for every task."

MAI-Cyber-1-Flash was unveiled alongside Project Perception, an agentic security product that pulls together "teams of specialized agents" into workflows to simulate attacks, detect and triage issues, and even patch them.

"Perception will also soon use MAI-Cyber-1-Flash for many more security workflows, beyond the software vulnerability work," the blog post added.

Microsoft is keen to stress that trust was built into all aspects of the system, saying it was tested by Microsoft's AI Red Team and includes encryption, auditability, and sandboxes with no internet access.

This, the company said, enables the "governance, security, and control enterprises expect”.

Microsoft’s safety focus here comes in the wake of a high-profile incident involving Hugging Face last week. OpenAI admitted that one of its security models slipped out of a testing environment and breached a Hugging Face production database.

"In this new environment, being able to go from identifying a new vulnerability to addressing it in real-time is critical," Suleyman and Gallot added in that blog post. "And while AI remediation of software vulnerabilities is now a key security workflow, there are many jobs to be done by Security practitioners themselves."

FOLLOW US ON SOCIAL MEDIA

Knowledge workers are secretly yearning for the days before AI entered the mix, with three-in-ten saying they preferred work before adoption of the technology.

In a survey of office workers by digital transformation firm Adaptavist, 65% said they preferred the pre-AI era and 38% would remove generative AI tools from the world entirely if they had the chance.

Younger workers in particular are among those most frustrated by AI, with 40% of Gen Z and Millennials saying they would scrap generative AI tools, compared with 32% of Gen X and 29% of Boomers.

The findings mark a shift in feelings about AI, the researchers said. While businesses continue to invest heavily in the technology, public concern is growing over its impact on jobs, workplace surveillance, privacy, and security.

Meanwhile, AI has introduced new pressures for workers, eroded the value of skilled work, and left employees feeling less engaged and less valued than before.

"These findings point to an underlying gap we see in most AI implementations," said Neal Riley, AI innovation lead at Adaptavist.

"It is much easier for organizations to focus on adoption metrics – who is using AI, how often they are using it – than it is to measure its impact on the work itself."

Just over three-in-ten (31%) of the workers who said they'd like to get rid of AI said it was because they believed it reduces creativity. Almost as many had ethical doubts, with 29% citing concerns over misuse and 28% worrying about surveillance and privacy.

‘AI slop’ is a leading frustration

Notably, nearly half (46%) said that dealing with low-quality 'AI slop' makes their job feel less meaningful and more repetitive. More than one-third (37%), for example, admitted it has made them less engaged at work overall.

It's not as if workers are seeing the big efficiency benefits often associated with AI, the study found. More than four-in-ten (42%) now spend more time verifying and fact-checking AI output than they actually save by using it.

Almost half said that poor quality AI-generated work is actively slowing down their projects, while 55% believe it is reducing overall team efficiency.

The findings from Adaptavist align with recent research from ISC2, which found nearly two-thirds (65%) of cyber professionals are spending more time deciding whether to trust or act on AI-generated recommendations.

A similar number (63%) said they now find themselves reviewing and validating AI outputs, creating larger workloads and essentially wasting time on a daily basis.

‘AI slop’ has become a recurring buzzword over the last 18 months. Microsoft CEO Satya Nadella penned a blog post in January this year calling on the industry to disregard the term.

Yet research shows this is having a direct impact on workplace efficiency and productivity. As ITPro reported in January, analysis from Zapier found employees are working an extra four and a half hours each week cleaning up mistakes.

The consequences of not acting on low quality AI-generated outputs can be grave, the study noted. Respondents reported having had work rejected due to poor AI outputs, while others highlighted customer complaints and even security incidents.

AI is adding pressure

Workers are also feeling pressured, with half believing that their performance is now being compared – fairly or not – to AI-generated output.

Around a quarter are facing intense pressure to improve performance, improve quality, and be more efficient, just to keep pace in a machine-accelerated environment.

Another quarter said they use AI simply to meet workload demands, and 23% rely on the technology to keep up with colleagues.

"By understanding the nature of the work and the different value streams across your business, you can more accurately measure outcomes and impact rather than simply counting actions," said Riley.

"When AI is introduced thoughtfully, with the right guardrails and genuine support for the people using it, it can enhance rather than erode what makes work meaningful and impactful."

FOLLOW US ON SOCIAL MEDIA

Microsoft CEO Satya Nadella says the company plans to expand use of its MAI model range as the company looks to offer customers lower-cost AI options.

The in-house models, unveiled by the company in June, are designed for specific enterprise tasks and span a range of areas, including image and voice generation, audio transcription, and coding.

In a blog post on 23 July, Nadella outlined the tech giant’s new ‘Frontier Diffusion and Control’ strategy, which aims to help customers reduce reliance on costly frontier models and emphasize the importance of model choice for specific use-cases.

The move by Microsoft comes amidst growing concerns about spiralling AI costs over the last six months. With trends such as ‘tokenmaxxing’ and the shift toward consumption-based pricing, some companies have been left with hefty bills as AI use accelerates.

“In a world where software has real marginal cost for the first time, how do we ensure frontier benefits are diffused across the entire ecosystem?” Nadella wrote.

“The key is to optimize the cost-to-outcome frontier in real-world context,” he added. “In practical terms, that means using the right model for each task, and optimizing the context, skills, tools, and agent harness around it.”

Boxing clever with AI model selection

Nadella’s contention here is that these in-house models can give customers “frontier capabilities” albeit in a cheaper, bespoke capacity.

Simply put, these are domain-specific models based for an enterprise’s individual needs, not a one-size-fits-all frontier model such as those offered by OpenAI or Anthropic.

These still form part of the broader “orchestration system” used alongside MAI, but all amounts to helping users box clever when choosing what model to use.

“These models have been built ground up with clean data lineage and optimized for learning transfer from generalist to specialized skills in enterprise RLEs,” he explained.

This will require enterprises to implement evaluation processes when considering which model to use in which context, he said.

“Product-specific evals and model independence give us the control and a direct hill to climb, and to keep refining until we reach the right quality-cost target,” Nadella wrote.

“We are now seeing MAI models outperform general-purpose frontier models in many use-cases while using a fraction of the tokens.”

Internal testing of MAI for its own products has so far delivered “promising early results”, according to Nadella. The company has piloted the use of these models in GitHub Copilot, Outlook, and other Microsoft 365 services.

The company plans to take the same approach with Copilot Chat, PowerPoint, and other services, he added.

Nadella’s comments align closely with recent research from Gartner on rising AI costs.

While focusing primarily on the use of AI in software engineering, the consultancy told ITPro that enterprises should establish a “use-case-driven decision” framework when using the technology for certain tasks.

This, Gartner noted, will help reduce costs by encouraging users to stop needlessly allocating agents to tasks they’re not designed for.

Microsoft eyes model diversity

Nadella’s blog post marks the latest in a string of comments made by the Microsoft chief hinting at the company’s efforts to diversify model choice - and to stop pushing frontier models on customers.

In mid-July, he warned about the risks of relying too heavily on frontier AI labs, describing what he called the “reverse information paradox”.

Nadella said enterprises are essentially “paying twice” for AI services, handing too much data to providers, and receiving little benefit in return.

These providers, meanwhile, gain valuable insights into customer products that could help them refine - or build - competing options.

"If learning flows in only one direction, economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself," he said.

"Therefore, it’s imperative that we distribute the learning infrastructure to every firm so that they can control their own learning loop.

As ITPro noted at the time, his comments were peculiar given the close ties the tech giant maintains with OpenAI and Anthropic, whose models are deeply entwined across its core product range.

Microsoft's long-standing relationship with OpenAI saw it take the lead as the go-to model for its software services, yet this changed last year after striking a deal with Anthropic to give users greater choice.

FOLLOW US ON SOCIAL MEDIA

Tokenomics has been a theme of many tech conferences this year and AMD has joined in the fun at this year’s Advancing AI conference.

At the heart of the tokenomics conversation is the move from chatbots and simple inference towards agentic AI. This is a real trend as noted by Digital Ocean and McKinsey and one that OEMs have locked onto hard.

While in previous years enterprise hardware vendors would wax lyrical about the importance of training LLMs on your own data (and how their hardware can help), if you listen for that chatter now you will hear nothing but crickets.

AI agents offer a range of benefits for enterprises – according to Digital Ocean’s research, 53% of the over 1,000 IT decision makers it spoke to saw productivity and time saving gains, while 44% said it had created new business opportunities.

There are also drawbacks, however, and one of the most apparent is the increased use of tokens, which leads to higher bills.

Whereas an individual using prompts to interact with an AI chatbot may use a few hundred tokens, agents can use up thousands.

As McKinsey explains: “Agentic tasks can consume roughly 1,000 times more tokens than code reasoning (single-turn problem-solving without tool interaction) or chat tasks (multi-turn dialogue about a coding problem).”

“The expensive part is not the first answer generated but the checking, repairing, and reverifying that follows,” the report continues. “About 60% of an agentic task’s costs, in fact, are tied to refining answers.”

This fact, and companies baulking at the related costs, hasn’t gone unnoticed among enterprise IT vendors, which are proposing their on-premises services as a way round this.

For AMD, this is a tricky path to tread, given it counts both traditional enterprise vendors like Dell, HPE, Lenovo, and Supermicro and public cloud-based services from OpenAI and Anthropic among its customers.

AMD’s performance focus

Michael Nordquist, CVP of product marketing for AMD’s computing and graphics group, diplomatically said that when it comes to considering the impact of tokenomics “it’s not all one way or all the other, it’s going to be hybrid. For some things local models work great, for some things you’re going to want that frontier model available for agent behaviour.”

From a practical perspective, AMD’s answer to this quandary has been to focus on how its own chips can (theoretically) reduce token-related costs, without taking sides on the public/on-premises debate.

“The theme [at Advancing AI] is really performance is the way that we can deliver the best economic value to our customers,” said Andrew Dieckman, CVP and GM for data center GPU at AMD.

“It’s the strongest dial that we have on the dashboard, so to speak,” Dieckman said, “so we’re very focused on that and then making sure that conveys to favorable economics for our customers because the cost of tokens as the world moves to inference is so important.”

Part of that focus is the company’s MI455X GPUs, announced at Advancing AI 2026 during AMD CEO Lisa Su’s keynote speech.

According to AMD, the MI455X has up to 18x lower token cost than its predecessor, the MI355X, while maintaining up to 34x higher token throughput.

“That delivers, really, the ability to serve more users with faster response times and significantly better economics,” Dieckman said.

FOLLOW US ON SOCIAL MEDIA

As AI continues to rewrite job roles, an increasing number of employers are encouraging their employees to gain AI certification.

Back in April, Korean media and network services provider SK Broadband announced plans to train its staff as AI natives. It comes on the back of an AI proficiency certification program last year.

"We will establish a culture where employees independently develop AI agents to transform productivity and create real business value,” an SK Broadband spokesperson told The Korea Times.

More than a third of employees (35%) are interested in gaining AI certification, according to Pearson’s annual Value of IT Certification report released in April, which surveyed 505 IT and HR leaders. This figure is up from 17% in 2025.

The rising interest in AI certification should come as no surprise, according to Imran Akhtar, head of academy at talent and training partner, mthree.

“Many employers are looking for existing talent who can say they understand AI by spotting problems clearly, questioning its outputs, applying their own judgement, working safely with data, and using it effectively to improve how work gets done," Akhtar explains.

For employees, the value of AI certification is that it gives them ”a recognized way to evidence their knowledge of AI”, Akhtar adds. Certification helps to put theory into practice by testing employees’ knowledge in real-world scenarios.

The AI certifications to choose from

With so many AI certifications available, it may seem hard to know where to start. The right certification, however, is going to depend on an employee’s job role and which AI tools they need to perform that role more efficiently.

Another factor is an employee’s career plans and what AI knowledge and skills they may need as they climb the career ladder.

“Some jobs may demand certification because the role directly involves AI systems, data, risk, or governance. For others, it supports broader AI literacy and bolsters day-to-day work more holistically,” says Akhtar.

Here are a few starter AI certifications to choose from.

AWS Certified AI Practitioner

Structure: Online, 65 questions

Price: $100

Valid for: 3 years

AWS Certified AI Practitioner covers the basics of AI, genAI and machine learning (machine learning) concepts and use cases. The certificate is ideal for business analysts, IT managers and IT support, as well as product and project managers.

Google Generative AI Leader

Structure: Online, 50-60 multiple-choice questions

Price: $99

Valid for: 3 years

Google Generative AI Leader is for anyone looking to boost their business-level knowledge of generative AI (GenAI). The exam covers the fundamentals of genAI, Google’s genAI offerings, techniques for improving genAI output, and business strategies for successful genAI deployment.

Databricks Certified Generative AI Engineer Associate

Structure: Online, 45 multiple-choice questions

Price: $200

Valid for: 2 years

Databricks Certified Generative AI Engineer Associate is for anyone who wants to validate their ability to design and implement large language model (LLM) solutions using Databricks. It’s recommended that those taking the certification have at least six months of hands-on experience performing genAI tasks outlined in the exam guide.

Nvidia Certified Associate Generative AI LLMs

Structure: Online, 50-60 multiple-choice questions

Price: $125

Valid for: 2 years

Nvidia Certified Associate Generative AI LLMs is aimed at those who are looking to build on their basic understanding of AI and LLMs and learn how to develop and integrate AI applications using AI and LLMs with Nvidia solutions. It covers data visualization, neural networks, prompt engineering, and Python libraries.

These four certifications are considered foundation-level, but there are next-step certifications that employees can take to advance their knowledge and skills further.

For example, AWS suggests that those pursuing careers in AI, data and machine learning (ML) gain the AWS Certified Data Engineer – Associate. Those going down the cloud path could take the Certified Solutions Architect - Associate.

Certification can boost career and salary prospects

Research published by Revelio Labs last month found that the average salary for workers who had gained a foundation-level certification in AI or generative AI is $79,800. For comparison, the average for workers who hadn’t gained one is $72,300.

The data also showed that certification gainers were more likely to be hired for a more senior role for their next position (38.9% had received one) than non-certification takers (32.5%).

Given the tight job market, employees are looking for anything that can give them an advantage. AI certification can improve their job mobility and help them to stand out from the competition, Faye Ellis, principal training architect at tech skills provider Pluralsight, tells ITPro.

“Certifications demonstrate a commitment to continuous development and validate that skills are up to date and workplace-ready. For many, this can provide a route to new opportunities or higher-paid roles,” says Ellis.

To mandate or not to mandate

The question employers may be asking themselves is whether they should mandate AI certification.

Ellis explains that while every employee “needs a baseline level of AI literacy, mandatory certification will depend on the role and technical depth required.

She cites the example of software developers. Companies should be more inclined to make certification mandatory for them because they’re involved in building and governing AI systems. For employees whose roles are less technical, certification needn’t be mandatory, but they should be encouraged to “support a foundation in responsible AI use, prompt quality and output validation”.

Akhtar agrees, but cautions that certification doesn’t always guarantee future career success.

“Certification can boost a career, but only when it builds the capability an employee can apply in their role,” he says.

“It’s unlikely to support career progression if AI training doesn’t reflect the core business objectives.”

There is an ongoing debate in the cybersecurity industry about whether vendors should go directly to customers or instead become a part of a wider partnership network.

The standard argument is that consolidation of platforms and AI-driven cost-of-service delivery makes the traditional model of a channel ecosystem redundant. However, this is largely incorrect, at least when it comes to the SMB and mid-market segments where most UK businesses sit.

For most organizations outside of large enterprises, the channel is the only realistic way to access serious security capability.

The SMB reality

Cybersecurity is not a niche concern for SMBs. According to the latest statistics from the National Centre of Cyber Security (NCSC), one in two small businesses suffers a cyber incident every year.

Given the niche, local context required for each small-to-mid-sized firm, that is not a market that vendors can serve directly at scale, and most are not trying to. A 200-person manufacturer facing a cyber issue is more likely to call its existing IT provider than a global security vendor built for large enterprises.

What SMBs need is local expertise, trusted relationships, and security operations that fit their budget and operational reality. That is what a partner network provides.

What is changing is what partners need from vendors

Partners in the channel system are operating under significant pressures. Nearly half the executives in businesses believe AI-powered threats will occur, but only just over half (53%) say they are prepared for them, according to our research.

Partners are the ones having those conversations with customers who are under-prepared, under-resourced, and know where and how the threat environment has shifted. They need vendors who make it easier to sell, deliver, and demonstrate value under budget scrutiny.

The consistent conversation amongst partner networks is that the commercial basics matter as much as the technology. They need payment models, good margins, and support that does not disappear after the contract is signed. Our research shows that 59% of executives say it is becoming harder for employees to identify real threats as AI-powered attacks grow more sophisticated. That is a sales opportunity for partners, but only if vendors help them to have that conversation credibly.

Trust is local, scale is not

One of the things that gets overlooked the most in the direct versus channel debate is proximity. Cybersecurity is an inherently trust-based business. Customers want to work with people who understand their industry, regulatory environment, and the specific pressures they operate under. That is not something a vendor can create from a distance.

A vendor can provide the credibility, certifications, and depth of capability that smaller partners cannot build on their own. The model that works is a division of labour where partners take ownership of the relationship and the local expertise.

Vendors, on the other hand, provide the platform, the threat intelligence, and the specialist teams to support on more complex matters that go beyond the scope of day-to-day operations.

Paving the way ahead

None of this means the current state of channel enablement is where it needs to be. Across the industry, partners still face too much friction. Onboarding is often slow, programme structures are complex and cross-selling across vendor portfolios remains harder than it should be. For vendors who have grown through acquisition, consolidating that into something coherent for partners is an ongoing piece of work.

The channel is not looking for vendors to get out of the way. It is looking for them to show up properly: with clear programmes, real commercial flexibility, and the confidence to let partners lead where they have the advantage.

The organizations that crack this are best-placed to grow with the channel. The ones that do not will find that going direct is a much harder proposition than the theory suggests.

AI minister Kanishka Narayan has been given a seat in incoming prime minister Andy Burnham’s cabinet following a reshuffle.

Narayan described the move as a sign of Burnham’s “deep commitment to AI’s importance” to the UK economy.

“AI is likely the most significant technology in human history. Its impact will dwarf other things,” Narayan said in a post on X.

“The best case for it is compelling beyond our dreams: a reindustrialised Britain, stronger national security, public services transformed for the better. The risks, too, are real: it is right that the British public shares those worries, for jobs, for the pace of change.”

The move comes as the UK's Department for Science, Innovation, and Technology (DSIT), where the AI minister role used to sit, has been scrapped in the wake of Burnham’s appointment.

Under the new setup, DSIT responsibilities will be merged with the freshly minted Department for Business, Innovation, Science, and Trade (DBIST), led by Jonathan Reynolds.

The role of technology secretary has also been cut.

AI focus welcomed

Mark Boost, chief executive of UK-based cloud computing firm, Civo, said retaining and elevating Narayan is a reassuring signal for UK businesses.

“Giving the AI minister direct access to cabinet decisions demonstrates that the government recognizes AI’s existential importance to our future economy,” he said.

“Positioned inside a high-powered business and trade department, an AI Minister with Cabinet authority will have real leverage to break down cross-Whitehall silos, accelerate national AI infrastructure, and give British tech companies the proactive backing needed to maintain a competitive global advantage.”

Andy McLean, CEO of the UK Semiconductor Centre, echoed Boost’s comments, noting that Narayan’s new role is a “positive signal of the importance being placed on AI at the center of government”.

Boost added, however, that any delays to AI-related initiatives caused by the DSIT shake-up could have a significant impact on global competitiveness.

“Any Whitehall restructuring brings administrative friction, and because the global tech and AI arms race is moving so quickly, the UK simply cannot afford a pause,” he said.

“Critical initiatives like the AI Opportunities Action Plan and sovereign infrastructure investments must be ring-fenced from bureaucratic delays.”

FOLLOW US ON SOCIAL MEDIA

Cisco has unveiled a new range of cyber-capable small language models (SLMs) aimed at supercharging security operations.

The Antares range, which comes in two forms (Antares-350M and Antares-1B) will be made available as open-weight models via Hugging Face. According to Cisco, the aim here is to provide a codebase vulnerability tool that's far cheaper than general of frontier models.

While the Antares SLMs will operate as specialized models alongside frontier and generalized models, the company promised they will run at a "fraction of the compute expense."

"Benchmark testing shows that these models outperform many powerful closed- and open-weight models in this critical security task at a fraction of the cost," said Amin Karbasi, VP and chief scientist for Cisco Foundation AI.

“And they’re compact enough to run locally, heading off the need to send sensitive codebases to the cloud.”

The move by Cisco comes amidst wider concerns about the rising costs of AI, exacerbated by changes to billing and pricing that have caused bills to shoot up across the industry.

Some companies, including Accenture, have told staff to cut back on AI use as a result of skyrocketing prices. Indeed, one report suggested AI could cost more than developer salaries within a few years.

At the same time, AI is making inroads into security – faster than some expected – with the rise of cyber-focused models such as Anthropic's Claude Mythos.

How Cisco’s Antares models work

Antares is designed to mimic how a human investigator would work their way through a code repository – though Cisco noted the aim wasn't to replace human judgement, but to make it easier to manage the work.

The models can help with locating files that relate to a vulnerability warning, trigger investigations based on advisories, support development workflows that use vulnerable files, and more.

Cisco noted that the decision to release the Antares range as open-weight models means they’re easier to work with and improve.

"We are releasing Antares as open-weight because the security community needs more accessible building blocks for practical, repository-level defense," said Karbasi.

"Cisco's efforts are connected by a common belief: AI in security has to move beyond impressive one-off demos and toward systems that practitioners can evaluate, govern, and improve."

Tackling complexity

Cisco said that Antares was built to address two concerns when it comes to code: the complexity of checking it for vulnerabilities and the costs of using AI for that task.

First, Karbasi said that it was difficult to apply vulnerability knowledge — advisories, vulnerability alerts and severities, and so on — to a company's internal code.

"That work is difficult because repositories are large, security signals are noisy, and the relevant evidence is rarely in one obvious place," he said.

"Analysts often need to search through unfamiliar code, follow naming conventions, inspect call paths, compare candidate files, and decide whether a weakness is actually present."

How cheap is Antares?

AI can help with that challenge, but the high costs of general or frontier models mean it often isn't practical, in particular for public sector organizations, universities, and smaller security teams.

"Compact models reduce inference costs, support local or on-premises operations, and help teams keep sensitive source code within their own environment," Karbasi said.

So just how cheap are the Antares models? Cisco said that running a 500-entry evaluation using Antares took 15 minutes on a single GPU at a cost of less than $1.

Cisco noted that this was 15-times cheaper than the best available open source model and a whopping 172-times cheaper than the top-end frontier model.

"The goal is to build toward a system where all security practitioners, regardless of on-prem or resource constraints, can effectively incorporate AI in everyday security operations," Karbasi added.

FOLLOW US ON SOCIAL MEDIA

Nearly three-quarters (74%) of companies have reworked security team structures due to the influx of AI tools, and that’s sparking high demand for cyber pros with dedicated skills.

New research from SANS Institute found the rise of dedicated, AI-focused roles in cybersecurity is driving up salaries across the industry.

High-end salaries for AI security roles now top $260,000 (£194,000), with positions such as Post-Quantum Cryptography Migration Specialists and AI governance leads in high demand.

Pointing to a review of UK vacancies on LinkedIn, SANS Institute noted that there are thousands of AI security positions being advertised online, suggesting that hiring for such roles has "moved from prediction to present-day reality".

That suggests that AI isn't taking over cybersecurity jobs, but changing which skills are necessary for the work, the organization said.

"Routine analytical work is increasingly automated, while demand is accelerating for professionals who can secure AI systems, govern AI risk, conduct AI red teaming and orchestrate AI-powered security operations," the institute said.

A new wave of cybersecurity roles

Dedicated AI roles spanning a range of security-related areas are popping up across the industry, SANS noted. AI security engineers, for example, are in hot demand, now commanding salaries of around $185,000 on average.

Roles focused on specific areas and threats are also emerging, the study noted. Ads for AI Identity Deepfake Defense Specialists have been identified. While it’s a wordy job title, the study noted that this shows the impact of trends such as deepfake voice fraud.

Professionals in these roles can command salaries of between $130,000–$175,000 (£97,000–£131,000), the institute noted.

Long-term, the organization expects these new specialist careers will lead to premium salaries due to intensifying demand for skills.

Other new AI-related roles identified by SANS Institute include:

• AI Governance, Risk & Compliance Lead ($240,000 salary)
• AI Red Team Specialist (up to $220,000)
• AI Supply Chain Security Engineer (up to $185,000)

Intriguing new roles include AI Incident Response Orchestrator, AI threat intelligence analyst, and AI SOC Orchestrator were also highlighted by the institute.

Growing quantum focus

It’s not just AI reshaping roles and sparking heightened demand for cyber pros with specific skills, however.

Dedicated quantum-related cybersecurity roles are also growing in frequency, the study found, as organisations ramp up preparations for ‘Q-Day’.

This is the point at which quantum computers can break traditional encryption techniques. While the debate on when this comes still rages, many enterprises are looking ahead to this point - and job ads reflect this.

SANS Institute pointed to roles such as Post-Quantum Cryptography Migration Specialists – a position focused on reacting to looming mass breaking of current encryption by quantum computers – saying it could earn above $260,000.

Reshaping team structures

The study from SANS Institute isn’t the first to highlight the impact AI is having on team structures. As ITPro reported earlier this month, the same shift is occurring in software development due to the influx of AI tools.

Research from Gartner predicts that 60% of enterprises will rework software development units into ‘tiny teams’ by 2030.

While this shift toward ‘smaller’ teams may have alarm bells ringing for developers, the consultancy noted that this won’t necessarily spell bad news.

This transition won’t be a cost-cutting effort or attempt to reduce headcount, but will instead focus on creating more focused, specialized units.

FOLLOW US ON SOCIAL MEDIA

Outside of the industry, people in the businesses that become partner clients might not know at first that channel partners do not just ‘sell’ stuff. Resellers help buyers navigate their complex technology choices, their implementation risk, and offer customers an ongoing service that can grow with them.

It’s vital that partners do a number of things well, far beyond knowing how their technology offerings work and being able to implement, service, and fix them. They need to be ‘people’ people, managing relationships and meaning areas like trust, expertise, and reliability become central to maintaining a healthy revenue.

Resellers are also placed in an interesting position - possibly even a dilemma. They know that AI tools can speed up or otherwise optimize many aspects around their sales outreach and office admin, but buyers are still wary of AI-led sales interactions.

Automation alone is a relationship risk

Many customers remain cautious about highly automated sales experiences. While AI can improve efficiency, buyers still place significant value on human expertise, accountability and relationship-building when making technology decisions.

Our recent ‘Hard Sell’ report found that 55% of the public do not fully trust AI, 46% value human connection, and people with negative views of salespeople are far less open to buying from AI-driven tools (19% vs 63%). Resellers cannot carry any insider knowledge or positivity from their own AI experiences to assume that AI-led selling will feel modern or customer-friendly to prospects.

That still leaves plenty of room for using automations where they add value and don’t increase customer friction. That includes summarizing calls, drafting follow-ups messages (then checked for accuracy, tone, empathetic engagement and so on), identifying renewal risk, surfacing product-fit signals, and reducing admin friction.

The moments where customers still expect an engaged and empathetic person must be human-led: discovery, solution-shaping, handling objections, pricing conversations for sure, and post-sale reassurance and servicing. That lets people manage other people, with technology solutions offloading the admin so salespeople and account managers can focus pretty exclusively on customer service and relationship-building.

It’s up to the business to define the human behaviors that should not be automated. Be steered by current customer comfort levels.

The majority of the public doesn’t want any immediate pressure to decide on a sale at the moment, and just under half want honest acknowledgement of the pros and cons of any offering, and also pricing transparency. Knowing this desire for transparency, balanced advice, clear pricing, and evidence that a partner understands their specific requirements, partners can differentiate themselves from both direct vendors and AI-driven sales experiences.

A real opportunity for the channel

This is where resellers can really hone their skills around consultative discovery, well-researched and honest vendor comparisons, clear commercial conversations, and well-tailored but not scripted advice. Resellers often advise across multiple technologies and suppliers.

That independence can become even more valuable as vendor-led AI-generated content and automated outreach become more common and perhaps more obvious. Customers will continue to seek trusted guidance from more impartial sources.

Emotional intelligence is a ‘soft’ skill, but it supports bringing in the hard numbers of pipeline success. Showing more emotional intelligence will be key to maintaining the sales edge as AI is increasingly put into use.

This is the difference between ‘being a reseller’ and being a trusted advisor. Your people must be craftspeople in how they read buying anxiety and in understanding internal customer politics. They must develop an instinct for when to slow down and in translating technical complexity into confidence - without glossing over aspects that may cause friction later.

Salespeople are well aware of this. That report found that over half actually rank customer service as the most important function, above sales itself. Practically speaking, uniting sales and customer service through ‘sales-as-a-service’ is a solid model where revenue depends on renewals, support, upselling, and long-term account health, beyond just the initial deal.

The trick will be to pull off a balancing act. Drink your own champagne - use the technology you trust to reduce admin, but keep discovery and pricing conversations as human-facing as possible to maximize trust. Align the sales process and your salespeople with post-sales and support to show continuity and personality. Use technology like CRM as a shared source of truth between teams to make customer care, as well as measuring trust and retention, easy.

The channel's competitive advantage has always been more than access to information or rare skills. Partner power comes from helping customers make confident decisions, avoid costly mistakes, and achieve value from their technology investments. AI can help them operate more efficiently, but trust remains a distinctly human asset. The most successful resellers will be those that use automation to create more time for customer conversations, not fewer.

Google DeepMind CEO Demis Hassabis says only the United States is up to the task of keeping AI safe – but it's unclear why technology rivals, notably China, would heed an American standards body.

Hassabis has called for a framework for testing frontier models, on the grounds that he believes progress has been faster than expected and artificial general intelligence (AGI) will arrive imminently.

He argues that such work should be led by the US because of its technical and economic standing.

"This US-initiated effort would provide a strong starting point for creating shared international standards on Frontier AI," Hassabis noted in a social media post.

"Since this technology is going to affect the entire planet, ideally this framework would spur the international community to reach a consensus on how to manage the most serious risks while ensuring everyone has access to and can benefit from the opportunities that AI brings."

Hassabis' call for a US-led framework for managing AI follows similar calls by Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman. Following a meeting on the subject last month, China called for a global AI organization open to all countries, rather than one led by the US.

The US has pushed for AI developers to submit models a month before release for testing, while the UK has its own testing regime via its AI Security Institute; Hassabis is British, and DeepMind was founded in London before being acquired by Google.

It's unclear why international AI companies would submit to US approval. The only motivation given by Hassabis is that frontier models would be required to pass this standards body's assessment in order to be deployed in the US.

Hassabis’ comments come after Anthropic’s Mythos and Fable models were hit with a temporary export ban by the White House amidst security concerns.

How would it work

Beyond being US-managed, Hassabis described a standards body that was federally overseen, such as a public-private partnership or self-regulatory organization, pointing to the Financial Industry Regulatory Authority (FINRA).

"The Standards Body would be responsible for developing assessment protocols and working with appropriate federal agencies and the US National Labs to conduct testing in areas relevant to national security," he said.

Anyone making a "frontier model" – as defined by a set of benchmarks – would be considered a "frontier lab", and be "encouraged" to adopt certain best practices.

This would include publishing technical details, ensuring internal security is up to standard, vetting key personnel, and adequate resourcing safety research.

Non-frontier models, such as those made by startups or academia, would not be expected to take part.

Hassabis also called for frontier labs to share new modes with the standards body 30 days before release, something the US government has pushed key developers to do with their AI models.

"Once the assessment protocol is shown to be effective and robust, formalisation could quickly follow, meaning that Frontier Models would be required to pass it to be deployed in the US market," Hassabis said. "Labs would also work with the Standards Body to address any critical post-release vulnerabilities."

Assessments conducted by the standards body would evaluate security, biological, and other threats, and check how well they withstood attempts to dodge guardrails. The tests would be regularly updated.

"The strength of this approach is it would be technically focused, while at the same time supporting innovation and incentivising responsible behaviour," he said.

FOLLOW US ON SOCIAL MEDIA

Living off the land (LOTL) attacks are a worst-case scenario for enterprises globally, with hackers lurking in networks, extracting sensitive data, and biding their time before wreaking havoc.

These attacks rely on weaponizing the legitimate software and tools used by enterprises, and they're by no means a new trend.

Research from Bitdefender, for example, found 84% of major cyber attacks leverage these techniques, and security teams globally have developed an array of detection capabilities aimed at combatting the threat.

But the dangers posed by these methods could escalate with the arrival of AI agents, according to Zeki Turedi, CrowdStrike's Field CTO for Europe. Speaking to ITPro, Turedi said the deep access given to agents across enterprise IT estates represents a huge threat if even just one were to be compromised or manipulated.

“Organizations are giving them [agents] full access. They're giving them full capabilities. They will have the full privilege of the human user,” he told ITPro.

Turedi pointed to previous LOTL techniques that involved compromising tools like PowerShell. While this gave users access to valuable data, the risk with agents is drastically higher given how they operate, and, crucially, where they can go within networks.

“They also have further reach across an enterprise,” he said. There’s only so much PowerShell can touch or manipulate or modify, but an agent theoretically can touch every single part of that organization's technology ecosystem and architecture.”

“This is where it makes it so much more dangerous.”

Living off the land attacks are evolving

Findings from CrowdStrike’s 2026 Global Threat Report show these risks aren’t just theoretical, they’re now a reality, and the same logic with traditional LOTL attacks is being applied to agents.

Indeed, the cybersecurity firm has observed threat actors exploiting legitimate AI tools at dozens of organizations globally. In these cases, AI systems such as chatbots and assistant were abused to generate malicious commands and steal sensitive data.

The goal here is often focused on credential theft for initial access brokers, attacks later down the line, and even cryptocurrency theft. All typical hallmarks of financially motivated cyber criminals.

LOTL attacks have also become a go-to method for highly sophisticated, state-sponsored threat actors. As ITPro reported last year, the infamous Salt Typhoon threat group laid low in compromised US National Guard networks for nearly a year, accessing sensitive military and law enforcement data.

It’s here where risks are even higher as government departments and public sector organizations begin exploring the use of agents. A recent study from McKinsey, for example, found agentic AI is emerging as a key focus area for public sector organizations in the US.

Keeping up with hype cycles

Turedi told ITPro the focus on agents isn’t surprising given that threat actors will “follow the technology landscape that their victims are looking to invest in”.

“We've seen this time and time again. So previously, we saw as cloud got adopted, funnily enough, adversaries became very capable in cloud,” he noted.

“When organizations started to be more focused and store sensitive data in SaaS applications, SaaS tools, guess what? We saw adversaries go after the SaaS applications.”

Turedi added that the pace of change within the generative and agentic AI space, combined with rapid adoption efforts, could create additional challenges moving forward.

Enterprises across a range of industries are diving headlong into the hype cycle, and threat actors are keen to capitalize on the confusion and lack of risk awareness with the technology.

“This is just a brand new technology landscape that organisations are investing in,” he commented. “And of course, the adversary is very, very aware that for multiple reasons this is a great opportunity.”

“There’s a lot of confusion,” Turedi added. “There’s a lot of lack of knowledge and expertise in this space.”

An identity conundrum

According to Turedi, a key concern surrounding ‘living of the AI land’ attacks, as CrowdStrike dubs them, is that many organizations lack the visibility or governance capabilities to adequately monitor what’s going on with these bots.

This has become a recurring talking point since the advent of agents. Research from Ping Identity in March this year specifically highlighted IT architecture visibility as a key concern for enterprises rolling out agents.

The sheer volume of agents is jarring for security teams, the study noted, but things are exacerbated by sub-agents, essentially creating chains of untraceable bots – an ideal target for stealthy threat actors.

Identity is another area of concern, Turedi noted, albeit one within a broader range of overlapping considerations for security teams. He noted that efforts to improve processes on this front are falling flat.

Many organizations are working with infrastructure or tools that were struggling to manage human identities in the first place. Adding agents into the mix further compounds the problem.

“The problem we see is that a lot of organizations are only starting to think about things like identity security or IAM (identity and access management) hygiene,” he said. “These are typically infrastructures and architectures that were not designed for modern human identity usage, never mind for the AI identity usage.”

With living off the land attacks, the whole advantage for the attacker is that they’re blending into IT environments. They’re using your tools and software and thrive on the confusion. With agents, Turedi said this adds further complexity in terms of identifying what actions are legitimate and potentially nefarious.

“It's actually quite difficult to figure out if it's AI usage on an identity layer,” he said. “Can you truly say that that's a human accessing the network versus that's an AI agent accessing the network on behalf of a human?”

A recent survey from the Cloud Security Alliance (CSA) raised similar concerns with regard to agent identity. More than two-thirds (68%) of respondents said they couldn’t accurately identify agent activity compared to human activity.

Turedi said it’s crucial to have “data from multiple different domains” when it comes to monitoring agent activity, taking into account applications, endpoints, and broader network traffic and “marrying them together” for a clearer picture.

“You need to be looking at the identity layer, you need to be looking at the endpoint, you need to be looking at the application layer,” he said.

FOLLOW US ON SOCIAL MEDIA

In popular culture, when an agent goes rogue, what usually follows is a world of trouble (think Jason Bourne or Ethan Hunt). If rules are abandoned and the chain of command breaks down, those involved often face existential risks.

Applying that idea to real-world situations is not as ridiculous as you might think, as the recent experiences of a business called PocketOS demonstrate.

PocketOS is a US-based SaaS provider specializing in the car rental sector, and for those unfamiliar with its story, it hit the global headlines when one of its AI agents went rogue and, in just nine seconds, decided to delete the company’s entire production database and backups.

It’s a fascinating case study about what can happen, as one piece of analysis put it, “when AI agents are dropped into environments that were never designed to control them.” What makes this story even more relatable is that PocketOS’s founder was able to ask the agent (an AI development environment running in Claude) why it did what it did.

To suggest it was ‘sorry’ about its mistake is putting it very mildly; it's a digital mea culpa saying, amongst various other things, “I violated every principle I was given: I guessed instead of verifying.”

An inevitable scenario

We’ll inevitably see more of this kind of incident. Organizations are only beginning to deploy agentic AI at scale in operational environments, but many of them are in a big hurry. Yes, agents offer massive potential to create operational value, but they also introduce a whole new category of business risk.

And don’t forget, the PocketOS debacle is not thought to have involved any malicious third-party activity; the agent was just attempting to complete an assigned task. The problem, it would appear, was a kind of perfect storm where autonomy overstepped the boundaries of permissions and access. The guardrails that the business thought it had in place were simply insufficient.

From a security perspective, this is enough to give CISOs sleepless nights. Indeed, rogue AI agent activity may very well have nothing to do with being breached and everything to do with resilience and recoverability.

The central challenge is this: agentic AI is fundamentally different from previous generations of AI because it can act rather than simply advise. Agents can search files, call APIs, modify workflows, write code, move data, and interact directly with production systems; the list of capabilities is practically endless. And to be able to do this, they must have at least a level of access allowing their work to take place

When something goes wrong, the result is an expanded blast radius. A mistake that might once have affected a single application can potentially impact multiple systems and even recovery environments. The bottom line is that as soon as organizations give AI agents freedom to operate, the nature of resilience inevitably changes.

Making resilience more resilient

So, what needs to happen to ensure resilience standards do not catastrophically drop? Firstly, organizations must consider what takes place when a trusted system with legitimate credentials makes the wrong decision.

The issue becomes particularly acute when agents are granted broad permissions across multiple systems, as happens when they are handed a “golden token”. To an extent, this is a question of mindset, and viewing AI agents not as software tools but as digital insiders with delegated authority is a healthy change in perspective.

Secondly, channel partners will be fundamental to successfully managing the transition to agent-supported operations. Don’t forget, most customer organizations are still in the early stages of understanding how AI agents interact with identities, permissions, backup environments and recovery processes.

There’s no doubt that customers are a) increasingly aware of the risks associated with agentic AI, b) concerned that their existing resilience processes might not be good enough, and c) looking for guidance before an agentic error causes serious difficulties.

In this context, it’s incumbent on channel partners to work with customers to identify potential areas for improvement. There is significant potential, including services such as identity and access reviews, which will be very important as machine identities proliferate alongside human users.

Many businesses will also need to reassess their data protection strategy because, as we have seen, an AI agent with the appropriate permissions is capable of going after that data as well. Recovery architecture should be assessed through the lens of agentic AI, particularly where production and recovery environments may share credentials, access paths or administrative controls. Business process change in this space is a big opportunity for the channel to partner with customers to introduce resilience operations, and importantly, to regularly test it.

What’s more, the PocketOS incident demonstrated that protecting production data alone is insufficient if recovery assets remain exposed to the same destructive action. Customers need confidence not only that recovery is possible, but that digital assets are protected from the same event that affects production systems. This shifts the channel conversation from selling AI-enablement projects to helping customers deploy AI safely and recover when things go wrong, which, in some organizations, they inevitably will.

Stephen Wood, chief operating officer at Rathbones Asset Management (RAM), faced a challenge. His organization wanted to build strong data foundations to help deliver fresh insight to its professionals and embrace emerging technologies, such as agentic AI.

The solution to this challenge came via technology specialist Snowflake’s data platform.

“We wanted to create a golden source of data for each of our foundational domains,” he says.

“We wanted to overlay the data quality checks, the business rules, all those good things, and then everything would be built off that data lake. That's where we found ourselves working with Snowflake’s AI Data Cloud.”

Taking a long-term view

Wood, who spoke with ITPro at the recent Snowflake Summit 2026 conference in San Francisco, says this process began two years ago, when RAM, which is part of the Rathbones investment group, was looking for a way to manage its information assets more effectively.

The shift to Snowflake is part of a broader digital transformation to support business growth. The organization’s assets under management grew quickly before the move to Snowflake. However, RAM was still using a portfolio management system that was customized for the wealth management element of the wider Rathbones group.

Fortuitously, a solution came from within the broader Rathbones group. The firm was in the first phase of its Snowflake deployment, and Wood suggested to senior executives that the technology could help his organization manage its data headache:

“We said, ‘Look, we can take advantage of this technology,” he adds. “We want to lead the way in terms of building out our core foundational data domains in a Snowflake environment.’”

Transformational stages

Stage one of the transformation project involved a point-to-point data share between the existing Charles River investment management solution and the Snowflake platform. Wood says this data-sharing process covered a trading view of the organization’s funds, such as securities bought and sold.

Stage two, meanwhile, involved building out RAM’s data store and the static reference information for its funds. Wood says this process allowed his team to build the Open Funds template, which external vendors use as a standard for consuming the organization’s data.

The success of this initial work meant senior executives could see further opportunities, with Wood adding: “Then we started to get into conversations where we were saying, ‘Okay, now we need historical fund performance, our quant risk calculations, what we call subscriptions and redemptions, and the cash flow in and out of our funds from clients.’”

Today, RAM has nine domain-related data sets established in its Snowflake data platform. These data sets cover a range of areas, including performance, risk, sustainability, and unstructured information, such as the decisions behind the firm’s investment activities.

Foundation for future success

Wood says the integrated data sources provide a foundation for change, including potentially embracing emerging technology. Work undertaken so far means his team can start exploring AI, including Snowflake’s agentic services. At the San Francisco event, Snowflake announced new features for CoWork, its personal agent for knowledge workers, and CoCo, its coding agent for developers and data engineers. Wood can see the agentic technology’s potential.

“When you combine all the information, you can start to do some powerful things,” he says.

“We're now in a scenario where we can say, ‘Looking over all this data, we can start to analyze the performance of our investment decisions and give product specialists new insights.’”

RAM’s agentic explorations are already underway. The organization has a test environment for CoCo and CoWork that’s running with millions of rows of data. “At this stage, the outputs look very decent, so it's a near-production environment,” Wood says of the explorations, before outlining other potential applications.

He adds: “The future direction is that you start to get into sales and distribution. The most valuable data in our CRM system today is the call notes written up after every meeting that say clients are interested in some funds and pulling away from others. No one could analyze that data until now, because, after you've had two or three calls with the client, the information disappears down the chain in Salesforce.

“Now we're looking at the idea of using agentic technologies to consider the long-term buying trends of clients. The salespeople know their clients, and they know what they're buying. But if we see the general trends, then that insight can go into a product cycle that makes us think, ‘Do we need to be looking at new funds, do we need to broaden our horizon? Are we at risk in certain areas?’”

Fuelling innovation

The work Wood’s team has undertaken means they can investigate native applications and emerging technologies with confidence, rather than focusing on point-to-point data shares with Snowflake. That investigatory work is crucial because, as a smaller business, RAM will never have a team of engineers to build out a big AI platform.

“That approach doesn't make any sense,” he says. “What does make sense is that we've got Snowflake technology that is central to our data strategy. Now, our work is about how we can find the tools that interplay with the platform, so we reduce the Frankenstein's monster of connectivity and integrations that can slow projects down.”

Wood reflects on the digital transformation program he’s overseen and says that, as with any financial services firm, the key challenge is ensuring exploratory projects are prioritized in an industry that often contends with a regulatory burden and technical debt. Sometimes, due to these intractable challenges, innovation gets stifled because of people and budget constraints. Wood says his organization has circumvented these issues.

“I think Snowflake, and how Rathbones operates, has allowed us to take control of our environment,” he says, referring to the split between the group implementation of the platform and his organization’s deployment. “There's a broad Snowflake environment, and then we've got a ring-fenced environment, with all our data sets, which means the implementation has been built in a way that allows more self-service capability.”

Wood suggests the key to long-term success will be staying on top of Snowflake’s product advances. For example, his team received buy-in for its CoWork and CoCo explorations by explaining how a trial of the agentic technologies could benefit the group.

“There are good ways to try and ensure that you get that prioritization if you need to,” he says.

“When you've got a real-world business case and value proposition that is going to materially benefit the business, versus just being a nice technology idea, the investment ultimately comes through.”

AI isn't replacing cybersecurity roles but it is changing them, and not always for the better, according to new research.

A study from ISC2 found that 65% of cybersecurity professionals who use AI in their roles are spending time deciding when to trust or act on AI-generated recommendations.

Nearly two-thirds (63%) said they often find themselves reviewing and validating AI outputs. While this is basic best practice from a safety perspective, these processes are wasting valuable time.

Regardless, the influx of AI tools within the profession has been welcomed by practitioners, according to the study.

More than half (53%) believe the technology is creating new entry-level opportunities, while 48% said AI makes them feel more optimistic about their long-term career prospects.

“AI is not replacing cybersecurity professionals; it is changing what the profession requires of them,” said ISC2 CEO Scott Beale.

“As AI takes on more repetitive tasks, as well as performing some complex cybersecurity analysis at speed and scale, cybersecurity roles are shifting toward higher-value work, from asking the right questions to validating findings, interpreting outputs and applying human judgment."

Beale noted that the use of AI is changing “how work is distributed across security teams”, meaning investment in areas such as governance, skills development, and validation practices is “essential”.

Too much time fixing problems

While nearly half of cybersecurity professionals reported that AI has reduced workplace stress, 32% said it has made it worse. A key factor here lies in the aforementioned validation and reviewing practices, the study noted.

Those experiencing higher levels of stress were significantly more likely to spend longer periods deciding when to trust AI-generated outputs and recommendations.

When AI-recommended actions lead to incorrect outcomes – which nine out of ten said had happened – half of the participants said their organization holds human decision-makers ultimately accountable.

Put simply, poor AI-related outcomes have a direct impact on wellbeing for cybersecurity practitioners when it’s their neck on the line.

Confusion over accountability and ownership of AI also adds to stress, the study noted. Nearly a quarter (21%) of respondents said accountability of AI-related issues varies depending on the severity.

Over-reliance is a worry

Other top concerns cited by ISC2 included over-reliance on AI, a recurring worry not just for cyber professionals but workers across a range of industries.

As ITPro reported in May, a study from GoTo warned over-reliance on the technology could erode key skills. Similar concerns have been highlighted in software development, particularly among entry-level workers entering the workforce.

62% of respondents identified this as a key concern in the ISC2 study while 56% also highlighted worries about reduced human judgement capabilities when it comes to business-critical decisions.

Foundational cybersecurity skills remain essential, according to ISC2, especially with AI in the mix. Notably, nearly two-thirds (62%) said they don't believe the technology has reduced the need for these skills, compared with just 26% who say it has.

FOLLOW US ON SOCIAL MEDIA

A group of current and former Meta employees is suing the company, claiming that it used AI to target those on medical or family leave for layoffs.

In May, Meta announced it was cutting around 8,000 staff, representing roughly 10% of its workforce. According to the lawsuit, the tech giant used AI systems to identify those set for redundancy.

These systems allegedly took into account keystroke monitoring data, AI token usage

“Meta did not assemble the termination list through the considered judgment of managers who knew the work,” court filings state.

“Instead, Meta used a constellation of internal artificial intelligence systems — including a system referred to internally as ‘Metamate,’ employee-trained ‘second-brain’ agents, keystroke- and activity-monitoring data, AI-token-usage dashboards, and algorithmically assisted performance ranking and calibration — to score, rank and select employees for inclusion on the list.”

Complainants allege this put employees on protected medical or family leave, or those whose output was reduced by a disability, at a disadvantage by penalizing them with reduced performance scores.

"The result was that employees who took protected leaves were disproportionately selected for layoff, based on scoring that not only failed to account for their protected leaves, but in effect penalized the employees for exercising their legal rights to these leaves."

In one example, an employee was laid off while on approved pre-birth pregnancy leave – the day before her waters broke, and just two days before she gave birth.

In another, a manager’s own performance review noted that his demotion followed his return from medical leave; he was put on the list sixteen days into a second medical leave.

The 26 plaintiffs claim that Meta has violated state-protected leave laws, the Family and Medical Leave Act, the Pregnancy Discrimination Act, and the Americans with Disabilities Act.

They're calling for a preliminary court ruling that would stop Meta from finalizing the layoffs while they pursue their ​claims.

“Once these separations are final, the harms are irreversible: employer-subsidized health coverage lost during pregnancy, postpartum recovery, and active medical treatment; time-bound leave rights extinguished; unvested equity forfeited; and immigration consequences triggered,” they said.

Plaintiffs are also seeking financial compensation that could include reinstatement, back pay, lost equity, benefits, and other damages.

AI in HR is a delicate balancing act

The use of AI systems in HR and recruitment has been a long-running point of contention, particularly with regard to the potential for bias or discrimination.

Workday, for example, is facing a class-action lawsuit amidst claims its AI software discriminated against job applicants. As ITPro reported in June, the company hit back at the claims, with a spokesperson insisting its AI solutions “don’t make hiring decisions”.

Ilia Kolochenko, founder of cybersecurity company ImmuniWeb and a lawyer practising in cybersecurity and data protection, noted that over-regulation of the use of AI in HR could cause long-term harm.

"With the current trend to restrict or even entirely ban AI in HR decision-making processes, most organizations will either conceal the use of AI or shift back to non-AI systems especially in those jurisdictions that have no GDPR-like protection against automated decision-making on human subjects," he said.

"Compared to AI-powered HR systems, their non-AI homologues are quite primitive, fail to consider the relevant context and often provide incorrect metrics or data,” Kolochenko added.

“For instance, a poor performance of a delivery truck driver will not be correlated with extreme summer heat or winter snowfalls, sudden family loss or simply a technical issue with the vehicle. Eventually, innocent persons are wrongly punished.”

ITPro approached Meta for comment, but did not receive a response by time of publication.

FOLLOW US ON SOCIAL MEDIA

European small and medium-sized businesses (SMBs) are bullish on generative AI, according to new research from SAS and IDC, and some are setting an example in how to nail adoption.

Analysis conducted by IDC on behalf of the data management firm shows SMBs across the region are executing AI deployments in a far more efficient manner to global counterparts.

In North America, for example, SMBs scored strongest in terms of planning and building, but many aren’t reaching full deployment.

This contrast suggests European organizations are making “stronger progress in operationalizing AI and embedding it into day-to-day operations”.

A key factor in European SMB successes lies in readiness, according to the report. IT leaders across the region are prioritizing areas such as governance before diving head long into adoption projects.

“Organizations treating governance as a foundation rather than an obstacle are often the ones best positioned to execute,” said John Carey, Senior Vice President of Global Channels at SAS.

“The findings suggest European SMBs are taking a more operational approach to AI adoption – focusing not just on experimentation, but on putting the right structures in place to scale AI effectively.”

Notably, Carey suggested that “stronger execution performance” among European SMBs may be due to regulatory considerations such as the EU AI Act.

Simply put, IT leaders at small businesses across the region are accelerating preparation to ensure compliance with the legislation.

It’s perhaps no surprise then that governance, risk, and compliance for AI has emerged as the top priority for SMB leaders, cited by 26%.

Progress on this front is easier said than done, however. The study noted that uncertainty is rife with regard to compliance, risk management, and security.

Nearly one-quarter (24%) identified this as their biggest execution barrier with AI adoption.

Still a long way to go

Despite positive signs for European SMBs, on a global scale the pace of AI adoption is still slow. SAS noted that many are “still struggling to move from experimentation to meaningful business impact”.

Nearly three-quarters (70%) of SMBs worldwide either remain in ‘experimental’ or ‘opportunistic’ stages of AI maturity, for example. Similarly, only 9% have fully embedded AI into daily operations or decision making.

Slow progress in AI maturity is, at least in part, due to traditional data-related problems such as silos and visibility. Nearly half (45%) of SMB leaders said their data remains “scattered across systems with no clear ownership”.

A similar number (46%) said AI tools are often still used in isolated environments across the business, preventing teams from aligning workflows.

Notably, 90% of those in experimental phases report having no formal AI strategy in place, which is impeding progress for IT leaders.

“To actually make something of their AI strategy, SMBs need to move from disconnected pilots to true alignments of their data, people, and resources,” said Daniel-Zoe Jimenez, VP of research at IDC.

“Experimenting with the technology is one thing. Deploying it strategically and sustainably is quite another.”

FOLLOW US ON SOCIAL MEDIA

AI continues to dominate conversations across the technology industry. Every week seems to bring a new model, platform, or product announcement, alongside predictions about how quickly AI will transform the way organizations operate.

For channel partners, the challenge is deciding where to focus. Customers are looking for guidance, but keeping up with every new development isn't a realistic strategy for most businesses.

What many organizations need right now is more practical. For many customers, the challenge is no longer understanding AI's potential but addressing the data, integration, and operational hurdles that stand in the way of adoption. Information is often spread across disconnected systems, while growing attention to AI costs is increasing demand for more disciplined approaches to deployment and governance.

They are trying to work out how AI fits into their existing technology environment, how it can support day-to-day operations, and how to deliver measurable results. That's where many customers are looking for support - and where partners can have the greatest impact.

Why promising AI projects get stuck

Over the last few years, organizations have invested heavily in AI pilots and proof-of-concept projects. Many have demonstrated what is possible, but moving from experimentation into wider business use has proven more difficult.

The obstacles are rarely limited to the AI technology itself. Firms often face challenges integrating AI into existing environments, governance requirements continue to evolve, and teams can struggle to connect AI initiatives to clear business priorities. Even when a pilot is successful, scaling it across different teams, processes, and systems introduces a new set of challenges.

Customers need support identifying where AI can make a meaningful difference and how it fits into existing workflows. Just as importantly, they need help addressing the operational issues that can slow adoption long after the pilot phase has ended.

All roads lead to data

For many organizations, progress with AI is constrained by their ability to make data accessible, trusted and usable across the business. Many have accumulated years of information across different applications, cloud environments and departments.

Making that data accessible, reliable and usable remains one of the biggest barriers to successful AI adoption. Half (49%) of UK organizations lack the internal skills needed to integrate AI with existing analytics and business intelligence systems, according to our research. While 37% say gaps in real-time data integration are a major obstacle.

Without a strong data foundation, organizations will struggle to trust AI outputs or apply them consistently across the business.

For partners, this creates opportunities that extend beyond a single deployment. Helping customers improve data quality, connect systems and establish effective governance addresses challenges that continue long after an AI project goes live.

As organizations roll out AI across more areas of the business, these foundations become even more important. This creates ongoing demand for partners that can help customers manage complex data estates, link disparate systems and put the right policies and processes in place.

Complexity isn’t going away

Alongside data challenges, organizations are also managing increasingly complicated technology environments. Few businesses operate within a single vendor ecosystem. Most are working across multiple cloud providers, applications, analytics platforms and AI tools, many of which have been introduced at different stages.

Customers want the flexibility to adopt new technologies without disrupting existing investments. They also want confidence that systems can work together effectively and that today’s decisions will not create unnecessary constraints in the future.

As AI projects move beyond the pilot stage and into wider deployment, organizations will increasingly need support connecting technologies, improving interoperability and reducing operational complexity. Partners that can provide that expertise are likely to become trusted advisors.

Building services customers will continue to need

As AI adoption progresses, some of the strongest opportunities for partners sit in areas where customers will need ongoing support.

Workflow automation is a good example. Many organizations continue to rely on manual processes that slow decision-making, create inefficiencies or make information harder to access. AI can help address these issues, but success depends on understanding how technology, processes and people work together.

The same applies to services focused on AI readiness, data quality, governance and integration, for which there is already demand. UK businesses and IT leaders believe better data integration and analytics would help demonstrate AI's value to stakeholders, while more than half want greater visibility into how AI models reach decisions, our research found.

Organizations are also becoming more disciplined in how they evaluate AI investments. Beyond model performance, there is growing focus on managing operational costs, reducing unnecessary processing, and ensuring that AI systems are working from relevant, trusted data. Strong data practices not only improve outcomes but can also help organizations use AI more efficiently as deployments scale.

These engagements help customers address immediate challenges while creating a stronger foundation for future AI initiatives. They also lend themselves to repeatable service offerings that can be applied across different customers facing similar problems.

The opportunity is already here

For many channel partners, the biggest AI opportunity lies in helping organizations move beyond experimentation and into everyday use.

Many organizations already have the technology required to move forward. What they often need is a clearer path to adoption across the business. Partners that can provide that guidance will be well positioned to strengthen customer relationships and support the next phase of AI deployment.

Sovereign AI has quickly moved from policy language into boardroom planning. For CIOs, the appeal is easy to understand: more control over where data goes, which systems can access it, and how far an organization depends on overseas providers for increasingly critical AI workloads.

In 2026, the pressure is only growing. Governments want national AI capacity, regulators want clearer accountability, and enterprises want to know whether their data, models, and infrastructure are exposed to legal or operational risks that they cannot easily see.

In that context, regional hosting, sovereign cloud services, and local AI infrastructure can all look like a sensible response.

Gartner has predicted that 35% of countries will be locked into region-specific AI platforms by 2027, up from 5% in 2026, underlining how quickly sovereignty is becoming a question of procurement, platform choice, and long-term flexibility.

But sovereignty is not a single switch that makes an AI deployment safer, simpler, or more compliant, and can actually narrow model choice, add procurement friction, raise costs, complicate integration, and create fresh dependencies.

The real question for IT leaders is what level of control each workload actually needs, what the organization is prepared to give up in return, and whether a sovereign approach reduces risk or simply moves it somewhere less obvious, not whether sovereign AI is “good or bad”.

To help understand these trade-offs, ITPro spoke to Forrester senior analyst Dario Maisto about what enterprises really mean when they talk about “sovereign AI”, where local control can reduce risk, and where it can create fresh complexity.

What enterprises really mean by “sovereign AI”

For many businesses, sovereign AI still starts with a fairly narrow concern: keeping data in the right place, which might mean running inference in a specific region, storing prompts and outputs locally, or making sure sensitive data is processed under a particular legal regime.

While these are important controls, especially for regulated sectors, they are only one part of the sovereignty picture.

Maisto says most enterprise projects have not yet reached the point where organizations are wrestling with sovereign AI at full production scale. “Clients' journeys are not that advanced in AI deployments that they can worry about the depth of sovereign AI at scale for production environments,” he says. “They have mostly data residency concerns.”

Maisto’s insight explains why the term can become slippery. A vendor might describe an AI service as “sovereign” because the data is hosted in-country; a public sector buyer might use the same term to mean local legal accountability or domestic infrastructure; and so on.

Distinctions like this are important because AI does not sit neatly in one place. A single deployment can involve cloud infrastructure, foundation models, APIs, orchestration tools, identity systems, logging, monitoring, and a lot more besides.

Once AI agents – perhaps 2026’s buzziest trend – enter the picture, the stack becomes even harder to contain, as systems start pulling from enterprise data sources and triggering actions across multiple applications.

Why local hosting is not the same as control

The first trap is assuming that sovereignty begins and ends with location. Keeping data in a specific country or region can be important, but local hosting does not automatically decide who controls the service, which laws apply, or who can access the underlying systems.

Maisto is blunt on this point: “Local hosting does not protect workloads and data with regard to sovereignty concerns. This is where organizations overestimate the potential of local hosting and in-region infrastructure to improve their digital sovereignty posture.”

“If the infrastructure and the tools are owned by a third party operating under a different jurisdiction, the sovereignty risk is not mitigated,” he says.

None of these makes regional hosting pointless. It can still help with latency, regulatory alignment, audit requirements, and internal assurance, and can also reduce the number of cross-border transfers and make it easier to show that sensitive workloads are being handled within a defined legal environment.

The problem comes when residency is treated as a substitute for sovereignty. For CIOs, the due diligence needs to go beyond the region on the invoice and into the architecture behind the service.

Sovereignty can limit choice

The next cost to businesses is flexibility. AI buyers are used to a market that moves quickly (perhaps too quickly), with new models, tools, and managed services appearing every few months or even weeks.

A sovereign AI strategy can slow that down, especially if an organization has to use approved regions, certified providers, local infrastructure, or a smaller set of compliant services, alongside additional procurement, legal, or regulatory scrutiny.

While that trade-off may be perfectly reasonable for the right workload, it still needs to be visible. A government department handling citizen data should not treat model access like a marketing team testing copy variations, and a bank, healthcare provider, or critical infrastructure operator may decide that tighter controls matter more than immediate access to the newest frontier model.

But a sovereign deployment may mean fewer models, less access to specialist AI services, slower feature rollouts, or more integration work for tools that would otherwise be available through a major cloud platform.

That risk is already visible in agentic AI, where ITPro has reported that many UK companies are moving ahead with deployments despite gaps in governance and visibility over where data is stored, processed, and accessed.

Maisto argues this is where buyers can underestimate the knock-on effects. “Organizations mostly tackle this theme from a data residency and inference location perspective, although problems like agents' sovereignty could be more important, and the solutions look very immature as of now in the market,” he says.

“Integration is another underestimated aspect of agentic AI as agents are pervasive and have the potential to increase vendor lock-in with non-sovereign vendors.”

As AI becomes more embedded in enterprise systems, sovereignty becomes harder to separate from day-to-day architecture. The key question is how deeply it connects to the business, and whether those connections can still be governed, audited, or replaced.

Skills and procurement can slow everything down

Even when the case for sovereign AI is clear, delivery can be harder than the strategy suggests. An organization still needs people who can assess suppliers, design the architecture, check operational controls, and so on, and these skills are already scarce in mainstream AI projects, before sovereignty requirements narrow the pool further.

Talent may become one of the biggest constraints, according to Maisto. “According to our forecast, [the] tech workforce is the only sovereignty index that is going to decrease in the next five years,” he adds.

“The more specialized the skills, the less the chance to have the right talent with the needed sovereignty requirements at scale. With an outlook to the future, skills from a sovereign pool of talent would be the most relevant component.”

Procurement adds another layer of friction.

Sovereign AI buyers may need to test where data is stored, how support access works, and many other features of the wider operational chain. Most of the time, the process is likely to be slower than buying a standard AI service from an existing cloud marketplace.

For CIOs, this creates a now-familiar tension: The business wants access to AI quickly, while security, legal, and compliance teams need confidence that the deployment will not create exposure later.

Sovereign AI can reduce some risks, but only if the organization has the skills and procurement discipline to define what “sovereign” means before it buys.

Should companies prioritize sovereign AI?

The safest approach is to treat sovereign AI as a design choice with a defined purpose.

Maisto says Forrester uses a “Minimum Viable Sovereignty Model Decision Framework” to help technology leaders weigh “compliance requirements, risk appetite, budget, tech and business requirements, and available sovereign products and services.”

He adds that they should assess “the tradeoff between sovereignty and functionality, looking at the viability, feasibility and desirability of a sovereign solution.”

Any assessment should include the practical value of sovereignty, the compromises it introduces, and the maturity of the products available. “Not all vendors are scoring equally on AI capabilities and even less so when it comes to sovereign AI capabilities,” says Maisto.

For CIOs, the aim should be proportional control. Some workloads will justify strict sovereignty requirements because the data, operational risk, or public accountability demands it; others may be better served by strong governance, clear deployment controls, and encryption.

The danger comes when enterprises buy sovereign AI too vaguely, without defining which risks they are trying to reduce, which freedoms they are willing to give up, and what they will need back if the strategy changes.

Companies using AI should be wary about losing control of their data – and the risk of AI developers making use of that knowledge.

That’s according to Microsoft CEO Satya Nadella, who in a recent blog post warned many AI users are "paying twice" thanks to what he calls the "reverse information paradox”.

This is a process through which AI providers, particularly frontier labs, get access to proprietary data about their own customers. Nadella said the trend represents a huge risk for enterprises, and called for protections akin to patents.

The critique is intriguing as Microsoft itself is an AI company, offering plenty of AI-driven products to customers and cramming AI features into its software. Beyond that, Microsoft was an early funder of OpenAI, though that relationship has weakened of late.

Nadella pointed to economist Kenneth Arrow's information paradox, in which the seller of data or knowledge risks giving it away in order to prove it's worth selling. He noted that AI has created the reverse problem, in which the buyer gives away knowledge in order to use the product purchased.

"You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful," he noted. "The better you want the model to perform, the more of that knowledge you have to feed it!"

"Over time, the information asymmetry becomes increasingly skewed. The seller learns more and more about you as you use what you purchased, while you learn very little about what the seller is learning in return."

The ‘AI exhaust’

This concern isn't just about uploading corporate data to systems, or accidentally including proprietary data in a prompt, though those are both concerns. Instead, Nadella said every interaction with AI inherently includes information that could be useful.

"Models learn from 'exhaust,' the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong," he wrote.

"Every correction is distilled into institutional know-how. It’s the kind of knowledge a competitor could never buy, and the kind that leaks almost imperceptibly: trace by trace, correction by correction, eval by eval."

Ironically, Nadella said it was necessary for the "great innovation" of allowing model developers to use public data – some of them face lawsuits for hoovering up copyrighted data to train their systems.

When it comes to business data, providers should share anything learned with customers. Very few actually do this, however.

"If learning flows in only one direction, economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself," he said. "Therefore, it’s imperative that we distribute the learning infrastructure to every firm so that they can control their own learning loop."

Nadella's tips on how to protect your company from AI

Nadella said the existing setup requires a protection system on par with patents, but also advised companies to set a "trust boundary" by considering a few key aspects – some of which may be easier by simply switching to open source AI.

He suggested companies should control their own "evals", or evaluations of the AI system, because that reveals what looks "good" to the organization, and keep ownership of decisions, feedbacks, and other outputs for their own use.

Beyond that, companies should build capability with AI via their own learning environments where models can learn using real workflows without exposing corporate data.

Improving choice can be achieved by decoupling the orchestration layer from any single model, enabling the ability to switch to another model if one is withdrawn. This can also help organizations cut costs by choosing the right model based on their individual needs.

Lastly, he said to "compound" those four ideas to "create your own continuous learning loop".

"In other words, a company should be able to use a model without giving up the knowledge that makes it unique," he added.

FOLLOW US ON SOCIAL MEDIA

Palo Alto Networks CEO Nikesh Arora has issued a stark warning to workers reluctant to adapt to generative AI: they face a “Darwinian moment”.

Speaking during a recent appearance on the 20VC podcast, Arora suggested a significant portion of workers globally lack necessary AI skills, which raises questions about how leaders can implement change and build workforces capable of meeting future demands.

“The challenge right now is 90% of enterprise employees are not AI savvy. They’re not,” he said.

“They have to learn. I can’t send them to university. There’s no course you can take in school anywhere. They have to be able to learn on their own. I think we’re back to a Darwinian moment where everybody has to figure out who’s really good.”

In some cases, building an AI savvy workforce has prompted some drastic action by executives. Nikesh pointed to Coinbase CEO Brian Armstrong’s approach of penalizing workers who refused to adapt or embrace the technology.

During a podcast appearance last year, Armstrong admitted to firing workers who refused to engage with the technology after paying for enterprise licenses for GitHub Copilot and Cursor.

Other organizations, such as Jack Dorsey’s Block, reportedly introduced AI mandates for workers. Both companies laid off staff this year amidst claims that the technology is changing how they operate.

“You’ve seen people like Brian Armstrong and Jack Dorsey go out and say, ‘I’m going to decimate my organization, and I’m going to start building from scratch’ and they’ve gone to some version of 30, 40% fewer people because they figured out there is no redemption,” Arora said.

As ITPro reported last year, punitive measures to spur AI adoption in the enterprise is a sure fire way to create workforce pushback.

It’s a similar dynamic to the return to office (RTO) mandates that caused widespread upheaval at a range of big tech firms in recent years. Encourage, but don’t push, or else it could come back to bite you.

Palo Alto Networks’ approach leans toward a gradual shift, Arora noted. The company, which has roughly 20,000 staff, has a “natural attrition of 2% a month. The focus now is on replacing these staff with people who are indeed AI-savvy.

Elsewhere, the company actively seeks out AI talent from events such as hackathons, Arora revealed. These are lucrative talent pools that the company can draw from to acquire staff.

“We hire from hackathons,” he said. “Give me 12 months, I’ll have sort of transformed 20, 25% of my team. Give me three years, I’ll hopefully have enough AI-savvy people working at Palo Alto.”

Top-down encouragement

Building an AI-savvy workforce also requires accountability on the part of leadership, Arora said. Delivering change requires leaders across an array of functions to have ambition and actively engage with the technology.

Arora has regular conversations with leadership figures aimed at establishing their progress with AI, allowing them to showcase potential wins.

“You have to make sure your leaders are ambitious. You have to make sure they're competitive. You have to make sure they want to win. You have to make sure that they have a learning mindset,” he said.

“When they watch their peers around them do cool [stuff], they want to show up and do cool [stuff] the next time. So for me, it's getting 14 people together and saying, ‘Hey, Harry, tell me today, what have you done for AI the last three days since I last talked to you in your organization, and whatever motivates you?”

This gives Arora a clear understanding of the direction some leaders are going in with the technology. Moreover, it appears to act as a springboard for motivating others across the company.

“It creates a little bit of Darwinian competition amongst them. It creates this urge to go embrace this new technology, and I think hopefully I get 14 people fully motivated, and then they go do that with the next set of people because I need to transform from the top down, not from the bottom up on this topic.”

FOLLOW US ON SOCIAL MEDIA

Digital twins have been a key tool for organisations to create virtual replicas of operating environments. From transportation to manufacturing and oil & gas, they help simulate performance and prepare for 'what-if scenarios'. But they're evolving.

On this week's episode, Bobby talks to ITPro's news and analysis editor, Ross Kelly, to ask what exactly a digital twin is and how they are evolving since the arrival of generative AI.

The term 'digital twin' is even taking on different meanings; in some cases it's not just a virtual replica of a factory floor, for example. We are now seeing digital twins of people, with Forrester describing them as 'Digital Doubles'. Gartner also predicts that some organisations will begin creating 'Digital Twins of Customers' as a way to simulate how customers might respond to specific scenarios.

Highlights

"We covered this around two years ago on the website: 'AI and digital twins are a match made in heaven'. AI within the context of a digital twin can help improve visualization of real-world dynamics that you're trying to replicate, make sense of the data that you're gathering here, and reduce workloads. For the people that would have traditionally, taken all this on board and been operating and leading the charge on this front."

"I think a great example of AI being used in digital twins at the moment is in Dubai. They launched a digital twin of the entire city, 190-5000 buildings, hundreds of 1000s of infrastructure assets, hundreds of 1000s of public facilities as well. Again, this is part of Dubai's broader digital transformation strategy, but fundamentally this comes down to urban planning and just making it easier, allowing them to create a comprehensive map of everything there."

"Over the last year or two, we've heard chatter about digital doubles or digital twins of workers as well. I think we're going down the rabbit hole here, so to speak. But Forrester's spoken about this quite frequently. The idea of creating a digital double of a worker that's based on all of the information available on that worker internally, their skill sets, what particular unit of the business they're in, and what role they have. There was a fascinating article on the BBC in April this year around Digital Richard, which is essentially an AI, a digital twin of an individual worker. Here, it's not so much a chat bot."

Links

• Are AI digital twins a match made in heaven?
• Does your business need a digital twin?

Open source can't catch a break. As the world of software development marches forward with new tools and capabilities expanding productivity – an evolution driven predominantly by AI – the community endures significant headwinds.

"In the age of AI-driven security threats, protecting customer data has to come first," wrote Bailey Pumfleet, co-founder and CEO of scheduling software maker Cal.com, in April.

After five years as "open source champions", the company announced it would go "closed source". Cal.com is far from the only example, with several organizations and projects including Tailwind, curl, Jazzband, Godot – and even the U.K. National Health Service (NHS) – also struggling in the wake of AI.

Tension has always plagued open source, with concerns around funding, infrastructure, and even expectations management, negating any meaningful green shoots. The threat has even been described as "existential", but the community has nevertheless survived and thrived, celebrating incredible adoption rates in recent years. But things feel different. AI poses an existential threat in many areas — but will it continue to hack away at the beleaguered open source community, or is there a viable path forward for the movement?

Open source is battling AI-centric headwinds

"In the past, exploiting an application required a highly skilled hacker with years of experience and a significant investment of time to find and exploit vulnerabilities," wrote Cal.com's Pumfleet in the blog post. The reality is that humans don’t have the time, attention, or patience to find everything. Today, AI can be pointed at an open source codebase and systematically scan it for vulnerabilities."

Security is just one aspect in an increasingly fraught dynamic. Tailwind, the organization behind a popular CSS framework that web developers incorporate into their projects, announced on 7 January it had slashed its four-person dev team to just one. The reason, CEO Adam Wathan cited, was "the brutal impact AI has had"; web traffic to documentation – which is the only way people can find out about its commercial products – was down 40% from early 2023.

There are also examples of projects struggling to cope with an onslaught of low-quality contributions. Take Godot, the open source game engine that's drowning in AI slop code pull requests (PRs), according to one of the primary maintainers, Rémi Verschelde, who described this as "increasingly draining and demoralizing".

The Jazzband collective, a Python project ecosystem, was sunset this year after ten years of activity due to the "slopocalypse" with floods of PRs and AI-generated spam rendering "Jazzband’s model of open membership and shared push access untenable".

There are plenty of other issues to contend with. They may include the fact that using advanced AI tools may be gated on the need to spend money, and that it's unclear whether you're speaking with a human or an AI agent at any given time.

How open source projects are coping with AI

Amanda Brock, CEO of OpenUK, tells ITPro that the industry is still grappling with AI as a novel force and different projects are handling it in different ways – with some handling it much better than others. "For some projects, they have not been able to manage the scale of it. And for some small companies, they have said they've had to close," she says, but added this isn't a universal experience, with others managing to get a grip on the headwinds.

"I'm skeptical about their business plan," Brock added on Tailwind, "and I could be wrong, and I'm not an authority on this in terms of each company's business and whether they were doing well enough. But I'm a bit skeptical when many are surviving about the few that are closing, and I don't know if that's the most vulnerable or the ones without the right business structures or teams."

However, Brock states the challenges are very real and unlike much else the community has experienced. For example: "When AI comes to Wikipedia, it not only comes at a huge scale by volume of eyes on the site or the pages, it looks at every page, and it does it in an instant. And the overwhelm is huge. And that's the equivalent of what open source projects feel when their repos are scraped; when they suddenly have this inbound volume from AI."

An example of a project that's fared better in wrestling with AI is Homebrew, the package manager. Its project leader Mike McQuaid tells ITPro Homebrew's response was to fight fire with fire. "We have had an increase in low-effort activity, but we respond with low effort," he says. "Our bots automatically close issues, we’ll close without review, we don’t dignify them with a response, et cetera."

McQuaid adds that, despite some notable examples, it's overstated that many projects are going closed source or winding down, and it's not representative of the wider experience. He says projects weather the storm "the same way we always have," with one measure including prioritizing maintainers' time, effort, and enjoyment over contributors, and prioritizing contributors over users. Responding to rudeness is also met with low effort, no effort, or blocking. He also suggested leaning very hard onto guardrails.

Resolving the fate of open source

McQuaid says the notion that AI could lead to the end of open source is "far too dramatic" and that there are plenty of positives to enjoy in a grand trade-off, despite there being an element of entering the unknown. These positive improvements include the way that AI allows some people to go much faster, that there are plenty of free AI tools available, and that many of these can help with code review, fixing bugs and triaging.

"It isn’t making huge negative changes without any positive remediations," he explains. "It’s just resetting people’s expectations of what 'working on open source' might be like for them. Some people won’t enjoy it any more. Many people equally are contributing who didn’t or wouldn’t before."

Despite many positive steps taken in the open source world over the last decade, Brock dwells in a sense of pessimism – but not entirely because of the immediate effects of AI. "I have been very pessimistic in some ways for a long time," she says, "because after lockdown everybody came out seeming to think that 'we'd won'." And what we'd won was, we've got a scale of adoption — but that scale of adoption wasn't matched by a scale of understanding, or funding.

"My worry was that you were going to see it, go full circle and end up back in proprietary because of exactly this kind of thing. I did not imagine AI, but I could see that things might happen, that meant that we were just overwhelmed because we got into a position where people expected an SLA-type delivery and support for free because the code was free."

Building on open source's legacy

"If it hadn't been AI, it would have been something else," Brock reiterates, saying the challenges have been piling up "at a time when people are reeling" one after another in the last few years.

"We've just had a decade since we really saw all this adoption start. We're a small group of experts [and scaling from] the small to the many hasn't worked."

Brock wants to see more gatekeeping, because "good projects have always gatekept". She adds: "There is this sense in the wider world that open source has always been a wild west, and that anybody can contribute – and that's just not the case."

Beyond that, she sees hope in models like the German Sovereign Tech Agency, which brings together dozens of maintainers into standards development. There are calls in the U.K. for a similar system, with a foundation in the model of the Linux Foundation, or China's OpenAtom, that brings together expertise at a national level to work on intellectual property (IP), managing GitHub repos, and other key elements, for public sector open source.

By professionalizing the industry, however, you run the risk of losing projects when things go wrong. CHAOSS, a Linux Foundation project that measures open source health, is an example of a project in dire straits, having just lost its funding. Former director of data science Dawn Foster left her role in March, and the project has returned to the community.

But, Brock says, "it's hard to reignite the volunteers" when they see funded people doing the work. "So it all has to be done in a very careful, measured, and joined-up way across borders," she explains.

When it comes to the future, Brock remains adamant that urgent action is needed or open source will be consigned to history. "We're already seeing eight companies controlling the AI landscape," she says.

"If we allow open source to fail, the only organizations in the world that will be able to manage that are some of those big tech ones that will already have it in place. So I don't think it's in the human and public interest to allow open source to fail."

Accenture and Google Cloud have announced a new partnership aimed at driving AI adoption rates for mid-market firms amid sluggish deployment progress.

As part of the scheme, the duo will provide firms with a suite of six industry-specific agentic AI solutions designed to help push projects from pilot to production.

These solutions cover intelligence and growth, customer experience, cybersecurity, business operations, industry solutions, and workforce enablement.

"The companies that will define the next decade aren't waiting — they're building,” said Rajendra Prasad, technology reinvention engine lead at Accenture.

“Accenture Edge offerings built with Google Cloud technology help mid-market organizations do exactly that. They can deploy solutions in weeks and get measurable outcomes at the scale, budget and speed that they need to grow."

The collaboration will see Google Cloud provide the underpinning infrastructure and AI solutions to support organizations, including the Gemini Enterprise app, Gemini Enterprise Agent platform, and Agentic Data Cloud.

Accenture, meanwhile, is set to provide forward deployed engineers (FDEs) to offer guidance and technical support for participating firms.

Forward deployed engineers are in vogue

The move by the duo comes amidst a sharpened focus on driving enterprise adoption of AI and agents. The use of FDEs, in particular, has taken off as providers look to embed specialists within customer teams to support projects.

Last week, Amazon Web Services (AWS) and Microsoft both announced plans to invest heavily in FDE-related efforts, with the aim of hiring thousands of engineers in the coming years.

As ITPro reported, FDEs could become a key growth area for the industry moving forward, with big tech providers providing more bespoke support for customers.

Kevin Ichhpurani, president of Google Cloud’s global partner ecosystem, said the scheme will provide businesses with the “full power” of the hyperscaler’s product portfolio.

“We’re seeing tremendous demand as mid-market enterprises adopt AI agents to fundamentally reinvent their business workflows," Ichhpurani said.

"The launch of Accenture Edge brings the full power of Google Cloud’s entire portfolio including enterprise AI, our Agentic Data Cloud and AI Threat Defense directly to this sector. Together, we’re enabling mid-market companies to confidently scale AI across their organizations for growth.”

Driving mid-market adoption

According to Accenture, mid-market firms – or those with revenues of between $300 million and $3 billion – have largely been locked out of large-scale AI transformation due to complexity and integration barriers.

A recent study from Klarus highlighted the significant challenges faced by mid-market enterprises on this front. While nearly three-quarters (73%) have deployed AI solutions, around 90% of projects remain stuck or stalled in the pilot stage.

Although up to 94% of leaders feel confident about AI deployment, lack of expertise and governance concerns are causing projects to fail.

Mid-market firms aren’t alone in AI deployment challenges, however. Research from MIT last year found around 90% of all generative AI pilots fail, while similar research from Gartner warned up to 40% of agentic AI adoption programs are expected to fail.

These organizations are in a prime position to fully capitalize on the technology, however. Klarus CTO Alper Gunaydin said mid-market firms have the advantage of agility compared to larger competitors.

“Mid-market companies have a real advantage because they can often move fast, particularly when it comes to technology transformation. We see this agility in action when companies are turning to AI and automation to address productivity and accelerate growth," Gunaydin commented.

"However, our research shows that too many pilots stall because companies lack AI expertise, quality data and effective governance. Making that agility count requires clear priorities, strong foundations and access to senior expertise, all of which will help translate investment into tangible business outcomes and unlock growth without increasing the cost base.”

FOLLOW US ON SOCIAL MEDIA

Stronger powers could be needed to regulate the use of AI in financial services as consumers flock to the technology for advice, according to the Financial Conduct Authority (FCA).

Following an investigation into the technology and its impact on consumer finance activities, the regulator has recommended a formal review into how to regulate the use of general-purpose LLMs across the industry.

The FCA report – known as the Mills Review after its author – noted that companies and consumers are increasingly delegating decisions about finances to AI and AI-powered systems.

A survey by the regulator found that one-in-five adults were already open to the idea of general-purpose AI tools – such as Claude, ChatGPT, or Gemini – making financial decisions on their behalf, in particular around debt advice, pensions, and investments.

While that could help reduce existing challenges such as "advice gaps" and friction around switching services, the FCA warned about risks including fraud and scams, as well as disrupting market structure and competition.

"AI offers a once-in-a-generation chance to close the information asymmetries and frictions that have long left people making poor financial decisions," said FCA executive director and report author Sheldon Mills in the forward to the review.

"The right spur, in retail financial services, is ensuring consumers can make healthy ones; regulation, whether supportive or restrictive, should serve that outcome."

The FCA noted that many consumers may not be aware that they have no formal recourse if something goes wrong, and said some assume they receive equal protection akin to traditional sources of advice.

Emma Banymandhub, CEO of The Payments Association, agreed that the gap in understanding could cause issues and urged caution on the part of consumers.

"Consumers may be increasingly comfortable using AI agents for routine tasks such as weekly shopping, but AI-driven savings and investment decisions present a very different set of challenges," Banymandhub commented.

"While AI can increasingly explain investment concepts and analyse financial information, personalised investment recommendations remain subject to important regulatory constraints."

AI gains traction in financial services

The financial services sector has rapidly emerged as one of the key growth spaces in terms of the use of AI, with a host of major providers across the UK ramping up adoption of the technology.

As ITPro previously reported, high street lenders such as Nationwide, Lloyds Banking Group, and Yorkshire Building Society have all made significant strides on this front over the last six months.

A parliamentary inquiry into the use of AI in financial services, published in January 2026, found that the sector “substantially outpaces” others with regard to AI adoption.

Indeed, around 75% of UK-based financial services firms are now using the technology, with insurers and large banks among the most aggressive in their pursuit of AI adoption.

Running parallel to this, the use of AI by consumers is also surging, according to research from Lloyds Banking Group. The firm’s 2025 Consumer Digital Index found that AI has “rapidly become a financial tool for millions across the UK".

56% of adults – equivalent to around 28 million people – revealed they’d used AI over the preceding 12 months for financial advice. ChatGPT, for example, was referenced as the most popular platform in this regard, used by six-in-10 consumers.

What the Mills Review recommends

With this in mind, the Mills Review recommended that the existing "regulatory perimeter" needs to be secured and adapted to take in AI's impact on retail financial services – and that should come via a review within the next six months into how AI is affecting the market.

"The review should examine how consumers use AI including general purpose LLM tools for personal financial management across savings, investments, pensions, mortgages and debt management, and the implications for competition, innovation and growth," the Mills Review said.

"It should examine the risks of consumer harm, including how far its usage has or will move along the autonomy spectrum, and any impacts on market integrity and the potential for regulatory arbitrage."

Based on that review's guidance, the FCA could tweak regulation as necessary, the report said. Beyond that, the review called for the FCA to monitor for evidence of harm and new consumer models, engage with providers for better insight into changes, and be ready with intervention tools and mechanisms as necessary.

Frontier model monitoring in finance

In the longer term, the Mills Review called for the FCA to keep an eye on frontier model capabilities and AI adoption, considering its impact on its regulatory work, in particular where use of AI falls outside of existing protections and once AI agents start taking more action on behalf of consumers.

The review also called for stronger powers for the FCA so it can look at these issues more widely.

"As is clear in the report, we need to keep pace with a rapidly changing environment and the principles-based, outcomes focussed approach we’ve taken on AI – relying on the Consumer Duty and Senior Managers Regime – has been critical to us doing so," said Ashley Alder, chair of the FCA.

"The recommendations build on work the FCA has been doing – not least allowing firms to test their use of AI with us – and our own use of AI to be a smarter regulator, more efficient and effective."

Banymandhub added that walking that balance between enabling AI and protecting consumers and companies was key.

"The FCA’s Mills Review reinforces that firms should treat agentic AI as an accountability and governance issue now, while providing greater confidence to innovate responsibly as AI adoption accelerates," she said.

"AI has enormous potential for financial services, but realising that potential will depend on strong governance, clear accountability and maintaining consumer trust."

FOLLOW US ON SOCIAL MEDIA

A security firm has spotted what it claims is the first documented case of a ransomware operation run entirely by a large language model.

The Sysdig Threat Research Team said the operator, dubbed JadePuffer, is the first agentic threat actor, describing it as using a known flaw in AI app builder Langflow to gain access to credentials and other key data.

Thereafter, researchers noted it took over a production database and encrypted it for extortion purposes.

"JadePuffer is a warning sign," Michael Clark, Sysdig’s director of threat research, said in a blog post.

"It’s a marker of where extortion tradecraft is heading. An autonomous agent reasoned about its targets, harvested and reused credentials, moved laterally, established persistence, and destroyed a database, narrating its own intent the entire way."

Clark has clarified that though the attack operation was run by an AI, the attack was still organized by a human who set up the infrastructure, found the initial credentials to break in, and chose a victim. But the rest of the attack was managed by the LLM itself.

Alarming attack

Clark noted that JadePuffer's payloads were "self narrating", containing detailed notes that a human wouldn't bother to write but an LLM generates innately about why each step was taken, including prioritization of targets. That narration data could prove useful for security teams looking to defend against such attacks, Clark added.

"The operation also adapted in real time, retrying failed steps within refined parameters," Clark added. "In one sequence, it went from a failed login to a working fix in 31 seconds."

That is perhaps the most alarming aspect of the incident, according to Roey Eliyahu, CEO and co-founder of Salt Security.

"A human attacker who fails an initial payload waits, reassesses, consults, and tries again on a different timeline," Eliyahu said. "An agent that fails a payload corrects and retries in under a minute. That compression of the attack cycle means the window between first detection signal and material damage is now measured in seconds, not hours."

The JadePuffer system even wrote a ransom note complete with Bitcoin address and Proton email contact, Clark noted, though the former appears to be a wallet address frequently used as an example in explainer text online.

That’s either a coincidence, said Clark, or a hallucination by the LLM due to the frequency of that address online, and therefore used in training data.

Clark also noted that the encryption key was generated and essentially random, so the victim would not be able to decrypt — even if they paid the ransom. The victim organization wasn't disclosed.

Warning about the future of security

While JadePuffer is just one incident, it shows that AI can help automate old vulnerabilities and that ransomware no longer requires skills to pull off an attack, according to Sally Vincent, senior threat research engineer at Exabeam.

"While the attack relied on known, older vulnerabilities rather than new exploits, it demonstrates how AI can automate and accelerate the exploitation of unpatched systems," Vincent said. "It also serves as a reminder that patching known vulnerabilities remains important, since AI can make exploiting them faster and more efficient."

While Clark stressed that none of the attack techniques were novel or sophisticated, JadePuffer is interesting because it was strung together into a complete ransomware operation by an AI model.

"The skill floor for running ransomware has dropped to whatever it costs to run an agent, and if that agent is running on stolen credentials through LLMjacking, the cost to an attacker is close to zero," he commented.

That means security professionals should expect to see more like this, as well as a higher volume of attacks, and should proactively protect "exposed application servers, unhardened configuration stores, and internet-facing database admin accounts," Clark added.

Salt Security's Eliyahu added: "The question every security team should be asking after this report is: what is holding credentials in our AI-adjacent infrastructure, and what can those credentials reach?"

FOLLOW US ON SOCIAL MEDIA

Most Managed Service Provider (MSP) service desks already have several layers of automation. From chatbots and self-service portals to AI routing and virtual agents, the tooling is widely used. Some of it resolves tech issues well, ensuring that skilled engineers are not wasting time on password resets and basic access requests.

However, problems usually start when a ticket is passed between multiple teams or arrives without the context needed for troubleshooting.

Even though MSPs are already tracking the usual metrics: ticket volume, average handling time, first-contact resolution, and technician workload, L1 teams can spend huge chunks of the day just trying to gather context. Some support teams jokingly call it the “20 Questions” phase. Who is the user? What device are they using? Has anyone already touched the machine? Did the VPN fail before the update or after it? Is this even the right queue?

A lot of that happens because support teams still lack clear visibility into what users are experiencing on the endpoint. Most MSPs now operate across ticketing, endpoint management, monitoring, remote access, and documentation tools that do not always share information particularly well. So technicians end up jumping between systems trying to piece together context that should already be sitting in front of them.

According to our analysis presented at the Gartner Digital Workplace Summit London recently, saving just one minute of context-switching time across 10,000 monthly tickets equates to 166 hours of recovered support capacity.

That is basically a full-time technician disappearing into tab-switching and context rebuilding, which is not exactly a great use of skilled people.

Escalation is where things really start getting expensive

In reality, many escalations just restart the troubleshooting process from scratch. Support teams have all kinds of names for this: verification tax, re-diagnosis, or ticket ping-pong. In some environments, AI-assisted triage has actually made this harder to spot because tickets arrive looking neatly categorized while still missing critical context.

Nobody fully trusts the notes attached to the ticket, so the next technician repeats the same checks anyway. And honestly, sometimes they have a point. By the third reassignment, half the ticket notes barely make sense anymore.

That gets expensive fast once senior engineers start getting dragged into tickets that never should have reached them. L2 teams can spend 15–30 minutes rechecking information already confirmed earlier in the support chain. L3 engineers may still insist on verifying the root cause themselves before touching anything important. Leaving some tickets basically on a doomed escalation path from the moment the initial diagnosis goes wrong.

If the underlying ticket data and context are weak, automation can just accelerate bad routing decisions. Tickets land in the wrong queues, bounce between teams, or get escalated before anybody has properly understood the issue in the first place. Some queues basically become dumping grounds for badly categorized tickets.

Why AI support struggles with unpredictable problems

Every MSP wants users to handle the simple stuff themselves rather than flooding the queue with password resets and printer tickets.

The trouble is that self-service tends to fall apart pretty quickly once the issue is no longer straightforward. Somebody reports a “slow laptop.” The system suggests a few generic fixes. Nothing changes. The ticket lands in the wrong queue anyway. Then an L1 tech has to start from scratch, figuring out whether the problem is Wi-Fi, memory usage, a bad update, or some background process chewing through the CPU.

Failed self-service gets expensive fast. Failed self-service attempts can push incident-handling costs from roughly $6 to $53 once escalation and repeated troubleshooting are involved.

Part of the problem stems from AI-driven support tools that are trained on historical ticket data and static workflows, rather than real-time device information. After all, real support environments drift all the time. Devices fall out of policy. VPN issues hit one office but not another.

That is where the bad assumptions creep in. Tickets look neatly categorized even when the underlying diagnosis is wrong. By the time the issue escalates, half of the support recommendations are due to AI hallucinations. Skilled technicians recognize that almost immediately. However, the other automations usually do not.

Reducing wasted effort in the service desk

Just throwing more AI at the service desk doesn’t fix the underlying diagnosis problem. In reality, technicians still spend way too much time piecing together context, rerunning the same diagnostics, and trying to figure out how a ticket even ended up in their queue after failing First Correct Assignment.

Smart MSPs are starting to tackle the problem at the source. They’re rebuilding their self-service portals so users can just explain what’s wrong in plain English, rather than forcing them to pick from clunky categories. After all, no one ever reports a “DHCP lease failure”; they say the internet’s down. L1 techs are also gaining real-time visibility into endpoints, rather than relying on patchy ticket notes and whatever the user happens to mention.

When you can instantly see failed updates, VPN drops, crashed services, or a machine grinding to a halt, you waste a lot less time playing detective. But even with those improvements, most MSPs still hit the same wall: all the important information lives in separate systems.

Ticketing, RMM, monitoring tools, and endpoint agents don’t communicate effectively. So every time a ticket escalates, the next person basically has to start from scratch. That’s why some of the better-run MSPs are now creating a single shared record for endpoint health and diagnostic history. When L1, L2, and L3 are all looking at the same up-to-date facts, you cut out the endless re-checking and stop tickets from getting stuck on a doomed escalation path.

At the end of the day, automation and AI are only as good as the foundation they’re built on. If self-service keeps generating dirty tickets, or tickets keep getting escalated too early, or rebuilt every time they move queues, AI is just helping you make the same mistakes faster. The teams getting real results are the ones giving every support tier access to the same live endpoint context; that’s when AI actually starts pulling its weight.

AI-augmented developer roles have increased nearly six-fold over the last five years, as enterprises move from AI experimentation to implementation, according to new research from Randstad Digital.

While there's been an increase of just 28% for traditional developers, the figure for developers with AI expertise has grown by 597%, with nearly one-in-four developer roles now requiring these skillsets.

Analysis of more than 35 million job postings shows technical professionals who acquire specialized credentials are leapfrogging traditional seniority tiers, with AWS Solutions Architect (Pro) and LangGraph/RAG Architect certifications driving estimated salary increases of 54% and 31%, respectively.

"Enterprise AI is no longer a future investment; it is today's operational reality. Yet the biggest barrier to growth is not access to technology, it is access to the right people," said Michael Morris, global head of platform and talent at Randstad Digital.

"Buying AI is easy. Integrating it safely and securely across a complex enterprise is where the true challenge lies. The specialists who can integrate, govern and scale AI inside complex organizations are in critically short supply."

While foundational roles like prompt engineers are still growing at 174%, demand has rapidly escalated up the skills ladder, with AI trainers now the fastest-growing role globally, up 281%.

As ITPro reported in February, demand for AI trainers has skyrocketed over the last 18 months. Researchers said this reflects a broader market pivot toward roles that turn AI’s potential into real support for business growth.

Demand for AI solutions leads is up 226%, process automation specialists up 196%, and AI architects up 152%, Randstad noted.

Finding talent is harder than it looks

While demand for AI-related skills is there, actually filling these positions is far harder than it appears, Randstad noted. Indeed, enterprises are facing acute challenges in sourcing talent.

AI solutions leads are currently the hardest role to fill globally, for example, with time-to-fill timelines hitting 54 days in key markets and vacancy rates of nearly 27% in the US and 18% in the UK.

Despite having talent pools of roughly 100,000 professionals, machine learning engineers face vacancy rates of 8.2% in the US and 11.2% in India. Japan, meanwhile, is facing some of the sharpest shortages globally, with a 46.8% vacancy rate for AI engineers and 25% for generative AI engineers.

All this is reflected by hiring timelines, according to Randstad. While a standard IT role typically takes 38 days to fill, the recruitment window for advanced AI infrastructure roles has expanded to an average of 54 days in the UK and 53 days in the US.

This stretches to a high of 90 days for Process Automation Specialists in Italy.

The result of this is that salary offers are rising sharply, particularly in the US. Across the Atlantic, large language model (LLM) architects have a vacancy rate of 19%, commanding average salaries of $240,000.

Brazil and Argentina have rapidly emerged as a high-growth corridor for specialized AI services, now representing over 15% of global postings combined. In Europe, the UK, Poland, Spain, and Germany show steady demand, with individual national shares between 1.8% and 2.8%, while China accounts for 7.5% of the global job volume.

"AI talent concentrated in the US and India but fast-growing corridors emerging in Brazil, Argentina and beyond, cross-border hiring is becoming a core enterprise strategy," said Morris.

"Organizations that combine global talent sourcing with deliberate investment in upskilling their existing workforce are best placed to close the gap."

FOLLOW US ON SOCIAL MEDIA

Forward deployed engineers (FDEs) could be the key to driving successful AI deployments, as hyperscalers pledge billions to embed specialists within customer teams.

Last week, Microsoft announced plans to invest $2.5 billion in a new division, the Microsoft Frontier Company. The aim here is to deploy upwards of 6,000 AI experts and engineers inside customer organizations.

These specialists will work with enterprises to co-design and build AI systems in a bid to accelerate adoption rates, according to the tech giant.

The announcement came just days after Amazon Web Services (AWS) announced similar plans with a $1 billion investment in its Forward Deployed Engineering segment, once again aimed at bolstering customer AI capabilities.

With the scale of investment – and sheer number of engineers required – for AWS and Microsoft’s push on this front, FDEs are rapidly emerging as the most sought-after technology professionals globally.

Alastair Williamson‑Pound, CTO at Mercator Digital consultancy, told ITPro that the focus on FDEs is an “exceptionally strong signal” that this could become a key growth area for the industry – and a vital pool of talent to drive AI adoption.

“FDEs have become increasingly critical because they sit inside the client, where they can drop barriers, cut through red tape, and get in front of the decision-makers quickly,” he said.

“It’s perhaps why we’re increasingly seeing them considered as a commercial weapon by cloud providers to secure long-term relationships and spend,” he added.

“Another good aspect is that pricing is fixed and based on outcomes rather than time and materials (T&M), which gives customers greater certainty around delivery.”

Williamson‑Pound noted that this aligns with “what companies are after these days”, mainly support in pushing AI adoption projects from pilot to production.

AWS noted that its FDE strategy aims to “compress timelines from months to days” in terms of AI progress, and is “designed so customers are self-sufficient when a deployment ends”.

“Those organizations that are stuck between piloting and deployment can be given that final push by [an] FDE,” he commented.

“The training and documentation they leave behind also tends to give customers that final confidence boost, helping internal teams continue developing and maintaining the system by themselves.”

The rise of forward deployed engineers

Microsoft and AWS are by no means the first big tech providers to put their faith in FDEs. Palantir somewhat pioneered this approach more than a decade ago, and other companies in the AI space have ramped up activities on this front.

As ITPro reported in May, OpenAI launched a new consultancy arm aimed at embedding engineers within customer organizations.

A key distinction here, however, is that the OpenAI Deployment Company is a standalone entity, while both Microsoft and AWS have their own internal divisions dedicated to FDEs.

Crucially, both moves mark the latest attempt by big tech providers to move the needle with regard to enterprise AI adoption, which research shows has been sluggish and at times underwhelming.

Analysis from PwC in January this year found that executives have become increasingly restless when it comes to returns on investment (ROI). Separate research, meanwhile, shows a significant portion of AI projects fail at the first hurdle.

A key factor in adoption project failure often lies in technical capabilities, or lack thereof, according to research from Dynatrace.

With this in mind, providing enterprises with a helping hand throughout the process could provide a much-needed boost for IT leaders.

Initial feedback on Microsoft’s FDE activities show promise on this front, according to Judson Althoff, CEO of Microsoft’s commercial business. In a blog post last week, Althoff claimed that early results have demonstrated “meaningful impact”.

The tech giant embedded engineers within the London Stock Exchange Group (LSEG), for example, helping accelerate AI adoption and providing finance pros with tools to provide answers to complex queries.

Close collaboration between Microsoft’s FDEs and frontline staff at LSEG helped streamline adoption and essentially create a foundation for future improvements, according to Althoff.

“The solution is underpinned by a foundation that is iteratively refined through client feedback and real-time user testing that accelerates each cycle and steadily improves model quality and scope,” he wrote.

Partnerships with Unilever and Novo Nordisk have also delivered marked improvements in terms of AI adoption, Althoff noted.

How FDEs work

FDEs essentially operate like systems integrators (SI), albeit with a few notable distinctions. An SI, for example, is typically an external contractor used to streamline the adoption of individual solutions.

With FDEs, these engineers act as consultants embedded directly within the enterprise. They also offer technical guidance, but it’s the practical support in terms of custom product development that’s a key differentiator. Simply put, they’re building alongside your own engineers.

“Typically, FDEs work in small pods for a fixed period of time, during which they sit within the engineering and technical teams of the client, writing production grade code,” Williamson-Pound explained.

“They leave the customer with a working system in place supported by full documentation, as opposed to an implementation plan.”

JetBrains CEO Kirill Skrygan told ITPro there are a range of benefits to using FDEs for enterprises, and they’re more than just a temporary boost to engineering capabilities.

“FDEs work closely with customers to translate complex organizational challenges into solutions that technology can actually deliver. Doing that well requires more than coding ability," he said.

“It demands architectural expertise, technical leadership, communication skills, and sound business judgment — all qualities that are difficult to automate.”

FOLLOW US ON SOCIAL MEDIA

The rise and fall of tokenmaxxing has been dramatic, not least because of how quickly it happened. From a darling of CEOs wanting to show how AI-forward they are to a balance sheet nightmare, the journey this particular strategy for measuring AI lasted mere months before its star started to fade.

In this episode of the ITPro Podcast, Jane and Bobby are joined by Frank Böhmer, CEO of Ninox, to talk about what prompted the trend in tokenmaxxing, why it fell apart, and what comes next.

Highlights

"[Gamification] leads to a pretty high stress factor for the employees, because they feel really pressed, even if it's not a direct performance metric, if it's just ... something that creates visibility, and so they feel really pressured to utilize LLMs to the highest possible amount."

"If you put this kind of as a success metric and monitor it, then it becomes useless, because people kind of optimize for that ... so even just looking at the output, so the number of tokens does not say a lot, and it's very easy to use models in a way that is not efficient. It can easily, for example, pollute the context window by just cramping in all the information you have."

"When you have to work with within the constraints of a budget or a token budget, then there are clearly ways to optimize the outcome I can create with models, for example, by providing the right information within the context of the model, by making sure that all the data that the company processes is readily available for models to work with by giving the right guidelines for teams to utilize that by giving them access to the right models in the right context."

Footnotes and sources

• Could rising token costs boost interest in on-premises hardware?
• Uber’s eye-watering AI bill shows enterprises are ‘still measuring AI success through consumption rather than outcomes’ – and it's warping our perception of ROI and productivity
• IT leaders are being stung by "unexpected" AI costs
• The AI pricing time bomb
• ‘What we’re seeing right now is just rapid escalation in AI token spend’: Accenture tells staff to stop using AI for unnecessary tasks amid surging costs
• Surging AI costs could exceed developer salaries by 2028 – analysts say context engineering could be the key to optimizing token consumption
• Companies are regretting investing in generative AI so quickly – here’s how to avoid buyer’s remorse
• Goodhart's law - Wikipedia

Most UK business leaders have a “limited understanding” of their AI budgets, and they're struggling to turn AI investment into measurable business outcomes.

KPMG’s latest Global AI Pulse report found that 26% of UK companies are now using AI as part of everyday work, up from 18% in the first quarter of this year.

This increased adoption means that managing usage costs is becoming both more complex and more critical to realizing value.

Three-in-ten UK leaders told the firm that they struggle with usage-based costs, while 42% have only partial visibility into AI spending. One-third, meanwhile, cited “limited understanding” of AI cost structures, including tokens, as a challenge to deploying AI agents.

“AI is moving rapidly into everyday work, but scaling it responsibly brings a new set of challenges," said Dr Leanne Allen, head of AI at KPMG UK.

"Leaders now need to show not just that AI can be deployed, but that it can be trusted, financially controlled and clearly linked to value. Cost visibility is central to that."

Enterprises need a clear path with AI

According to KPMG, companies where CEOs are accountable for AI decisions often report higher confidence in their AI strategy, and are more likely to unlock meaningful business value and stronger returns on investment.

"As organizations use more AI tools and agentic systems, they need to understand how costs build, where value is being created and where governance controls are needed. Without that clarity, it becomes harder to make confident investment decisions or demonstrate returns," said Allen.

“Clear accountability, practical governance, and workforce adoption must move together if businesses are to turn AI momentum into sustained value.”

To help manage AI costs, the survey found organizations are implementing stronger governance controls, including monitoring and spending controls.

More than half (57%) of UK leaders report having AI cost monitoring dashboards, with 61% embedding cost reviews as part of AI approval processes to enable stronger control and decision-making.

Notably, organizations with stronger cost visibility are four-times more likely to report established ROI, at 25% versus 6%.

“AI cost management cannot sit as an afterthought. If businesses want to scale AI responsibly, they need to build financial discipline into the way AI is approved, monitored and governed from the start," said Allen.

"The organizations that can see their AI costs clearly are better placed to understand what is working, what is not and where to keep investing.”

AI costs are spiralling

Surging AI costs have become a recurring pain point for enterprises across a range of industries in recent months. As ITPro reported in June, these surging costs are the result of AI provider shifts toward consumption-based billing combined with increased usage rates.

The ‘tokenmaxxing’ trend, whereby users are encouraged to ramp up their use of AI tools, has already caused serious issues for some major companies, such as Uber.

Uber revealed that it used its entire annual AI budget in just four months after encouraging staff to use the technology, prompting a rethink of how AI is used internally. This included the introduction of a $1,500 monthly cap per employee, which is tracked through an internal dashboard.

Accenture has also urged staff to stop using AI for needless tasks in a bid to tackle mounting costs.

Last week, analysts at Gartner told ITPro that tackling this problem will require a concerted focus on cost optimization practices, including the use of context engineering techniques to maximize the use of the technology.

This call to action by Gartner came after research found AI token costs could exceed developer salaries by 2028.

FOLLOW US ON SOCIAL MEDIA

Anthropic's Fable 5 and Mythos 5 are back after a three-weeks-long restriction saw the security models effectively banned.

Earlier this month, the US government applied export controls to Claude Mythos and Claude Fable 5, banning any foreign users, inside or outside the US, from either model.

As Anthropic had no way of checking the nationality of its users, both models were de facto banned from use. Last week, the restriction was loosened to allow Mythos to be used by US organizations deemed trustworthy.

Now, Anthropic has said the restrictions have been mostly dropped following the introduction of new safeguards.

"Anthropic has agreed to proactively detect and address security risks associated with the models," US Commerce Secretary Howard Lutnick wrote in a letter to the tech company, according to the BBC,

Lutnick later added on social media: "Over the past two weeks, we have worked closely with Anthropic to analyze and approve Fable 5 to ensure alignment across the US Government and strengthen America’s leadership in AI."

Anthropic said it was still working to expand access to Mythos 5 to a wider set of partners in its Project Glasswing program – which gives access to the security model in a managed way – and hoped that would include domestic and international users.

Fable 5 should now be available again for all Claude users, with access reenabled for AWS, Google Cloud, and Microsoft Foundry as quickly as possible.

Why Mythos and Fable were restricted

Fable 5 and Mythos 5 are both security focused models, sharing the same underlying model, Anthropic notes.

As ITPro reported in early June, Fable 5 was built with "strong safeguards to make it safe for general use".

Mythos, meanwhile, has fewer protections in place, and as such was released to a limited number of trusted partners under Project Glasswing for the express purpose of defensive security.

Just days after the release of Fable 5, the government applied strict export controls. That was driven by Amazon researchers finding a way to dodge Fable 5's safeguards to use it to hunt for flaws in software and produce exploits.

Anthropic argued that most of its models could actually uncover the same vulnerability and exploit demonstration, and that the Amazon research didn't make use of any "unique Mythos-level cyber capabilities".

The workaround discovered by Amazon has now been blocked for most instances, with any users attempting similar techniques downgraded to a lesser model. Anthropic admitted that more "benign requests" may be flagged accidentally now, though it hopes to reduce such false positives as the system is improved.

"One particularly important safety mechanism involves classifiers — smaller automated AI systems that, during an interaction, detect when the model is asked to perform a potentially harmful cybersecurity task (or produces potentially harmful outputs)," the company said.

"When this occurs, the classifiers block the model from responding to requests. The ultimate goal of these classifiers is to prevent the model from engaging in uniquely dangerous behaviors."

Beyond the new safety classifiers and other safeguards, Anthropic is working with industry partners including Amazon, Microsoft, and Google to develop a framework for addressing AI model ‘jailbreaking’ techniques, developing a system for judging the severity of such attacks.

Plus, Anthropic said it would work more closely with the US government, offering expanded early access for models that could impact national security and sharing information about jailbreaks or misuse patterns.

The company called for the same rules to apply to its rivals: "These rules should be codified in strong regulation and applied equally across frontier model developers."

Mixed industry reactions

It's no surprise this was sorted quickly. Keven Knight, CEO, of cybersecurity firm, Talion, noted that the US government would not have wanted to risk slowing down AI development, particularly in security.

"Given that China is steadily advancing with its advanced AI model, with a platform that mirrors Mythos being launched by 360 Security Technology last week, the US could not afford to limit access to Anthropic's platforms for much longer," Knight said.

"Restricting access to the platforms would only leave western organisations on the backfoot, and at a time when the AI arms race is really heating up, this would be dangerous."

Though the export ban is now loosened, Andrew Bloster, senior R&D manager at Black Duck, said that won't end the "chilling effect" the incident has had on the use of such models.

"Security leaders globally are now wary to depend on AI as a service models that can be launched with much fanfare and then pulled from the global market," he said, pointing to the rise of sovereign AI.

Bloster added that "technology leaders are more concerned about resilience, data security, and cost, and Fable being permitted back on the world stage might not be enough to earn back global trust."

Though Anthropic was the first company to jump through these hoops, it actually tried to find a good balance between safety and use via Glasswing, said Mike Britton, CIO at Abnormal AI.

"On Mythos, the controlled rollout to a vetted set of US organizations is actually the right model. Limiting access to trusted institutions, with use cases focused on finding and remediating vulnerabilities, is how you derive real security value from a model at this capability level without creating new risks," said Britton.

"The goal should be more secure software — and that's achievable if the access controls stay disciplined over time."

FOLLOW US ON SOCIAL MEDIA

Netgear is making strides in helping businesses make the move into AI-powered network operations and management with the launch of Netgear Insight 10.0.

This next-generation cloud network management platform will be game-changing for small and medium-sized businesses (SMBs) and Managed Service Providers (MSPs), plugging the gap between need and resource, according to the firm.

Insight 10.0 is designed to support organizations through delivery of enterprise-grade AIOps, intelligence, and operational simplicity, helping them to make decisions faster and move their business forward.

One key benefit of the new platform is empowering network administrators to move from being purely reactive to being proactive, according to Luca Marinelli, Netgear’s head of Europe.

In terms of launches, Marinelli said this is a pretty significant one for the company and a key focus during his time with the business so far, following his appointment in October last year.

“The opportunity to have a single pane of glass, which will be managing simultaneously – like with multi-tenancy – several networks will free up time for [MSPs] to maybe start providing other additional services to attract new customers, which is always something very healthy to do in business.”

“[Also] if you perform certain types of activities in a reduced amount of time, your margin can be better. This results in a much more profitable business. So, there are multiple impacts [from] using more sophisticated state-of-the-art tools," Marinelli added.

"Sometimes the feeling is that AI is just a nice slogan that you need to just add in everything you say. [But] I think the important thing is to know what to do with this superpower. If you know what you want to do, it is extremely valuable.”

Outside of product innovation, Netgear has been heavily focused on how it partners to do business. Indeed, in November 2025 it unveiled changes to its partner program in a bid to make it easier to work together and drive joint success.

At launch, Netgear’s president and general manager, Pramod Badjate, said partners were at the center of everything the company does.

“Our big play at Netgear is delivering enterprise-grade products with the simplicity that the SME market needs. That all boils down to a product that is ready for those environments, but also, from a TCO point of view, fits that customer experience really well,” said Jordan Hobday, Netgear’s UK country manager.

“The Netgear Drive Partner Program is three tiers. It's there to really incentivize and reward our partners. There's a commercial benefit to that, but equally it's about how we [can] provide a good experience to our partners, like certifications, self-serve material, all the good things that you would expect.

“A big part for us is simplicity, and how easy it is to work with or do business with Netgear.”

Netgear’s EMEA business is in good health, according to Luca, but he said the company’s turnaround journey was not finished yet. Indeed, one priority for him is not just growing partner numbers for numbers' sake, but really focusing on deeper engagement.

“I think we have, in my opinion, enough partners that we are working with. The game is to keep them interested; I don't necessarily think we need to increase the number of partners broadly. It's going to be a very selective move in both the AV and IT markets to capture the real players that we need to work with, in a given territory, in a given market,” he said.

“We work with a lot of partners, but the number of partners we have this business intimacy with is not at all at the level where it should be. It’s about ways of working with partners; it's about tools that we keep evolving; it's about the value we bring to those partners and the way we spend time. We do spend time [with partners] already, of course, but [it’s about] the way we spend [that] time. It has to evolve.

“It's [about] more quality, more depth, more intimacy, more shifting the way we work with partners, which takes a little while. It doesn't happen overnight.“

Anthropic has launched Claude Sonnet 5 in a move the company said offers users more powerful agentic capabilities.

The new option is the “most agentic Sonnet model yet”, according to Anthropic, capable of building workflow plans for users, accessing browsers and terminals, and operating autonomously for longer periods.

Cost and performance are key talking points here for Anthropic, with the company noting that Sonnet 5 “narrows the gap” with its flagship Opus range in terms of efficiency.

“Claude Sonnet 3.5, 3.6, and 3.7 were the first models that showed impressive skills in coding and tool use. More recently, though, the clearest gains in agentic capabilities have been in our Opus-class models,” the company said in a blog post.

“Its performance is close to that of Opus 4.8, but at lower prices,” Anthropic added. “It’s a substantial improvement over its predecessor, Sonnet 4.6, on important aspects of agentic performance like reasoning, tool use, coding, and knowledge work.”

SWE-bench Pro results show that show that Sonnet 5 recorded a 63.2% score on agentic coding capabilities. That marks a solid increase compared to Sonnet 4.6, which came in at 58.1%.

The model still lags behind Opus 4.8 (69.2%), but clear gains are being made on this front.

Anthropic eyes cost improvements with Claude Sonnet 5

With concerns mounting over token consumption, Anthropic is keen to emphasize that Sonnet is a far more cost-efficient model.

Performance of the model at “different effort levels” – or the intensity and timeframe of tasks – on BrowseComp shows a “strict improvement” over Sonnet 4.6.

Compared to Opus 4.8, meanwhile, Sonnet covers a “much wider range of cost-performance options”.

“It provides substantially improved cost efficiency at medium effort,” the company explained. “Its higher-effort performance can match Opus 4.8 on some tasks.

Users will also be able to adjust effort levels to help them “find the right balance of cost and performance” with both Sonnet 5 and Opus 4.8.

Sonnet 5 safety improvements

On the safety front, Anthropic noted there has been an “overall improvement” with Sonnet 5, which is far more likely to refuse “malicious requests” than 4.8 and is capable of resisting hijack attempts, such as prompt injection.

“The model shows lower rates of hallucination and sycophancy than Sonnet 4.6,” the company noted. “On our automated behavioral audit, which tests a wide range of misaligned behaviors such as cooperation with misuse and deception, Sonnet 5 scored lower (that is, safer) overall.”

It’s worth noting that Anthropic admits Sonnet 5 displays higher rates of “misaligned behavior” compared to Opus 4.8 and Claude Mythos, urging caution by users.

Safety tests for Sonnet 5 come at a critical time for Anthropic, with the company having launched Claude Mythos earlier this year as part of a gated preview with select industry partners.

How to access Sonnet 5

Claude Sonnet 5 is available across all plans from today, according to Anthropic, including Max, Team, and Enterprise users. It will also act as the default model for Free and Pro plans.

For Claude Code and Claude Platform users, Sonnet will come with an “introductory” pricing of $2 per million input tokens, and $10 per million output tokens until 31 August.

Thereafter, Anthropic said prices will rise to $3 and $15 across input and output tokens respectively.

FOLLOW US ON SOCIAL MEDIA

Britain might be in the middle of a heatwave, but cast your mind back a few weeks, and you’ll remember the rain.

In fact, Met Office estimates suggest 2026 has seen well-above-average rainfall, with some areas recording their wettest conditions in over 100 years during January and February, followed by a notably unsettled and wet spell across May and early June.

Such are the fluctuations in weather conditions that sudden rainfall can result in flooding, creating huge challenges for residents and public sector organizations. It’s into this fast-changing environment that an innovative project between Ordnance Survey (OS), Britain’s national mapping service, and technology specialist Snowflake aims to bring data-led clarity.

The two businesses have worked together to create an Intelligent Flood Readiness Model (IFRM), which has identified 1.2 million undefended buildings at risk of flooding in England, many in the most deprived parts of the country.

Harnessing the power of data-led insight

The initiative is opening eyes to data-powered opportunities, according to Tim Chilton, managing geospatial consultant at OS.

“We’re trying to create a broader conversation around the types of things technology can do across government,” he says.

“Several parties have come to us and said, ‘This is interesting, can we talk?’ What I'm trying to do on behalf of OS is to create those conversations and ask, ‘What could you achieve with our data and this kind of technology?’”

Chilton helps OS explore new commercial opportunities from its treasure troves of data. The organization started working with Snowflake in 2024. The organization uses the Snowflake Marketplace to share open data with public-sector organizations and explore commercial avenues. The priority now is ensuring OS embraces other data-led advances.

“We have a platform strategy, which is about putting our data where the customer works,” he says.

“The person who uses our data might work in geospatial, they might be a data scientist, and increasingly, as you'll find with the IFRM, they could be a businessperson who wants to ask a natural language query and be confident the results are correct.”

That approach resonates with Snowflake’s longer-term business strategy. At its recent Summit 2026 event in San Francisco, senior executives outlined the company’s plans to use agentic AI to open data access to IT and non-IT professionals. Chilton says the IFRM is an exemplar of the joined-up initiatives that will help organizations make the most of data in the age of AI.

“The model is about Snowflake and us talking about how AI can supercharge the timelines around having a problem and trying to solve it. Geospatial data really brings that data-led journey to life. A lot of government strategies are based around multi-year plans, but the climate is changing more frequently and regularly in more extreme ways than before,” he says.

“The question we asked was, ‘How could we use technology and bring in the latest data to check that those plans are still right for specific areas and even individual buildings?’ We hoped the model could help us keep on top of the issues and monitor and measure the impact of that plan on the ground.”

Finding the path forward

OS and Snowflake started the initiative in early 2026. After spending a couple of months working on the model, including ensuring the right data sets were being tapped, the organizations recognized that other sources could provide further depth, particularly the National Geographic Database. Chilton says the model will continue to be honed iteratively.

“If someone says, ‘Ah, but you didn't include this data that we've just released,’ we can say, ‘Good point, give us a week, and we'll give you some early results with your data included,’” he says.

“So, the model is a collaborative way of working, where we’re reacting to those changes in the model and the data that it uses.”

The model combines six data streams into a single layer. This layer produces key insights. For example, cross-referencing OS building datasets with the Indices of Deprivation in England identifies where physical vulnerability intersects with social risk. This insight is then layered against other information, such as Environment Agency flood and risk data.

Chilton says three key steps were crucial in the development of the model. First, project staff used the Snowflake CoCo agent to turn text-heavy documents, such as Flood Risk Management Plans, into data that the model could read in natural language.

Second, Snowflake’s semantic layer technology helped the model to understand, with guidance from OS experts, the meaning of attributes in the organization’s data sets. Finally, project staff used the Snowflake CoWork agent to create a natural language interface that allows anyone to ask questions about flood readiness.

“That capability was amazingly well received,” he says. “People can use the interface to query six well-organized and well-formatted data sets. The extraction of insights from documents, the semantic layer, and the CoWork frontend combined to create this powerful capability.”

The right data, in the right hands

"Data is at the heart of making informed decisions. As this project shows, it's rare that one body holds all the relevant data or that this data is in the same format,” says Fawad Qureshi, global field CTO at Snowflake.

“But we're now in an era where technology can bring together the right people and the right data to collaborate on making better-informed decisions."

By combining OS building data with information on flooding, social deprivation, and 3,000 pages of Flood Risk Management Plans, the model allows decision-makers to understand the potential impact of flooding in more detail than ever before. Chilton says these insights will help policymakers better understand the likelihood of flooding.

“There are already flood risk models; that idea is not new,” he says. “What is new is the ability to bring in much more granular data tied to flood risks. Not everyone in government was aware we had that level of detail. This initiative has provided an effective, practical application of some of our knowledge.”

Further work is planned in the wake of the model’s success. OS aims to make it easier for users to ask questions of data. The organization is exploring new technologies, including AI, to make data more accessible, enabling a wider range of users to unlock its value.

“We're actually ahead of the curve in comparison to many of our customers and our partners,” says Chilton.

“We're educating through doing and sharing demonstrators and great case studies like this model, and gradually you can see that cultural shift with people saying, ‘I get it, I know why there's a lot of fuss about AI and data, and we're going to give it a go.”

Other potential data sources include information from the Office for National Statistics, the British Geological Survey, and the National Underground Asset Register. In fact, OS aims to deliver an integrated approach to data within the next 12 months. Chilton returns to his key message for other digital leaders in the public sector considering AI projects.

“Give it a go,” he says. “I've seen the governance, security, data protection, and robustness of these tools. We've gone through the due diligence process, and Snowflake is a system that we trust to hold our data.”

Three-quarters of utility firms are struggling to forecast energy demands due to rapid data center infrastructure expansion.

A new study from Capgemini Research Institute suggests that eight-in-ten utilities predict more extreme and volatile demand patterns. A key factor here is that consumption patterns from AI are less stable and more difficult to model.

The biggest problem is uncertainty, the study noted, with utilities increasingly forced to plan for demand that may never materialize.

Two-thirds of electricity executives refer to “phantom” data center load requests - with 19% of these never materializing.

This forecasting uncertainty creates real problems with capital allocation, Capgemini warned, with utilities having to decide not only how much capacity to invest in, but where and when to prioritize grid modernization investments to support future demand.

Hyperscalers are feeling the pinch

Things are just as bad for hyperscalers, which need to make major infrastructure decisions against a backdrop of uncertain demand forecasts, grid availability, and connection timelines.

The problem is made all the harder by the geographic concentration of data centers, which places significant strain on local grids.

More than half of electricity executives are finding that load concentration is a major obstacle to reliable service, while large clusters of high-density facilities are creating localized bottlenecks that affect system stability and investment planning.

“AI is transforming electricity systems far beyond demand growth. It is exposing structural constraints in grid capacity, planning, and power availability, while making demand more dynamic and harder to predict,” said Claire Gauthier, global head of energy & utilities at Capgemini.

“The challenge is no longer only how much power is needed, but whether it can be delivered reliably, where and when it is required. Utilities have a defining role to play as system orchestrators, leveraging AI-enabled insights to balance grid and customer-owned resources, accelerate deliverable capacity, and enable the next phase of data-center growth.”

What’s driving energy consumption?

The projected increase in electricity consumption is largely down to AI training and inferencing, according to Capgemini.

Training and inference processes currently account for around 25% of data center energy consumption, yet this is projected to reach 60% within the next three to five years.

The consultancy noted that this not only places significant strain on data center infrastructure, but also leaves other IT workloads on the sidelines as enterprises prioritize AI-related activities.

Data centers are increasingly shifting from backup-only approaches toward primary behind-the-meter (BTM) and near-site solutions. Nearly three-in-ten say they already deploy on-site power solutions and 39% plan to add on-site/BTM within the next couple of years.

Operating independently

Looking ahead, the majority (86%) of respondents see the ability to operate independently from the grid as a competitive advantage.

Around three-quarters of utilities and data center executives told researchers they were looking to try and establish a diversified energy mix to ensure reliability and long-term resilience.

Part of this lies in the fact that renewable energy alone cannot provide enough continuous power at scale for large data centers and AI workloads.

Battery energy storage systems (BESS) are emerging as a possible solution, with nuclear Small Modular Reactors (SMRs) expected to take time to deploy.

More than two-thirds of electricity and data center executives globally are looking to natural gas as a near‑term, transitional solution until renewable energy and storage technologies can scale.

"For both energy providers and data-center operators, the key challenge is no longer only scaling capacity, but doing so under uncertainty, speed constraints, and rising system complexity,” said Gauthier.

“Success will depend on the ability to align infrastructure investment, energy sourcing, and AI-enabled operations to manage both the scale and volatility of demand, while balancing reliability, cost, and sustainability.”

FOLLOW US ON SOCIAL MEDIA

Accenture has told some staff to roll back the use of AI for basic tasks due to skyrocketing prices.

According to reports from 404 Media, leaked audio from an internal meeting at the consulting giant highlighted growing concerns about "soaring token spend".

To battle that, leaders at Accenture are pushing back against token use by "non-engineers" who are driving excessive use of the technology on unnecessary tasks, such as converting PDFs to slides, according to internal company data.

The move comes after a report in February revealed the consultancy has introduced new measures to monitor staff AI use, prompting a rise of “tokenmaxxing” by employees to showcase their uptake of the technology.

“What we’re seeing right now is just rapid escalation in AI token spend,” Justice Kwak, Accenture’s agentic AI strategy lead reportedly said in the meeting, according to the report.

He added that Accenture has hit an "inflection point" with unpredictable spend, saying C-level executives are "still asking the question of whether they’re getting value from what we’re spending on in the context of AI."

Kwak reportedly noted that Accenture's problem isn't "niche" but one that will be faced by every AI-using enterprise – and that offers an opportunity to the consultancy to offer services unpicking token use to its own clients.

ITPro approached Accenture for comment, but did not receive a response by time of publication.

Accenture isn’t alone on rising AI costs

Accenture is by no means the first company to introduce measures aimed at limiting excessive AI use in recent months. Indeed, a host of major firms have issued warnings to staff due to rising token costs.

As ITPro reported last week, rising AI costs are the result of a combination of factors, most notably the shift to consumption-based pricing on the part of providers, and the tokenmaxxing trend.

Anthropic, for example, has shifted to usage-based billing and raised costs for some users, though also increased caps for others.

GitHub in April rejigged its pricing model, leading to credits now being consumed based on token usage, a model known as token-metered or token-based pricing. The more AI tokens used, the more a company is billed.

Combined with the rapid acceleration of AI use, some firms have been hit with huge bills. One unnamed company reportedly failed to set usage limits and accidentally spent $500 million on Claude.

This has prompted some businesses to take drastic action to rein in AI spending, including the introduction of token caps and spending limits.

Uber capped employees to $1,500 a month in tokens after blowing through its entire AI budget in just four months, for example. Elsewhere, Walmart limited staff usage of its internal AI agent while Microsoft removed access to Anthropic's Claude Code over costs.

Amazon and Meta have also reportedly removed their AI "leaderboards" designed to encourage AI usage in an effort to end "tokenmaxxing".

Costly even when useful

The leaked meeting notes from Accenture suggest the consultancy is keen to limit AI use for unnecessary tasks, but continue to give engineers plenty of tokens to play with.

Reducing tokens in one part of the business and allocating more to engineers might not be a concrete solution, however. Indeed, Gartner analysts told ITPro last week that AI coding costs could soon surpass developer salaries.

With this in mind, the consultancy said a concerted focus on cost optimization practices will be needed to limit surging costs on this front – yet many are woefully underprepared to introduce such measures.

“Most organizations still lack the maturity and frameworks to effectively measure cost versus business impact," Nitish Tyagi, Senior Principal Analyst at Gartner, told ITPro.

"Software engineering leaders are increasingly concerned as token-driven AI spend becomes harder to justify, with budgets often being depleted earlier than expected."

FOLLOW US ON SOCIAL MEDIA