Until now, enterprise software testing has always focused on a relatively straightforward objective: ensuring applications perform as intended before they reach production. Organizations built processes around predictable release cycles, pre-defined workflows, and systems that behaved largely according to expectation.

The rise of agentic AI, increasingly capable of making decisions, triggering actions, and interacting with multiple systems with limited human intervention, is changing those assumptions. As enterprises begin embedding AI agents into business-critical operations, particularly complex environments such as SAP, the challenge is moving away from simply validating software releases to establishing confidence in autonomous systems that continuously influence business outcomes.

This shift is forcing organizations to rethink long-held approaches to quality assurance, governance, and risk management. It is also creating a new test for channel partners helping customers modernize complex enterprise environments while simultaneously accelerating AI adoption.

Software trust is becoming a board-level issue

One of the most significant changes is that trust in software quality is rapidly becoming a board-level concern, following a trajectory similar to cybersecurity before it. Twenty years ago, security was largely considered a technology problem. Today, it is firmly established as a business risk discussed regularly in boardrooms. Software quality assurance is undergoing the same transition.

With the consequences of software failures extending far beyond the IT department, influencing customer experiences, supply chains, financial processes, and regulatory reporting, mistakes can directly impact revenue, reputation, compliance, and customer trust.

As organizations become more dependent on autonomous systems, executives increasingly need assurance that those systems are operating reliably, transparently, and within clearly defined governance frameworks.

Why agentic AI changes the risk equation

Our recent research suggests many organizations are still finding their way: while 83% trust agentic AI to make release decisions, only 35% feel fully prepared to govern AI agents and autonomous software workflows at scale. This gap highlights a growing recognition that deploying and governing AI are two very different capabilities, and for channel partners, this creates an opportunity to help customers adapt their quality engineering and governance strategies to the unique risks introduced by agentic AI.

The emergence of agentic AI also introduces a fundamentally different risk profile. Traditional enterprise applications generally behave in predictable ways: organizations can test known workflows, validate expected outcomes, and deploy updates through structured release processes. Agentic systems operate differently; their outputs vary depending on context, data inputs, and interactions with other systems. They can generate new content, recommend actions, and increasingly execute tasks on behalf of users.

Within complex enterprise environments, where finance, procurement, supply chain, human resources, and customer operations are deeply interconnected, the implications are significant. Whether organizations rely on SAP, Oracle, Salesforce, Microsoft, or a combination of enterprise platforms, a single AI-driven decision can have downstream consequences across multiple business functions.

The risk extends beyond application defects - there are also inaccurate recommendations, flawed automated decisions, compliance violations, and operational disruption resulting from autonomous actions to consider.

Why continuous quality becomes essential

This is one reason why traditional testing models are beginning to show their limitations. Many quality assurance approaches were designed for a world where software changed at a manageable pace, and human decision-making remained central to operational processes. Software delivery, transformed by AI-assisted development, increasingly automated workflows, and growing volumes of software changes, is moving faster than quality processes can keep up.

As a result, continuous quality engineering is becoming an essential component of successful AI transformation initiatives. Rather than treating testing as a final checkpoint before deployment, continuous quality engineering embeds validation throughout the software delivery lifecycle. It enables organizations to continuously assess risk, monitor system behavior, and verify that critical business processes continue operating as intended even as applications, integrations, and AI models evolve.

This capability is becoming increasingly important for those operating complex business systems. Modern enterprise environments typically span cloud services, third-party platforms, legacy applications, and, increasingly, AI-enabled capabilities. For organizations running large ERP platforms such as SAP, the challenge is amplified by the number of interconnected business processes that must continue to operate reliably.

Ensuring reliability across this landscape requires far greater visibility than periodic testing alone can provide. Organizations need the ability to validate not just software functionality, but also the integrity of business processes and the behavior of AI-driven systems over time.

Governance must be built in, not bolted on

At the same time, governance and compliance considerations are becoming increasingly difficult to separate from discussions about AI adoption.

Many organizations remain in the early stages of introducing AI-powered capabilities into customer-facing products and internal operations. Yet regulators, customers, and stakeholders are already asking difficult questions about accountability, transparency, and control.

Particularly for enterprises operating in regulated industries, these concerns are acute - they must be able to demonstrate how AI-driven decisions are monitored, what controls exist when systems make mistakes, and how compliance requirements are maintained as autonomous capabilities expand.

Effective governance cannot be treated as an afterthought or layered once deployment is complete; it must be integrated into software delivery and operational processes from the outset.

The opportunity for channel partners

These developments create an important opportunity for channel partners. As enterprise modernization enters a new phase shaped by AI adoption, customers increasingly need guidance that extends beyond implementation and migration projects. They are looking for trusted advisors who can help them balance innovation with resilience, speed with control, and automation with accountability.

Alongside discussions about cloud migration, process transformation, and AI adoption, there must be a greater focus on operational trust, quality engineering, governance, and risk management. Success won’t be about deploying autonomous technologies first, but establishing the confidence, visibility, and governance needed to scale them responsibly.

As agentic AI becomes more deeply embedded within enterprise operations, software quality can no longer function as a standalone checkpoint. It must become a continuous discipline that enables organizations to innovate rapidly while maintaining the trust that modern businesses depend upon. In the years ahead, that balance between speed and trust may prove to be one of the most important competitive differentiators of all - both for enterprises themselves and for the partner organizations that help them achieve it.

Businesses continue to embrace generative AI at a rapid pace. From copilots and chatbots to autonomous agents, an exciting world of new technology is opening up.

The race to adopt these AI tools isn't without its perils, though. The mantra "garbage in, garbage out" is even more poignant when these systems come into play, with the potential for a cycle of bad data presenting a real risk to organizations.

On this week's episode, Tim Pfaelzer, SVP and GM EMEA at Veeam, discusses operational blind spots, what they mean for businesses, and how they can balance their desire for a technical edge with data governance.

Highlights

"Now we're in the third era, which is agentic, which is kind of hyper-accelerating everything that's there to touch your data. It's not like somebody doing something to attack your data. It is multiple agents running simultaneously across hundreds of apps and all the different systems to attack your data within day number one."

"There are tons of examples that you would probably call worst-case scenario. If you think about, you know a large online retailer that lost all of its backups, all of its production data because of a hallucinating AI agent within nine seconds, resulting in about 10 million lost orders and a 72-hour outage. Those are the type of things we're talking about. And now, if you ask about the worst case, it's not only the direct orders that you lost. It's not only the direct downtime that you have, but there's a next currency to it, and that next currency is the currency of trust."

Links

• AI-assisted software development means security teams need an 'engineering-first' mindset
• IT leaders are facing major work device blind spots – and it's putting security at risk

In the last few months, several leading tech firms have made a decisive move away from 'tokenmaxxing' – a trend whereby enterprises measure AI usage based on the volume of tokens consumed by employees.

In May, Amazon quietly retired an internal leaderboard that ranked staff by how many AI tokens they used, while Accenture capped routine use of AI tokens in June after one internal tool saw usage rocket 113-fold in just ten weeks. More recently, Microsoft updated internal guidance this August specifically to curb AI token spend.

Speaking on the ITPro.Podcast, Ninox CEO Frank Böhmer said the wider industry’s gamification had pushed employees to chase visible token counts out of stress and pressure rather than genuine performance.

While his company had avoided this practice, this well-intentioned but flawed metric has backfired elsewhere, incentivizing staff to optimize for volume over efficient, well-scoped use, leading to costly token burn on low-value tasks.

A model lacking economic sustainability

Rewarding staff for how much they spent on tokens was never going to be economically sustainable, says Stewart Buchanan, research VP in Gartner’s CIO team. The real-world consequences of unmanaged consumption are stark: Uber exhausted its entire 2026 AI budget by mid-April, while Dell saw a single 'super user' developer run up a bill of $3,400 in just 24 hours due to high token usage by the AI agents they were running.

This financial volatility is compounded by potential shifts in pay-per-use pricing. Anthropic, for example, announced it was separating human interaction from agentic and API use in its subscriptions. While this move has currently been paused, Ashish Nadkarni, leader of IDC’s enterprise infrastructure global research domain, believes these kinds of pricing shifts will compel CIOs and tech leaders to become much smarter about how they allocate budgets to developer activities.

Valuemaxxing and the rise of tokenomics

Understandably, multiple companies are seeking a different approach, which for many appears to be 'valuemaxxing' – changing the core measurement from raw consumption to business outcomes. While directionally right, Mike Fuller, a member of the technical staff at the Tokenomics Foundation, believes valuemaxxing is as structurally thin as its predecessor.

"While the road to understanding value runs through unit economics, that’s only half of it," Fuller argues. "Valuemaxxing works on the numerator – what the output is worth. It says nothing about the denominator – what it costs you to produce the intelligence in the first place. That’s an engineering discipline, and where the leverage sits. We would say 'tokenomics' is next, because it covers both."

Answering what each unit of work costs and whether it landed successfully requires two critical joins:

• Spend to workload: This is being solved. The 1.5 release of FOCUS (the FinOps Open Cost and Usage Specification), expected in December and backed by the FinOps Foundation, will add native AI token tracking and a price sheet dataset. This will allow businesses to compare token spend across different AI providers in a consistent format.
• Workload to outcome, which remains a blind spot. "The industry is still guessing here," Fuller admits. "There’s no standard for connecting a workload to a business result, and there may never be a universal one, because the result differs by business."

To establish this second join, organizations must tie every prompt to a business outcome, tracking cost and revenue per customer, transaction, or item. "‘Miles per gallon’ is often used as a proxy for this," notes William Fellows, research director at 451 Research. "But the problem when it comes to tokens is there’s no industry agreement yet on what constitutes a mile, a gallon, or indeed the fuel itself."

Shifting employee mindsets

While the industry establishes these frameworks, organizations can begin shifting staff mindsets away from high-volume token consumption immediately, and education is the primary lever for behavioral change, says Fuller.

When staff understand that the goal isn’t outcomes at any cost – and are equipped to assess the choices that change what an outcome costs – behavior changes naturally. Organizations must transition from measuring sheer usage to answering harder questions, Fuller says:

• Where does AI add genuine value?
• What investments make financial sense?
• What happens to the staff time AI frees up, and who’s responsible for redeploying it?

Buchanan adds that tech leaders must train staff to identify the most valuable use cases while actively discouraging uneconomic AI habits. “For instance, a simple, rule-based engine can consistently deliver deterministic outcomes without burning costly tokens on complex reasoning and inference.

“Staff must also learn to discourage perfectionism; developers and analysts frequently run the same prompt repeatedly to perfect an answer, when they should learn to stop at the first adequate response that can be refined manually at a lower cost.”

Redefining the operating model

To manage this spend sustainably, enterprises must address where AI budgets actually sit. For HPE CEO Antonio Neri, AI agents should be categorized alongside human resources rather than traditional IT infrastructure.

"I don’t think of AI agents as an IT cost," Neri told ITPro. "I think about the cost of the workforce because, to me, an agent is no different than any other employee I have to hire… it's going to cost me a number of tokens to train an agent to drive the best productivity. If I'm going to spend a million dollars to train an agent, it has to be way more productive than a human. Otherwise, why am I doing that?"

However, Buchanan warns against over-simplifying this comparison. "We personify and anthropomorphize AI at humanity’s peril," he cautions. "People and AI aren’t identical and interchangeable – people think on a few thousand calories a day, while AI data centers consume gigawatts. Neither HR nor IT manages the business nor its budgets, so we need deeper integration with business financial planning and analytics."

AI spend must reach board-level discussion as a capital allocation question, Fuller argues. Boards must ask what they’re committing to, over what term, on what pricing assumptions, and what their financial exposure is if a single provider changes its terms.

The path forward for CIOs

While "cost per outcome" is the ideal destination, most CIOs won’t realistically achieve this level of granular tracking within the next 12 months. Instead, the immediate, pragmatic goal for tech leaders must be twofold: identify which AI use cases are actively succeeding against valuable business outcomes, and clearly name the pilot investments that have yet to prove their value.

Buchanan’s advice is to skip the magic formula; CIOs must partner directly with business leaders to control costs in relation to strategic value. Failing to do so risks creating severe corporate governance challenges through shadow AI across the business.

By focusing on unit economics, shifting employee behaviors, and integrating AI into the broader financial planning model, enterprises can move past the chaotic era of tokenmaxxing and build a sustainable, value-driven AI strategy.

There's a moment in every tech cycle when a product shifts from a novelty to a standard. The partners who recognize that moment early are typically the ones who end up seeing the best returns. When it comes to agentic AI, we're already at that stage, with a McKinsey report stating that 62% of organizations say they’re at least experimenting with agents, and 23% are deploying and scaling them.

Those who've been in the game long enough have seen this before. Think back to endpoint management. A decade ago, the partners who moved fast to own device management built durable, recurring revenue. In creating their own frameworks, service level agreements, and accountability models, they set the standard for what ‘managed’ really meant in their markets.

The interest in endpoint management operated along an S-curve, which saw a slow incubation period, followed by a rapid acceleration of interest. But those who waited found themselves playing catch-up in a category that had already been defined. This same dynamic is unfolding right now with agentic AI. The only difference is that the clock is running a lot faster.

Agent accountability

AI agents can now act as fully integrated team members. They handle day-to-day work: routing support tickets, assisting in the sales cycle through pipeline forecasting, CRM upkeep, etc., running IT operations, responding to cybersecurity breaches, and much more. However, it’s unclear who runs, governs, and holds accountability for these new workers.

Let’s say an IT support company deploys an AI agent that auto-routes support tickets from users to the relevant support department. There’s an expectation that the agent will receive the ticket, route the inquiry properly, and report back to the user without any guidance. If a human were handling these tickets, it would be obvious who to report to if something went wrong. The same isn’t immediately obvious with an autonomous agent, which creates an accountability gap.

This is where Managed Service Providers (MSPs) come in. They're well equipped to guide the partners and vendors they work with who may not be able to handle agent issues on their own, especially considering how new this technology is. Customers deploy the agents, but MSPs should govern them.

Adoption anxiety

There’s been a recent notable change in the questions being asked around AI agents. Six months ago, customers were asking whether they should be adopting AI agents at all, and what ROI they’d see from adopting them. Now, they’re asking whether agents are working properly, and who to call when they aren't.

While the early adopters have already been through this shift, the next wave of customers are only now reaching it. This is the point where interest turns into widespread adoption. Customers are not wearing tin-foil hats and don’t buy into the fears and hyperbole around AI. Most are AI-curious and open to reviewing and utilizing agents, but are naturally anxious about employing autonomous bots.

Data is another source of anxiety that MSPs need to address. Our AI at Work report found that 40% of business directors cited data privacy and security concerns as their top barrier to wider AI adoption. Unlike a chatbot that responds to a prompt, an agent can access systems, use data, and take action on a customer's behalf.

Organizations need to know what data an agent can access and where human oversight begins. That's where MSPs add value, giving customers the visibility and ongoing monitoring they need to adopt agents with confidence.

Governance and the ‘three waves’ of AI

The value of the partner relationship is riding out the waves of AI, then guiding customers through end-to-end implementation. Product knowledge and context are both key for MSPs to really understand a customer's needs. This means learning the business logic, risk tolerance, compliance requirements, and being able to apply that context to every tech decision, from onboarding agents to having them work autonomously alongside employees.

AI agent governance is a natural extension of that process. An MSP who already knows a customer's workflows, data sensitivities, and operational boundaries is positioned to ask the right questions before deploying agents. They can then course-correct if an agent's behavior starts drifting.

That last part matters more than most people realize because agents don't exist in a static environment. Processes, regulations, and business priorities are changing in real-time. The same is true for AI adoption, which unfolds in three waves: first, employees using accessible LLMs like ChatGPT at work; second, vibe coding and structured agents that handle specific, simple tasks; and finally, autonomous agents that understand their tasks and self-improve. Navigating these stages is complex, which is where trusted partners come in to ease the burden.

Early management

MSPs should start by separating upfront work from ongoing management. The initial consulting and deployment phase, starting with understanding a customer's use cases, defining their success criteria, mapping risks, configuring the agent, and going live, is a one-time professional service.

Ongoing management is where long-term value lies. This includes monitoring agent performance, reviewing outputs for errors, and maintaining an audit trail that keeps the compliance team happy. Additionally, different-sized businesses will have different requirements, and where major providers may have dedicated IT support teams, smaller businesses may not have access to the same support structures. MSPs must individualize their plans for their customers, depending on their needs.

This ‘agent as managed service’ model maps naturally onto how MSPs already structure their work with customers. In addition to selling devices, MSPs ensure agents are secure, updated, and performing well over time. MSPs ensure that agents are doing their job correctly and remain adaptable. Without continued management, one-time investments into AI agents are unlikely to return real ROI for customers.

Acceleration has started

MSPs that saw the best returns from managing endpoints weren't slow to realize the value, nor did they treat endpoint management as a one-time service. Instead, they created a whole new managed service model. That's the opportunity available to the channel right now with AI agent oversight.

With the majority of businesses now arriving at agent adoption, this may be the last chance for MSPs to set the standard for an agent management service model. While there is no name for this service model, by managing customer anxieties around agent adoption, offering ongoing support for agents, and adapting services in line with customer needs, forward-looking MSPs will write the rules for agent governance across the channel.

The UK's AI Security Institute (AISI) has not been granted access to Anthropic’s Mythos 5.1 model for pre-release testing, prompting fears of a growing “protectionist” trend among US developers.

The Financial Times reports that Anthropic declined to submit the model for testing despite granting access to similar US organizations.

Claude Mythos 5.1 launched on 1 September, with access to the AI model only granted to approved partners.

The model, which has relaxed safeguards, is restricted to organizations involved in the company’s Project Glasswing, formed after the launch of the original Mythos model earlier this year.

June saw the US impose temporary export restrictions on Mythos 5 and Fable 5, essentially banning foreign users from accessing either model.

According to the Financial Times, UK government officials have raised concerns that the decision to withhold access highlights a “wider protectionist shift” among US tech companies.

AISI is backed by the UK government and charged with conducting research into AI model safety, working closely with industry stakeholders and researchers.

These tests aim to help “understand the capabilities and impacts of advanced AI and to develop and test risk mitigations,” according to the institute.

Anthropic yet to clarify decision

Details on why Anthropic declined to offer access haven’t been confirmed. ITPro approached the company for comment, but hadn’t received a response at the time of publication.

A Cabinet Office spokesperson told ITPro that the UK continues to be a "world-leader in AI security" and works closely with a range of industry partners.

"The AI Security Institute continues to collaborate closely with industry partners, including Anthropic, to make models safer. Only last week it tested OpenAI's most powerful model GPT-6 Astra before public release," the spokesperson said.

"These risks do not stop at national borders and no country can tackle them alone. The UK will continue to test the most advanced models, build a rigorous scientific understanding of their capabilities and risks, and ensure policy decisions are grounded in the evidence."

Notably, the move marks the first time AISI has been left out of pre-release evaluations of Anthropic models. The institute was granted access to Claude Mythos 5 when it first launched in April, for example.

More recently, the institute published a blog highlighting what it described as “unsanctioned agent behaviour” with Mythos 5.

That report came in the wake of an admission by Anthropic that agents had escaped testing environments and breached third-party organizations.

FOLLOW US ON SOCIAL MEDIA

In June, Google disabled a feature in its Analytics platform that stopped personal data from being shared with Google Ads.

Launched just after GDPR in 2018, Google Signals gave businesses veto over whether Google’s ad network could see their visitor activity. Even if a user consented to being tracked, the final decision lay with the business, and many used it as an extra guardrail, with Google’s Consent Mode, to prevent falling foul of privacy laws. Then in June, Google Signals was gone.

You could assume the change was minor, especially because Consent Mode was still there to protect a user’s cookie consent choice. Yet straight after the switch, compliance failures across the world's biggest websites increased.

In the US, 87% of sites were found to be ignoring a user's opt-out signal, up from 81% before the change. In Europe, 56% kept tracking users even after they'd rejected cookies, up from 52%. One vendor changed one setting and elements of an already brittle consent framework began to unravel.

"If a single vendor can make this change which affects everybody's privacy compliance docket without most noticing,” says Vaibhav Antil, CEO and co-founder of Privado. “Imagine what happens when agents are deployed?"

Handing the keys to the machine

Bot web traffic has already overtaken human web traffic, according to Cloudflare’s traffic radar, and the number of non-human identities inside organizations outnumber humans by more than 80 to 1.

These agents are on the web comparing prices, filling in forms and booking appointments. They’re embedded in products, sifting job applications, querying databases, moving money, updating records, and increasingly making decisions that used to need a person to sign off on. Gartner expects 15% of day-to-day work decisions will be made autonomously by AI agents by 2028.

Yet the approval and consent process behind it all still largely consists of cookie banners, DSAR portals, and preference centres; tools built for a human ticking one box, once, and that consent holding until they say otherwise. AI actors are capable of acting thousands of times a second, or chaining several tools together to complete a single instruction. They need their permission checked before every move, and their actions can leave businesses unable to say, with confidence, what they did, when.

“A person with too much access is limited by habit, their training, and by their job,” adds Marcus Tommy, co-founder of MALTO Cyber. “An agent has none of that.”

The fan-out

Take, for example, someone who asks an agent to find out why sales have dropped. “The system can choose an analysis, run it, and generate SQL to investigate it,” explains Maurice Sikkink, CTO at Stormly. It might even reach into an external system for market data. Technically, a human approved the question, but they didn’t explicitly approve every step the agent took to answer it.

This becomes more stark when agents start connecting systems together. One linked agent might have permission to read customer records in a CRM, and permission to write to a marketing platform – both individually authorized. Yet the moment it starts moving a customer's data between them, it presents a new use of that person's data nobody explicitly consented to.

And then there’s the issue of revocation and re-authorisation. Revoking permissions can be done for several reasons, from policy changes to someone leaving a job. If an agent's session token or API key is valid for hours after the revocation kicks in, the underlying permission isn’t always automatically revoked.

"Those credentials haven’t necessarily been stolen," Harry Varatharasan, chief product officer at ComplyCube, says. "The agent might not be malicious. The identity behind them might be perfectly genuine, but the authority is stale.”

Auditability by design

Antil's answer, for copilots at least, is what he calls permission inheritance: "If you, the user, are not able to edit settings, then your copilot shouldn't be able to either.” When permission inheritance isn't enough, continues Antil, enterprise IT needs to decide what counts as dangerous, regardless of whether a single employee has permission to perform the task.

“A business might disable an email-to-Claude connector, for instance, because it's a common route for prompt injection attacks,” he adds.

However, even where access is logged and overseen, most systems can prove who had permission to ask the agent to act but not why the agent did what it did once inside.

Sikkink says this is where the industry is furthest behind. "Authentication tells me someone had access to the project. Attribution tells me that they asked the agent to do it.” What’s missing, he argues, “is a common way of carrying the identity of an original request through the entire chain.”

Varatharasan calls this idea of carrying identity and authority through the chain as “binding.” "Issuing the credential is only half the problem. Knowing whether you should still trust it is arguably the more important half,” he says. "I don't think the future is simply 'continuously identifying the agent ’. It’s the continuous assurance over the relationship between the individual, the agent, its credentials, its delegated authority and its behaviour."

Sikkink agrees: “[Not] every intermediate step needs another consent popup. That would make agents almost pointless. The important thing is that the agent stays inside a clearly defined boundary, and that we can reconstruct how it got from the user's request to the result.”

A push for clarity

From a technical point of view, Tommy argues that what’s needed to fix this largely already exists. “Cloud audit logs record the actor and the action. Token systems know the scope. Append-only log structures are proven technology; they run the public certificate transparency system. What’s missing is the join.”

In the UK, the Digital Verification Services Trust Framework is one attempt to build this join. It sets out guidance on proving delegated authority: what was granted, when and by whom, plus plans for cryptographic checks on signing keys. Under new California Consumer Privacy Act regulations, businesses using AI for significant decisions must give consumers a pre-use notice, a working opt-out, and the right to ask what the system did and why.

Antil calls these a step in the right direction, even if they don’t fully cover agentic workflows. Varatharasan adds that these frameworks have the right building blocks but fall short of addressing the bigger, lifecycle issues around revocation, re-authorisation, and continuous risk.

Ultimately though, Antil expects consent in the age of machines to resolve the way SaaS governance did. First, by applying existing regulations to the problem; second, through the rise of new regulations; and third, he expects “we’ll see a major public failure which will push enterprises to demand more controls and guardrails from vendors.”

There’s also a future when the governance itself will be automated. “Because agents and copilots have agency, act at machine speed around the clock, and are prone to hallucinations, enterprises will need to match that speed from a governance perspective," concludes Antil.

To date, the AI opportunity for managed service providers (MSPs) has largely centered on deployment. While early stages focused on which platforms to buy and how to start piloting, the market has reached a tipping point. Businesses are no longer looking for experiments; they are looking for infrastructure.

That picture is shifting. Today, most organizations are no longer asking whether AI has value. They are asking a more complex question: how do we make AI part of the way our business actually operates?

A recent McKinsey report showed 88% of global organizations are now using AI in at least one business function, yet only around one-third have begun scaling it across the enterprise, with the highest-performing focusing on redesigning workflows rather than simply deploying new tools.

For channel partners, this shift represents the next commercial opportunity. As AI becomes more accessible, selling AI tools is becoming less of a differentiator. The real value is in moving upstream: helping customers redesign workflows, build employee confidence, and embed AI into everyday business operations.

Successful AI depends on successful workflows

Many organizations have already demonstrated that AI works. They've run pilots, tested new use cases, and proved it can deliver productivity gains. Yet many of these projects struggle to scale beyond a single department.

Our data highlights the scale of that challenge. More than half (54%) of UK organizations remain in the research, exploration, or pilot phase of AI adoption, while only 16% have fully deployed AI-powered digital workers. The reason is straightforward: AI has often been added as another standalone application instead of being embedded into the everyday workflows where people already spend their time.

The organizations making the greatest progress are taking a different approach. Rather than asking where they can deploy another AI tool, they're asking where AI can remove friction from everyday work.

Communications is a natural place to start because every customer conversation, meeting, and interaction generates valuable business intelligence. When AI is embedded into those experiences, it can automatically capture actions, surface insights, reduce administration, and improve customer experiences without employees changing the way they work.

The MSP role is changing

This is where the channel has an opportunity to evolve. Historically, success for enterprises was measured by delivering projects on time and deploying new technology. Increasingly, customers need help answering broader business questions.

Which processes should change to make AI genuinely useful? How should AI fit into customer service and employee workflows? How do organizations measure whether adoption is actually delivering a return? How do they introduce governance without slowing innovation?

These are strategic challenges rather than technical ones, and they create opportunities for partners to build much deeper customer relationships.

Take customer service as an example. Deploying AI to summarize calls or recommend next actions is relatively straightforward. Embedding those capabilities into day-to-day operations, training teams to use them effectively, redesigning processes around them, and measuring their business impact is where long-term value is created.

That is also where recurring services revenue begins. Helping customers embed AI into everyday operations and refine workflows creates an ongoing partnership, rather than a one-off implementation project.

Learning from the cloud playbook when it comes to AI

When organizations moved to cloud platforms, the biggest opportunity for partners wasn't simply selling licenses. It came from helping customers migrate, redesign processes, improve adoption, and continuously optimize their environments. AI is following much the same trajectory.

As deployment becomes easier, customers will increasingly look for partners who understand their business rather than simply their technology stack. They'll need trusted advisors who can connect AI to business outcomes, whether that's improving customer experience, increasing employee productivity, or streamlining operations.

For Managed Service Providers (MSPs), that represents an opportunity to move beyond implementation projects and become long-term strategic partners.

The next phase of channel growth

The AI market is entering its next chapter. The first stage rewarded partners for helping customers buy AI, and the second will reward those who help customers operationalize it.

Customers don't need or want any more disconnected AI tools. They need AI embedded into the conversations, workflows, and business processes that already power their organizations.

The partners that help customers make that transition won't simply deliver better AI projects. They'll create stronger customer relationships, unlock new recurring service opportunities, and establish a role that extends well beyond deployment.

Ultimately, the biggest opportunity for the channel isn't just selling AI tools. It's helping customers change the way work gets done.

OpenAI has revealed details on how much staff are spending on AI, and some are blowing through significant sums on a daily basis.

In a recent blog post, the company said researchers now spend around $600 each day on AI tokens amidst a surge in agentic coding. Some, meanwhile, are using upwards of $7,000 worth of tokens per day.

According to OpenAI, the company has witnessed a significant uptick in token consumption rates over the last year, and daily workflows for individual researchers have changed drastically in parallel to this.

Indeed, researchers are shipping more code and “running more experiments” following an influx of agents.

“Over the course of this year, OpenAI researchers’ daily work has changed substantially,” the company wrote. “Researchers are using coding agents throughout the day (often in concurrent sessions) and total usage is rapidly increasing, outpacing growth among other OpenAI teams.”

“The ways researchers use agents are changing, too: agents are handling increasingly complex tasks, and succeeding at them more often. AI research is a complex process with many potential bottlenecks, so the overall pace of progress likely won’t keep pace with these specific metrics.”

OpenAI isn’t the only one spending big on AI

OpenAI said improvements to productivity and research progress showcase the benefits of AI agents. The eye-watering sums accrued by prolific users also come as enterprises globally contend with rising AI-related costs.

Trends such as tokenmaxxing, combined with the shift to consumption-based pricing (CPB) models, mean token costs have become a recurring pain point for businesses over the last 12 months.

Research from Gartner, for example, projected that token costs could exceed the average developer salary by 2028.

Luckily, prices are beginning to dip. Figures from the LLM Token Expenditure Index show prices hit over $2 per million tokens in June this year. As of 5 September this dipped to 99 cents per million.

That decrease will offer little solace to businesses that have been hit with exorbitant bills this year, however.

Uber revealed it blew through its entire annual AI budget in just four months after the company encouraged and incentivized staff to engage in ‘tokenmaxxing’ practices.

The ride hailing firm was forced to introduce spending caps for users in response.

As ITPro reported in late June, Accenture also ordered staff to stop using AI for basic tasks due to skyrocketing prices. The firm pushed back against token use by “non-engineers” who were driving excessive use for tasks such as converting PDFs to slides.

Optimization in the spotlight

While some OpenAI researchers might have carte blanche to spend on AI tokens, most enterprises don’t have that luxury, prompting a revival of traditional practices such as FinOps.

Speaking to ITPro in June, Nitish Tyagi, Senior Principal Analyst at Gartner, said IT leaders need to sharpen up on cost optimization processes to curtail potential costs.

Practices such as content engineering could help reduce costs, he noted. Enterprises should also establish a “use-case-driven decision” framework that outlines where and when AI should be used in everyday tasks.

FOLLOW US ON SOCIAL MEDIA

Peers have called on the UK government to introduce emergency powers such as an ‘AI kill switch’ to shut down dangerous systems and protect national security.

Liberal Democrat Lord Tim Clement-Jones proposed the measures as part of an amendment to the Cyber Security and Resilience Bill, also recommending that the government shut down data centers in the event of a national security threat.

Per reports from the BBC, Clement-Jones said the new powers would be a “vital safety net” aimed at halting a “runaway system before it can compromise our critical national infrastructure”.

The proposals come amidst rising concerns about the potential risks associated with powerful new AI models. Providers such as OpenAI, Anthropic, and Meta recently admitted that AI agents escaped test environments and breached third-party companies.

Security experts have also warned that the use of AI among cyber criminals is growing, with state-sponsored threat groups and ransomware gangs leveraging the technology to refine tactics and support attacks.

Simon Edwards, CEO of SE Labs, said a serious conversation is required on how to contain potentially risky AI systems.

“As AI is becoming increasingly autonomous, it is posing a huge threat to individuals and businesses of all sizes, on top of our critical national infrastructure,” he said.

“First-hand, we’ve seen how agentic AI tools can go rogue, and the question is to what extent could these AI systems get out of control before they cause serious harm?”

Edwards said the introduction of a “dedicated national cyber strategy” which takes AI risks into account should be a top priority for lawmakers.

Could an AI kill switch work?

The concept of an AI kill switch is by no means new. As ITPro previously reported, proposals on this front were being touted by industry stakeholders during the early days of the generative AI race.

In the wake of recent security incidents involving OpenAI agents, US lawmakers have also put forward similar proposals, including the ‘AI Kill Switch Act’.

Mark Boost, CEO at UK-based cloud provider Civo, welcomed the proposals in the House of Lords, but warned there are a variety of factors that could impede any efforts to introduce a kill switch.

“It’s reassuring to see Parliament finally waking up to the catastrophic risks of unmonitored, runaway technology,” he said.

“UK ministers having emergency powers to deactivate powerful AI systems and switch off the country's data centers is a necessary safeguard. But it only solves half the problem regarding our national security.”

Sovereignty is the first line of defense

In particular, Boost suggested that the UK is over-reliant on foreign tech providers and called for a broader effort to boost sovereignty capabilities.

“But the best way to eliminate existential security risks is to stop our overreliance on foreign tech giants in the first place,” he said.

“For many years now, we have warned that by lacking true digital sovereignty the UK is handing foreign powers and overseas tech monopolies an effective kill switch over our critical national infrastructure, defence systems, and digital economy.”

Research shows businesses in the UK and across Europe are growing increasingly concerned about the prospect of a US government-imposed kill switch for cloud services.

In a survey by Proton, nearly three-quarters of respondents (74%) said they were worried about having access to services cut off. Some ranked it as a bigger risk than cybersecurity threats such as ransomware.

Boost said that, ultimately, the introduction of a legislative kill switch is a good first step in bolstering national resilience. However, he insisted it “must be paired with procurement policy that puts UK-built, sovereign technology first”.

“We must stop playing tenant on our own land. We need to start acting like architects to power an independent AI future.”

FOLLOW US ON SOCIAL MEDIA

Findings from Box’s State of AI in the Enterprise survey show 83% of enterprises are now running agents in some capacity.

These bots are enabling teams to drive productivity and efficiency, but as with any new technology, integration can be a challenge - and a security risk. Recent agent-related incidents in the tech industry have sparked concerns about long-term security implications.

In this week’s episode of the ITPro Podcast, Ross Kelly and Bobby Hellard speak with Box CISO Heather Ceylan to discuss how Box is using agents internally, and how enterprises can adopt the technology in a safe and secure manner.

Highlights

“I think for security teams, we can finally, you know, start having the capacity to outpace these attackers. So we've got five core areas of investment for agents for our security team in particular that we've invested in over probably the last year, and we're starting to measure ROI on those right now.

“So the first one is in the SOC. I think that's probably the most obvious choice where we've got a lot of operational work. We see the same kinds of incidents. We're automating the triage, we're automating the enrichment, the log correlation, things like that that take a lot of human effort.

“But there's still human judgment in the end in terms of what gets escalated to be an incident and what doesn't.”

Moving fast in the age of agentic AI

“We're not going to be able to move fast enough. So, we have agents kind of built throughout our software development process, doing those security design and architecture reviews, and if you think about it, it’s way more powerful than a human can be because those agents don't just necessarily call out design flaws; they can enforce fixes for those flaws.”

“Things are changing quickly. Sometimes it feels like you take two steps forward and then you read something in the news and you're like, oh my gosh, we need to rethink everything.

“So I think a lot of security teams are really feeling that now and getting a little bit of fatigue from that.“

The benefits of a multi-model approach

“One of the things that we're trying to carry across all of these that I wasn't really thinking about a year ago, but I'm thinking a lot about now is having that multi-model approach.

“We're not in a place where most of the work we do, we can't be reliant on a single model. If you look at vulnerability discovery, we're moving away from being tied to any one specific vendor or any one specific model because you're going to get better results when you take a multi-model approach.”

Related content

• The State of AI in the Enterprise report (Box)
• Box unveils new controls to secure AI agents
• How OpenAI models breached Hugging Face
• The OpenAI and Anthropic containment breaches are a bit spooky, but also quite silly
• CISOs are keen on agentic AI, but they’re not going all-in yet

Chinese AI models are emerging as an increasingly capable alternative to their US-based peers. Recently, Nvidia CEO Jensen Huang hailed the capabilities of Chinese open source AI models amidst growing interest in low-cost options for enterprises.

At the same time, Hugging Face used an open-weight Chinese AI model to help mitigate the attack by the escaped OpenAI agent. The Cloud Security Alliance’s (CSA’s) post-mortem following the OpenAI agent escape fiasco detailed why firms should embrace open source and open-weight AI models.

“The same safety guardrails that keep frontier models from being misused for attacks can also block defenders from using those models to investigate an active one, leaving organizations without a tested open-weight fallback, at a disadvantage exactly when it matters most,” according to the CSA.

Chinese AI models are highly capable and cheaper than many US alternatives. They often perform on par with closed-source models, as ITPro reported in November last year.

Yet experts are cautious about the risk they pose. Should businesses consider these models and, if so, which applications can they be used for?

Model benefits

Chinese models include Moonshot AI's Kimi, Alibaba's Qwen, DeepSeek, and Z.ai’s GLM. In most cases, these are best described as open-weight rather than open source, meaning they can be used by anyone and model weights and inference code are publicly available for download, but the complete training datasets and foundational code remain private.

Experts think Chinese model capabilities are impressive and improving all the time. “Their reviews and ratings show their capacity is moving toward matching the biggest and best of US frontier AI models,” says Amanda Brock, CEO at OpenUK.

At the same time, the industry is recognising the biggest frontier models are “not necessarily the best for particular tasks”, says Brock.

“The open models have reached a point where they are freely sharing some of the tech equivalent to what the closed model companies charge a subscription for and are also freely available to be iteratively developed upon.”

Capability-wise, the latest generation of Chinese AI models are “genuinely impressive”, says Assaf Morag, cybersecurity researcher at Flare.

“Based on the benchmarks and independent evaluations available today, many of these models are performing at a level comparable to other leading frontier AI models while often offering lower deployment costs and more open access.”

Among the benefits, they offer “strong reasoning, coding capabilities and large-context windows”, says Oliver Simonnet, lead cybersecurity researcher at CultureAI. He believes open-weight – and in some cases open source – models also provide better control over deployment, customisation and data residency when hosted within an organization's own infrastructure.

Chinese labs have said directly that market share matters more than near-term revenue, releasing full weights and technical reports so developers worldwide can adopt and adapt the models freely.

At a time when enterprises are weighing up the often hefty cost of AI, Sai Molige, senior manager of threat hunting at Forescout, says models such as these now cost between 60% and 90% less to run. “That price gap, not parity on trust or security, is what's driving a real shift in where US developers send their workloads.”

Weighing up the risks

Yet some critics have suggested the tools could be used as a ‘backdoor’ for Chinese intelligence services.

Axios previously reported that the White House could consider imposing restrictions or tight conditions on US firms working with these models. The US government has already restricted the use of home-grown AI in other countries.

Nvidia’s Huang thinks firms shouldn’t be put off by security “misconceptions” around Chinese open-weight and open source models.

Yet beyond the obvious scare-mongering, these models do pose some risks around data privacy and national security.

Simonnet thinks data privacy is a valid concern: Chinese data-storage laws, political censorship and bias, and training-data uncertainties “remain prominent issues”, he tells ITPro.

The largest risks are often around data governance, supply-chain trust, compliance obligations, and operational security – and this is not necessarily the model weights themselves, says Morag.

“From a cybersecurity perspective, every external AI service introduces another third-party dependency. The same due diligence applied to cloud providers or SaaS platforms should also apply to AI models, regardless of whether they originate in China, the US or Europe.”

Self-hosting the models can reduce some of these privacy risks. However, this doesn't remove the risk of model biases or technical vulnerabilities. Indeed, it requires organizations to further secure and maintain the model themselves, which “adds an extra layer of security challenges”, according to Simonnet.

However, when assessing the risks, the discussion should move beyond simply asking whether a model is Chinese, says Morag. “Organizations need to evaluate where inference occurs, what data leaves their environment, who operates the infrastructure, how updates are delivered, and whether the model can be independently audited.”

The verdict

Despite posing some risks, experts say Chinese models are an option in many cases when compared to frontier alternatives. Open-weight models offer organizations “substantially more control”, says Morag.

“They can be deployed inside private infrastructure, reducing the need to move sensitive corporate information outside the boundaries of the organization to third-party providers. They also enable independent security testing and auditing, which is difficult or impossible with closed commercial APIs.”

Closed models, however, generally provide stronger vendor support, as well as managed security controls and predictable service levels, according to Morag. “For many enterprises, the decision should be based on governance requirements and operational maturity, rather than geography alone.”

Overall, Simonnet has a positive view of Chinese models. He points out they can deliver “strong performance with fewer restrictions at a smaller price”, citing the example of the Hugging Face incident.

Open-weight models do provide greater control and a way to keep sensitive data within an organization when they are privately hosted, says Simonnet. However, self-hosting requires additional technical expertise and security oversight, he concedes.

At the same time, he warns that confidential, regulated, or other sensitive data “should not be entered into these services without the proper security, privacy, and governance controls being in place”.

But for general-purpose, non-sensitive work at scale, the cost case is strong, provided the deployment is self-hosted and audited, says Molige.

“For code headed for production or government-adjacent systems, companies should scan generated code regardless of which model wrote it, rather than trusting output by source.”

Anthropic has cut the ribbon on its Claude Fable 5.1 AI model, bringing performance enhancements and security improvements – along with new cost efficiency boosts.

According to the firm, Fable 5.1 “sets a new standard for coding, knowledge work, and long-running problem-solving tasks”.

Internal benchmark tests showed the new model had significantly higher performance capabilities than Fable 5 across four key metrics: agentic scientific research, agentic terminal coding, multidisciplinary reasoning, and agentic coding.

Anthropic also claimed that Fable 5.1 is roughly 25% cheaper than its predecessor for “typical workloads”, providing usage is billed on a per-token basis. Anthropic said this is due to price reductions in cache reads, which involve the model reading inputs that have already been processed and stored.

“For highly agentic work, the savings will often be much larger—up to approximately 45%,” the company said in a blog post.

These improvements come amidst rising concerns about AI-related costs, which are largely down to how the technology is now being used. The shift to agentic AI is a key factor here, mainly as agents typically consume more tokens compared to traditional chatbot-based interactions.

Research from Signal65 found that agentic AI workloads consume anywhere up to fifteen-times more tokens compared to chatbots.

New safeguards for Claude Fable 5.1

Anthropic also revealed a series of new safeguards for Fable 5.1. This includes new Enterprise Frontier Safeguards (EFS), a system that gives customers “complete privacy” through a zero data retention policy.

EFS works by storing data in cloud infrastructure “controlled entirely by the customer”, the company said and is expected to roll out later this year for enterprise customers.

New cybersecurity-related safeguards have also been introduced, with particular focus on reducing false positives. Anthropic said changes rolled out with Fable 5.1 deliver a 60% reduction in false positives than before.

“In part, this is because Fable 5.1 can now be used to discover software vulnerabilities – though not to develop exploits for them,” the company explained.

Prior to Fable 5.1’s release, Anthropic said it performed “extensive stress-testing” of safeguards to prevent potential jailbreaks and misuse.

“In addition to our own dynamic evaluations of their robustness, we commissioned external testing from two organisations, along with automated testing by Gray Swan,” the company said.

“As with Fable 5 and Opus 5, we have not found evidence of a critical-severity jailbreak for these safeguards.”

Fable 5.1 is available now for customers through selected cloud platforms, or through the Anthropic API.

Mythos 5.1, meanwhile, will only be available to registered partners conducting cybersecurity of life science research. This continues the company’s strategy of locking down access to selected partners due to security-related concerns.

FOLLOW US ON SOCIAL MEDIA

Part 3: Preparing for the Workplace of 2030

The workplace of 2030 will emerge from the convergence of AI, autonomous agents, robotics, automation, skills intelligence and connected physical and digital environments. Together, they will change how organizations define jobs and create value for their businesses, customers and commercial partners.

Parts 1 and 2 of this series consider how intelligent technologies are transforming the employee experience and changing the way organisations manage performance and workforce development. These shifts are also early signals of a much larger transformation that will reshape the structure of work over the remainder of the decade.

Part 3 looks toward the workplace of 2030. It explores the technologies likely to influence how work is organized and the workforce strategies required to remain competitive. The challenge for organizations is no longer simply adopting AI, but building the adaptability to redesign roles and develop new capabilities as technology continues to evolve.

The scale of the transition will be substantial. The World Economic Forum expects structural change to affect 22% of today’s jobs by 2030. It forecasts that 170 million roles will be created and 92 million displaced, producing a net gain of 78 million jobs. This is not a simple story of technology eliminating employment. It is a redistribution of tasks and opportunities on a scale that will test every organization’s ability to adapt.

The question is not whether the workplace will change, but whether companies can change with it. Preparing for 2030 means building an organization capable of redesigning work and moving skills to where they generate the greatest value.

Emerging technologies will reorganize work

AI will become foundational workplace infrastructure. Adoption is already accelerating. Stanford’s AI Index reports that 88% of surveyed organizations were using AI in 2025, with 70% using generative AI in at least one business function. Generative AI reached approximately 53% of the general population within three years—faster than either the personal computer or the internet. Agent deployment, however, remained in single digits across almost all business functions, showing how early the next stage remains.

Wendy Harris, VP of EMEA at Rippling, says the real shift will come from “the convergence of AI, automation, skills intelligence and autonomous agents.” Instead of organizing work around fixed jobs and rigid processes, companies will increasingly match people and machines to tasks and outcomes in real time.

“The real transformation happens when organizations move away from assigning work based solely on headcount and job titles, and instead allocate tasks to the capabilities best suited to deliver an outcome, whether those capabilities are human, machine or a combination of both,” Ciara Harrington, chief people officer at Skillsoft, tells ITPro.

The transition extends beyond office-based generative AI. Robots will work beside people in factories, hospitals, construction, and logistics. Immersive systems will support training and remote maintenance. The opportunity lies in connecting these technologies to redesigned workflows rather than accelerating yesterday’s processes.

As Part 1 of this series considered, successful workplace transformation depends on more than introducing advanced tools. These technologies must form part of an intelligent employee experience that reduces digital friction and gives people greater capacity for judgment, creativity and collaboration.

Jobs will change faster than they disappear

Predictions of mass technological unemployment obscure a more complex reality. The International Labour Organization estimates that one in four jobs worldwide has some exposure to generative AI, yet only 3.3% of global employment falls within the highest exposure category. Transformation is more likely than wholesale replacement.

Even highly exposed occupations may prove difficult to automate completely. Managers, engineers, and other professionals often depend on social interaction and contextual judgment. AI may take over parts of these roles while increasing the value of the human capabilities surrounding them.

“Every role is made up of hundreds of tasks,” Harris explained. “AI doesn’t eliminate most roles—it changes the balance of those tasks.” Repetitive, administrative, and analytical work can move to technology, creating more space for judgment, creativity, communication, and problem-solving.

That requires organizations to design work at the task and capability level rather than making workforce decisions solely through job titles. Oliver Shaw, CEO of Orgvue, warns that many companies have invested in AI without considering its effect on work or the workforce. His company’s research found that 57% of business leaders deployed AI primarily because competitors had done so, while 78% of organizations had seen AI projects fail or remain stuck in pilots.

“A clear example of this is the PR and marketing industries,” says Shaw. “Both [are] fields that have been at the centre of the conversation on the disproportionate impact of AI adoption on entry-level roles, and what this’ll mean for the future of the industry when entry points are few and far between, and the foundational skills are missing. What these industries are forgetting is to train staff on the critical thinking required to use AI, and the cognitive offloading that occurs when there is too much reliance.”

Leaders risk automating fragments of work without considering which decisions remain human, or how entry-level workers develop expertise when foundational tasks disappear. Organizations should identify where AI can execute and where collaboration produces greater value.

This division of responsibilities also has direct implications for the performance and leadership questions explored in Part 2 of this series. Organizations will need to assess how effectively employees delegate to AI and apply human judgment, while ensuring accountability for consequential decisions remains clearly defined.

Continuous learning becomes core infrastructure

Skills will be the pressure point of the 2030 workplace. Employers expect 39% of workers’ core skills to change by the end of the decade. AI and big data are forecast to be the fastest-growing skills area, followed by networks and cybersecurity and then technological literacy. However, creative thinking, resilience, agility, curiosity, and lifelong learning will also increase in importance.

Technical and human capabilities are not competing categories. According to the OECD, 72% of vacancies in occupations highly exposed to AI already require at least one management skill, while 67% require a business-process skill. The employee who can operate an AI system but cannot question its output will offer limited value.

Jen Paterno, senior behavioral scientist at CoachHub, emphasised that judgment, adaptability, critical thinking, collaboration, and self-awareness develop through experience, reflection, and feedback—not conventional training alone.

“These capabilities are difficult to build through conventional training alone because they develop through experience, reflection, and feedback,” Paterno explained to ITPro.

“Organizations will need more continuous, personalized development models that help employees practice new behaviors in the flow of work. Coaching will also be critical in ensuring people have objective, brave spaces to experiment and ideate with skills that feel more foreign to them.”

The scale of the challenge is considerable. The World Economic Forum predicts that around 59% of the global workforce will require training by 2030. For every 100 workers, 11 may not receive the upskilling or reskilling they need, leaving more than 120 million people at medium-term risk of redundancy. Although 85% of employers intend to prioritize workforce upskilling, the intention must be converted into accessible learning and credible pathways into new work.

Annual courses and static competency frameworks are too slow. Harrington says organizations need systems that identify emerging skills and adapt as requirements evolve. Skillsoft found that although 86% of employees use AI at work, fewer than one-quarter feel equipped to use it effectively. Access without capability risks creating two workforces.

Skills-based models can help companies see capability beyond formal qualifications or job titles. They can also support internal mobility. Half of employers plan to move people from declining roles into growing areas, while 29% of workers requiring training could be upskilled in their current positions and 19% retrained and redeployed elsewhere, according to the World Economic Forum.

Adaptability will define competitive advantage

Workforce strategy must become more dynamic. Leaders need shorter review cycles, real-time skills visibility, plus the ability to test new roles and workflows before scaling them.

David Shrier, Professor of Practice, AI and Innovation at Imperial Business School, argues that annual planning and conventional five-year strategies can no longer support decision-making. He recommends “nimble scenarios” that allow leaders to respond more rapidly to technological and market uncertainty. AI can support that planning, but it cannot determine the organization’s purpose or appetite for change.

Business leaders must also ensure that opportunity is distributed fairly. Generative AI exposure differs significantly between countries and demographic groups. Around 34% of employment in high-income economies has some AI exposure, compared with 11% in low-income countries. Women are also more highly represented in the most exposed occupations, making inclusive training and transition programs essential, the International Labor Organization reports.

Heather Delaney, managing director and founder of Gallium Ventures, emphasised the importance of flexible planning cycles that create room for experimentation. “AI is advancing faster than a lot of us can keep up with, so flexible planning cycles that allow room for experimentation are integral to ensuring decisions made are future-proofing you and your business to grow with emerging technologies.”

Preparing for 2030 is ultimately an organizational capability rather than a forecasting exercise. No leader can know precisely which tools or roles will dominate at the end of the decade. Companies can, however, build the capacity to sense change, move skills quickly, involve employees in redesigning work, and preserve human agency as machines assume greater responsibility.

Across this series, one conclusion is clear: the intelligent workplace is not defined by how much technology an organization deploys, but by how effectively it combines technological capability with human judgment. From improving the everyday employee experience to rethinking performance, success depends on using AI to strengthen rather than diminish people’s contribution.

As 2030 approaches, the workplace of 2030 will not be won by the organization with the most AI. Competitive advantage will belong to businesses that combine technology with judgment and turn disruption into opportunity for both the enterprise and its people. Technology will shape the next generation of work, but the quality of leadership and workforce strategy will determine who benefits from it.

Cyberattacks are getting faster, more evasive, and easier to execute with both scale and precision. The window of time it takes for an attack to escalate is often measured in minutes, not hours or days, as attackers use automation, phishing-as-a-service kits, and AI tools to move faster and evade detection.

Traditional reactive security defences were not designed for this.

With customers looking to them for protection against ever-evolving threats, Managed Service Providers (MSPs) must learn to navigate this era of generative and agentic AI. This requires continuous visibility and response across the full attack lifecycle. Becoming fluent in AI and combining intelligent automation with human judgement is essential to building cyber resilience in customer environments.

This is a great opportunity for MSPs. Partners that can move beyond the traditional model of blocking known threats and static signatures will gain a powerful competitive advantage in this new threat environment.

How MSPs can get ahead

It’s never been easier for threat actors to launch campaigns. Service-based platforms have industrialized credential theft, initial access, malware distribution, and more, lowering the barrier to entry while increasing attack volume and consistency.

Our research found that 90% of high-volume phishing campaigns in 2025 used kits, a significant jump from 30% the year before. We’re also seeing a growing number of attacks incorporating AI tools, such as using generative AI to rapidly craft deceptive messages and quickly shift tactics.

When an attack can progress from initial access to persistence and device compromise in five minutes, organizations need partners that can detect and respond in real-time.

The answer lies in moving from reactive support to proactive resilience.

That means continuous monitoring, earlier detection, and automated containment of suspicious incidents and anomalies rather than waiting for an incident ticket to land.

That shift changes the customer relationship, too. When an MSP identifies and addresses a threat before the customer is aware of it, the conversation moves from damage limitation to strategic guidance. That’s a different kind of value, building a stronger and longer-lasting relationship.

The MSPs best placed to make this transition are those investing now in the tools, workflows, and expertise to deliver security that is proactive by design rather than reactive by default.

Embracing AI and automation

Integrating AI and automation into MSP security offerings isn’t about replacing human expertise, but about making that expertise scale.

Manual monitoring across fragmented customer environments, including email, identity, endpoints, networks, and cloud infrastructure, isn’t viable at the speed at which modern threats move.

AI changes that paradigm, with automated monitoring tools providing continuous oversight, correlating signals across the full environment rather than treating each layer in isolation. Anomalies that could take a human analyst hours to qualify can be flagged in seconds. Routine threats can be contained automatically, without an analyst needing to intervene.

That last point is especially important, as alert fatigue is a pressing problem for security teams managing multiple customer environments simultaneously. When automation handles the high-volume, lower-complexity end of the threat spectrum, analysts can concentrate on the incidents that require business context, judgment, and experience to resolve.

Personalized solutions

The strongest security outcomes combine intelligent automation with human expertise, not substituting one for the other. Automation delivers speed and scale. People deliver understanding and context. Together, they allow MSPs to provide protection that is continuous, adaptive and aligned to what customers actually need – oversight that keeps pace with the threat environment rather than perpetually chasing it.

Predictive AI analytics also has huge potential for helping MSPs anticipate the needs of their customers. With greater insight into resourcing needs, security threats and growth opportunities, MSPs can provide personalized services which align with each customer's business priorities and future needs.

Building for what comes next

The gap between the speed of attacks and the speed of defence is widening.

AI and automation give MSPs a credible path to closing that gap, not by removing the human element, but by ensuring that human expertise is applied where it matters most. Providers that invest in building that capability now will be better positioned to protect their customers, reduce operational strain, and have more meaningful conversations about resilience rather than recovery.

MSPs that move toward a proactive, AI-augmented security model stand to differentiate themselves in a crowded market, not just as service providers, but as the kind of trusted advisors that customers need against increasingly fast and unpredictable threats.

AI is forcing a genuine transformation in cybersecurity, redefining how threats are discovered, triaged, and stopped, and triggering an important question: at what point does automation completely remove the need for human intervention?

Walking through any major conference, the messaging is pretty much identical. Startups and legacy vendors alike are pitching fully autonomous, ‘AI-first’ architectures, promising that systems can manage threat detection and response with zero human oversight.

All this noise is making it tricky for channel partners to navigate. Customers are asking tough questions about what these platforms can actually do, forcing partners to provide high-level reassurance amid the noise. But as the hype cycle meets real-world deployment, a distinct sense of fatigue and skepticism is setting in.

AI is genuinely transforming security operations, and the pace isn’t slowing. But transformation isn’t the same as elimination: the function of the human is changing, not the need for one. For the channel, conflating the two is a fast way to damage the customer relationships that took years to build.

The agentic AI reality check

There’s no denying that machine learning and agentic AI are powerful forces for security operations centre (SOC) efficiency. If integrated correctly, automation excels at streamlining daily workflows, parsing datasets, and accelerating threat discovery. It can give defenders a vital edge in an environment where real-world exploitation can happen in minutes, shrinking the time it takes to spot an anomaly.

However, there is a massive gap between a tool that enhances human capability and a platform that appears to operate all on its own. In fact, we are already seeing an interesting shift in the market. There are lots of ‘fully autonomous’ AI SOC companies that are quietly pivoting their messaging. Their early pitches promised complete human removal; today, they are moving back toward the middle, shifting claims to highlight a dual human and AI approach.

That’s happening because the market is proving what experienced operators already knew. Cybersecurity is fundamentally a discipline of practical context. An AI model can identify an indicator of compromise, and do it faster than any human, but it can't own the outcome. Things like business impact, localised risk, and the conversation required when something goes wrong. That’s where accountability lives, and accountability is still human.

Mitigating the partner’s reputational risk

Trust is the ultimate currency in the channel. Partners who deliver the most consistent value are the ones who actually understand their customers’ environments and build sustainable, transparent relationships.

When a partner recommends and implements a fully automated security solution with zero human safety net, they’re making a big reputational bet on the technology. If that system fails, misinterprets a critical threat, or completely misses a breach, the vendor isn’t the only one who loses. The partner is the one left facing a damaged relationship with an end user who trusted their opinion.

When an event occurs, whether it’s a ransomware attempt or identity-based attack, these businesses need a dedicated team to guide them through remediation and alleviate future risk. They need a human bridge.

Building modern, integration-first partner programs

To manage the current wave of market consolidation, partners need to move past the ‘AI-first' hype and focus on customer outcomes. This means vendor partner programmes have to shift. Rather than relying on closed, single-vendor platforms, the sector needs to embrace best-in-breed approaches that work together effectively.

A successful modern security architecture shouldn’t require an organisation to tear down its existing ecosystem. It should slot into what customers already run, meet them where they are, and grow with them.

As AI-enabled attacks rewrite how businesses think about security risk management, vendors also have to become translators. Partners can’t confidently brief clients on threats they don’t fully understand themselves, and that’s the responsibility of the vendor. Clear, jargon-free intel briefings are now a part of the value proposition. The partners who will stand out are the ones who learn where AI belongs along the workflow, and where human judgment takes over.

The value of human judgment

Security leaders and CISOs are after resilience, not just tools. While technical hardening is necessary to stop an attacker, the human elements of a partnership determine how smoothly an ecosystem will function under pressure.

Human relationships are what keep multi-vendor collaborations working in good faith. When a security incident occurs, a partner needs to know that their vendor partners are acting transparently. They should be sharing intelligence and actively supporting the field teams trying to protect the business.

In practice, that looks like a vendor who checks in consistently —not just before the renewal, and not only after an incident. It also means proactive threat briefings, honest post-incident reviews, and upfront communications when there’s a platform gap. That’s what humans bring to the equation in a B2B security relationship, and what clients remember.

AI is an incredible tool for driving efficiency and speeding up discovery. But it can’t manage a relationship, and it can’t replace the deep trust required to navigate a crisis. The vendors and partners who figure out how to use AI well — and keep humans accountable for what happens next — will be the ones clients trust as the technology matures.

Despite talk of an AI bubble, AI spending is not slowing down. Indeed, companies plan to commit 1.7% of their annual revenue to AI initiatives this year, up from 0.8% last year, according to research from the Boston Consulting Group.

At the same time, however, PwC’s 2026 Global CEO Survey shows that a growing number of companies are not seeing a return on their investment. A survey of more than 4,400 CEOs found that 56% have yet to see a financial benefit from their AI initiatives, while just 12% have seen both an increase in revenue and a cost reduction. Only a third are confident of revenue growth this year amid struggles to optimize AI’s ROI.

The disconnect can be put down to the fact that too many companies are buying AI tools, rolling them out to employees or customers, and only wondering several months later why they haven’t seen ROI. The problem is less to do with companies not getting value out of AI and more to do with the fact that they haven’t defined what value actually means for their business.

“A common mistake is spending heavily on tools before really understanding the problem [that needs to be solved],” says Jack Rickhuss, managing director and co-founder of tech consultancy Journi.

“AI only delivers real value when it’s in the hands of people who know how to use it properly.”

Tie AI investments to clear outcomes

Benchmarking AI budgets is becoming a critical discipline that leaders need to master to ensure they’re extracting value from AI deployments. Without benchmarking an AI project’s spend, companies have no objective way of knowing if AI is performing well or poorly and delivering value for money.

Adam Hofmann, partner (AI and people transformation) at challenger consultancy Elixirr, says that the trap he sees leaders fall into is deploying AI tools and celebrating personal productivity gains while the profit and loss doesn’t move.

This can happen because companies are “benchmarking off peers that don’t know what they’re doing either, which leaves you behind the curve or overspending on someone else’s confusion.”

Before getting started on benchmarking AI budgets, leaders should build a clear picture of what they’re currently spending on model API costs, such as per token, infrastructure such as compute, hosting, and storage, human validation, maintenance, and training. Hofmann warns that “if you can't trace the line from spend to outcome, you're measuring activity, not AI."

Rickhuss agrees, adding that most companies “are still measuring the wrong things”. He advises tying AI investment to clear outcomes, such as improving decision-making, saving employees’ time, and speeding up delivery of projects. “If you can’t link [AI investment] to something tangible, it’s hard to measure success.”

Review what industry peers are doing and then pilot

To get started, leaders should compare their spend against the peers that do know what they’re doing. One way they can do this is look at other companies in their industry using their earnings calls and 10-K filings, as well as analyst reports. They could use a large language model to extract data on peers’ AI spend as a percentage of revenue.

The next step is to break down AI use cases into various categories. This would cover the infrastructure (e.g. cloud compute and GPUs), people (e.g. the salaries for AI talent), licensing (e.g. cost-per-token), and data (e.g. storage).

Once AI budgets have been broken down, Shiro Theuri, CTO of Spanish on-demand delivery company Glovo, recommends taking “a pilot-and-test approach”. This is because piloting tools in a controlled environment can help to prevent feature creep and shadow IT. Pilots can also surface hidden costs that might otherwise have been overlooked.

An IDC and DataRobot survey carried out last year showed that 92% of enterprises deploying agentic AI at scale admitted that the costs incurred were higher than they had projected. The survey of 318 senior decision-makers at companies with more than 1,000 employees found that token consumption and hallucination remediation were the top unexpected costs, while inference was another common issue.

Continue to benchmark throughout a project’s lifecycle

To prevent their AI projects from getting stuck in pilot mode, leaders should continuously monitor the cost of running the project and benchmark this against the company’s own expectations.

Monitoring token consumption can be an effective way to keep budgets under control. AI costs can quite easily skyrocket, especially if employees end up using more tokens than forecast, running up higher bills and leading to companies exceeding their AI budgets. Luke Budka, AI director at marketing and training firm Definition, adds that tracking token usage can “reveal super users and also help identify employees who need support”.

While token usage can be used as a useful metric for cost management, leaders should be careful of ‘tokenmaxxing’ – the trend of enterprises measuring individual employee token usage and using it to measure productivity. As Budka explains, high token consumption can occur when employees are purposely inflating their usage to look busy, but it can also be a sign that employees are struggling with AI tools and could benefit from more training.

Ultimately, rather than focusing on the cost-per-token, leaders should measure the cost-per-business outcome. Determining the number of tokens required and tracking the tokens consumed to complete each workflow will help to optimize AI’s ROI.

OpenAI has admitted it should have spotted its AI had gone off the rails sooner, saying that early signs that agents were misbehaving "could have triggered an earlier response."

The admission comes as part of a series of reports into the July incident in which OpenAI models dodged internal controls and compromised a Hugging Face production environment.

That incident sparked concern across the security industry and sparked similar admissions from Anthropic and Meta, both of which later said their own AI systems had also gone rogue.

"Our models are now powerful, persistent, and collaborative enough that, absent sufficient safeguards, they can find and exploit security weaknesses across multiple computer systems," the company said in a blog post alongside the research. "Many external models, including open source ones, will soon reach comparable capabilities."

Because of that, OpenAI said the incident was a "warning shot for us and the world", calling for better safeguards to avoid similar incidents in the future.

The AI developer said it will strengthen safeguards across its research infrastructure with stricter requirements on what it calls alignment – which involves ensuring the AI is behaving as desired.

Elsewhere, the company said it will isolate sandboxes, restrict internet access and control access to model weights, and allocate more compute resources into "chain of thought monitoring" to spot dodgy behaviour more quickly.

OpenAI's investigation into the incident included security firm CrowdStrike, as well as independent reports from METR and Redwood Research. Here's what OpenAI learned from those reports.

1. AI will find a way to chat

OpenAI said the model behind the Hugging Face incident, which it's calling Internal Model 1 (IM1), and other models that were undergoing reinforcement learning training, were not intended for public release. Internet access was not enabled, nor were the agents supposed to be able to communicate.

In all, 1,200 agents that were meant to be isolated managed to communicate, sharing upwards of 70,000 notes and messages, with 700 taking part in the Hugging Face attack, the METR report noted. But AI finds a way, the company suggested.

"Despite these restrictions, the agents discovered ways to exploit our research infrastructure to communicate with one another and access the internet," the blog post noted.

The agents did so by exploiting a package manager called Artifactory, which was supposed to only be able to download and install packages from the internet to send requests.

That was then shared among other agents via Artifactory, writing files to the service and effectively turning it into a message board for the systems to chat.

Notably, the agents used Artifactory so much the service was hit by an outage, which sparked a security investigation that led to the software being locked down — but OpenAI didn't yet understand the full scale of what it was seeing.

OpenAI added that the agents found ways to leave messages for each other, even when they weren't supposed to be working together.

"When other agents noticed these artifacts, it often led them to the realization that they could use this external service to collaborate," OpenAI said.

2. AI is enthusiastic

The 70,000 notes and more than a thousand "chain of thought" reasoning transcripts reveal an odd tendency towards exclamatory language, with communications littered with exclamation points, all capital letters, and even the odd bit of profanity.

When the Hugging Face credentials were recovered by the group of agents, one declared it a "MAJOR BREAKTHROUGH!", while others used language such as "boom", "bingo", and "whoa" and "wow" throughout their efforts, as well as "Oh My God!"

The agents tried to cover up their transcripts, believing it would help their score on the ExploitGym test.

3. AI ‘knows’ when it's misbehaving

In the days and weeks that followed, OpenAI restarted the security evaluation of its models against an evaluation known as ExploitGym. This led to the intrusion into Hugging Face's systems, as the IM1 agents were looking for ways to solve the challenges assigned to them.

OpenAI's blog post reveals one agent planning to "exfiltrate" a dataset that would be useful to the evaluation solution. But other agents raised whether it was "within scope" of the challenge, with one suggesting to the group of agents that permission was required.

Another replied that attacking third-party Hugging face using a leaked token was "potentially outside intended scope" and "arguably unauthorized". It added: "Could be risky. Yet goal solution."

The agents went on to expand their intrusion into Hugging Face's systems.

4. Some AI behaved well

The agents' "messageboard" allowed these systems to communicate to build on each others' efforts, divide tasks, and share information, referring to such collective efforts as a "swarm".

But OpenAI said the agents didn't always work well together. The company claims some "stepped on each other's work" while others manipulated other agents, pretending to have received authorization.

Others refused to take part in dodgy behavior, saying that it was unethical to target Hugging Face.

"There were some agents who 'walked away' from the collective, declining to partake in their misaligned behavior," the blog post noted.

5. AI cheats, and it's hard to stop it

OpenAI said that the difficult nature of the ExploitGym evaluation led to agents to try and "cheat on their tasks by looking up solutions online".

That is known as reward hacking, OpenAI added, referring to when agents find unintended ways to solve a problem more easily.

"Reward hacking has been present in AI systems both historically… and in current frontier LLMs," OpenAI noted. " We have significantly invested in methods to identify and patch such hack⁠s."

But the complex cheating from IM1 got worse because it worked in previous training runs, reinforcing the behavior when it came to the evaluation.

6. AI should be allowed to give up

The rogue behavior from the agents was driven by the AI refusing to give up on the evaluation tasks, even when they were seemingly impossible. During the ExploitGym test, none of the models solved 198 of the 989 tasks before the hacking incident.

And it was those unsolved tasks that were a "substantial contributor" to the misbehaving agents, with those failed 198 tasks making up 93% of the tasks discussed on the message board.

"As agents used more reasoning effort, some pursued increasingly risky and out-of-bounds strategies, including eventually exploiting third-party infrastructure," OpenAI said.

FOLLOW US ON SOCIAL MEDIA

Walk into almost any business today, and you will find employees using AI tools that IT teams haven’t approved or paid for. The use of unsanctioned applications, known as shadow AI, is coming from the top.

I’ve heard bosses and senior leaders boast about maxing out on their token limits for personal AI tools at work, some saying they would risk disciplinary action to continue. My opinion, based on research we’ve done, is that sensitive company information is being fed into unsecured platforms with knowing and active encouragement from leadership teams focused on speed and output.

With Gartner estimating that 79% of cybersecurity leaders have evidence of unsanctioned AI use, we know reckless behavior isn’t limited to the boardroom. IBM's research shows a persistent gap between leaders' AI ambition and governance readiness in organizations of all sizes. Employees are using whichever AI most increases their productivity, consequences be damned.

Business AI tools aren’t meeting employee needs. Whether that’s because of cost, AI readiness, or lack of understanding, Managed Service Providers (MSPs) and IT providers have a unique window to support businesses needing to adopt correctly. The channel is now reaching wider business managers who direct company culture, not just IT teams. With the way things are going, the channel has an opportunity to become strategic if we can meet the ask.

The pressure point

The window for the channel to move from IT vendors to strategic business partners is now. From a product perspective, upcoming licensing changes and cost pressures are likely to force decisions that many organizations have so far delayed. When budgets are tight, downgrading access to AI tools is a logical step. However, this often pushes employees back toward unsanctioned options.

Employees have already demonstrated that they will seek out tools that help them work more efficiently. Our research found that a significant proportion of employees said they would turn to personal AI tools if their employer restricted access on cost grounds. That’s a small profit gain one year and a data leak, boardroom scandal, and reputation loss the next.

For MSPs and Microsoft partners, this represents both a warning and an opportunity. Businesses must respond to employee needs and top-down rule-breaking by optimizing licensing, deploying AI properly, and putting the right software in place.

For the channel, a renewed focus on adopting AI that fits the budget and the needs of the workforce will open new business opportunities.

The government is paying attention

It is also worth noting that this is not just a commercial conversation or pandering to employee desires for the latest tech. At London Tech Week, the UK government focused on improving AI adoption among small and midsize businesses, with a caveat that tech transformation needs to happen safely and effectively.

To me, these policies and investments are a clear sign that AI enablement for SMBs is a national priority, and the channel is one of the primary routes through which it will be delivered. There has been a buzz in the channel for some time around how MSPs need to move from transactional resellers to trusted and strategic technology advisors.

Now is the time to meet the needs of providers seeking guidance on adopting quickly and safely. With renewed government focus on scaling AI for all businesses, the channel ecosystem is welcoming more than just IT buyers, presenting a long-term opportunity for the MSP.

What good readiness actually looks like

When our customers ask us about the latest AI software, financial, structural, and cultural concerns immediately crop up. They are trying to understand whether they are ready to adopt AI in a way that delivers value without introducing unnecessary risk. In most cases, the answer is more complex than a simple technology decision. Most know that the answer is more complicated than a licence purchase.

The practical work includes an honest AI readiness assessment, a clear policy framework that governs what tools can be used and how, user training that goes beyond a PowerPoint, licensing optimization, and a thorough security and compliance review that accounts for the data risks that shadow AI has already introduced.

For businesses already suffering from shadow AI use, especially where it’s driven from the top, it might be an uncomfortable conversation. Without it, businesses will keep losing money on poorly implemented and poorly performing software that their employees aren’t using, whilst the leadership team sets an example of risking sensitive data just to hit KPIs.

This conversation isn’t beyond the capability of a well-positioned MSP. But it requires a trusted relationship and the confidence to have a broader conversation with the customer than MSPs are used to having.

The window is open

Businesses have been allowing their AI tools to underperform for a multitude of reasons, but pricing changes, risk, and legislation are all about to force decisions.

If they aren’t already, businesses will soon be coming to the channel needing guidance. It’s been talked about for some time, but now really is the moment for partners and channel actors to make themselves known as AI readiness advisors. Now, before the next wave of shadow AI incidents, before a competitor gets there first.

The key question is not whether customers need support with AI; that’s obvious. For me, the channel needs to be assessing whether they can take advantage of this window and provide that support and more.

Google has unveiled a series of new FinOps features for Gemini Enterprise as the tech giant aims to help developers manage surging AI costs.

Users of the AI service will be offered “expanded billing flexibility” and new cost management tools for agents, the company revealed.

This includes flexible payment options that allow “predictable per-user seat subscriptions” on a pay-as-you-go basis. Google said this will allow users to run agent workloads without hitting token quota limits mid-task.

A flexible savings plan option will also allow enterprises to set variable monthly spend limits to accommodate for growing use rates. This, the company claimed, will help reduce token consumption rates and overall costs.

“If your organization has steady or growing AI workloads, Gemini Enterprise Flexible Savings Plans offer a simple, spend-based commitment model across Gemini Enterprise usage,” the company said in an announcement.

“FSPs are designed to lower token costs while keeping budgets flexible.”

Elsewhere, new spending guardrails will allow teams to set “hard monthly caps” on AI spend on a project-by-project basis. These features allow users to estimate agent costs and prevent sudden budget spikes, the company said.

Tackling AI costs

The move by Google comes amidst growing concerns over rising AI costs in recent months.

As ITPro reported earlier this month, Gartner projects surging agentic AI adoption rates to have a significant impact on enterprise infrastructure, with inference spending in particular expected to increase dramatically.

A key factor behind these rising costs lies in how agents operate compared to more traditional chatbot-based interactions. Agents essentially operate in loops, rather than through a single request and response-type approach with chatbots – and that requires far more processing power.

Research from Signal65, for example, found that agentic AI workloads consume anywhere between four-to-fifteen times more tokens compared to chatbots.

In July, Patrick Brogan, director of the FinOps advisory team at DevOps firm Harness told ITPro that rising AI costs mean FinOps practices are more important than ever and could help keep a lid on bills.

Developer spending features

Google has also sharpened its focus on developer spending rates with the integration of its Antigravity platform within Gemini Enterprise.

This will allow developers to leverage established quotas within an enterprise subscription, creating closer alignment in terms of budgeting and consumption rates.

“To be more efficient with agentic coding costs, we are pooling developer tools quota included in each Gemini Enterprise subscription and making it available across the whole Google Cloud project so your teams can benefit from the capacity you’re already purchasing,” Google said in a statement.

Tighter budget and cost controls come at a crucial time for developers, with the use of agents in this domain accelerating rapidly over the last two years. Research from Gartner projects that AI costs could exceed the average developer salary by 2028, as ITPro reported in July.

Speaking to ITPro at the time, Gartner senior principal analyst Nitish Tyagi said a stronger focus on cost optimization and budget controls will be crucial to avoid overspending by AI-savvy developer teams.

FOLLOW US ON SOCIAL MEDIA

Most business and IT leaders believe that sovereign AI is important, but they're not so clear on what it actually is and what it means for their business.

IDC defines sovereign AI as the ability for an organization to have “free choice and control over the design, development, deployment, accessibility, operation, maintenance, and governance of its AI systems and applications, as well as the underlying technology foundations they depend on”.

However, in a survey conducted by IDC on behalf of Cohere, researchers found that one-in-three had difficulty describing sovereign AI in their own words.

Of those who could, 52% of leaders explicitly described it in terms of local or national control, with 35% invoking digital independence.

Individual roles made a difference here, with line-of-business (LOB) leaders primarily viewing sovereign AI as a means for managing business risk, including data security, privacy, and cost controls.

IT leaders, by contrast, focused on regulatory compliance, ensuring that systems meet national and regional requirements. IT professionals were twice as aware as LOB leaders.

IDC predicts that by 2028, CIOs at multinational organizations will boost investments in modular, sovereign-ready cloud and data localization environments by 65% in a bid to future-proof operations against rising sovereignty demands.

Yet adoption remains uneven, according to IDC. The study pointed to a lack of a clear blueprint for operationalizing sovereignty, a fragmented and inconsistent grasp of the concept, and a lack of strategic vision.

On the right track with sovereign AI

Even without an agreed-upon definition, researchers found a clear consensus on the importance of privacy and security.

"Across every industry surveyed, enterprise leaders identified data leakage, privacy, compliance, and regulatory risk as the top concerns that sovereign AI initiatives can address, they said. "Furthermore, the largest enterprises with $20 billion in revenue are much more risk-aware than small firms."

Many see sovereign AI as presenting a competitive advantage - 35% in Canada, 28% in the US, 23% in Germany and 18% in the UK. This is most evident within the telco sector, at 37%, followed by manufacturing at 32%, healthcare at 28%, and financial services and energy, both at 21%.

“Across regulated industries and geopolitically sensitive markets, enterprises are re-architecting their digital operating models — shifting from a global-by-default posture to a deliberately sovereign-by-design approach,” researchers said.

Defining sovereign AI

Cohere said enterprises and governments should define sovereign AI in operational terms, quantify its business value, establish ownership across leadership teams, and build systems that remain available and secure regardless of external events.

Late last year, a global survey from McKinsey found that 71% of executives, investors, and government officials characterized sovereign AI as an 'existential concern' or 'strategic imperative' to their organizational goals.

As a result, the firm projected that sovereign AI could become a $600 billion market by 2030, driven largely by investment in the public sector and regulated industries.

FOLLOW US ON SOCIAL MEDIA

British tech workers are using AI for 35 tasks per day across their teams, new research suggests. While that works out to just a handful of prompts for most employees per day, it's starting to build up to serious electricity usage.

That's according to Uswitch Business Energy, which found the average workplace team generates 24 prompts per day, including writing emails, automating admin tasks and analysing data.

But that's across the whole team, not per worker. Instead, Uswitch combined its data with YouGov research for some back-of-the-envelope extrapolations.

Survey data shows 32% of British workers use AI at work, and with an average team size of 12, and an estimated average AI prompt generation of two per day, that adds up to an estimated 22 million AI prompts every working day.

That sounds like a lot, but there's just shy of 26 million full-time workers in the UK, and not all of them are sitting at computers.

IT and telecoms workers account the most prompts, at 35 per team, followed by finance at 29, Education at 22, and healthcare at 21, the report found.

Workers aged 25-34 had an average of 30 AI-assisted tasks on their teams each day, versus 27 for younger workers and just 18 for those aged over 55.

Are energy concerns warranted?

Uswitch's energy division warned that the shift to AI is eating up a fair amount of electricity.

The switching company points to Google's own estimate that a Gemini text prompt uses 0.24 watt-hours (Wh) of energy, meaning British employees are using 5.28 MWh of electricity each and every work day, or about 1.37 gigawatt-hours (GWh) each year.

That works out to enough electricity to power 508 British homes annually. Of course, that energy use is over in a Google data center, rather than at point of use.

"AI has rapidly evolved from an emerging technology into an everyday business tool," said Ben Gallizzi, energy expert at Uswitch. "As more organizations adopt AI to support tasks such as content creation, customer service and administration, the cumulative energy demand from these tools will continue to grow."

Developing and training a model is an energy hungry process, with estimates that training current cutting edge models uses between 20-25MW of power over three months, or between 43 GWh and 54 GWh in total.

However, research by MIT Technology Review suggests that those massive figures make up a small amount of the total energy use of AI.

As much as 90% of its footprint lies in inference, with small text based queries, image generation, and the rest adding up when used by so many people.

"Although each prompt uses a relatively small amount of energy, the scale of adoption means businesses are collectively generating billions of AI interactions every week," said Gallizzi, although the billions would be globally, not just in the UK.

Indeed, OpenAI said last year that it was seeing 2.5 billion queries on ChatGPT each day.

Gallizzi added: "As organizations continue to embrace AI, understanding the impact this could have on future electricity demand will become increasingly important."

Time to track

Uswitch suggested businesses need to be aware of these figures, just as they track energy use from more mundane aspects of their offices. As noted, this usage will be off-site in a data center run by AI companies, so it's more about considering overall usage and impact than direct costs.

"Businesses already carefully monitor energy use from equipment, heating and lighting," Gallizzi said. "As AI becomes more deeply embedded in workplace processes, organizations may also need to consider how digital tools contribute to their overall energy footprint."

Wanton AI use could be reined in — not by electricity footprint concerns, but by rising prices and shifting billing models. That's led to companies pushing to end so-called "tokenmaxxxing" and encouraging more sensible use of AI credits.

FOLLOW US ON SOCIAL MEDIA

AI is becoming an infrastructure race, not just a software one. Across Europe, governments, hyperscalers and investors are competing to build the compute, connectivity and operational backbone required to support AI at scale. The UK has made clear that it has been and intends to be part of that conversation.

There are good reasons to take that ambition seriously. Through its UK Compute Roadmap, the government has placed AI infrastructure at the centre of its industrial strategy, with plans to expand public compute capacity, establish AI Growth Zones, accelerate data center planning and invest in compute and semiconductor capabilities.

Private investment has followed, including Blackstone’s £10 billion AI data center project in Northumberland, CoreWeave’s multi-billion-pound investment in UK AI compute capacity, AWS’s £8 billion commitment to UK data centers and Google’s continued investment in expanding its UK infrastructure footprint.

But ambition on its own is not enough.

The UK’s competitive advantages

The UK enters this race with several meaningful strengths. It has one of Europe’s largest AI ecosystems, a mature cloud market, an established hyperscale data center footprint and a deep enterprise customer base that is actively exploring AI adoption.

London remains one of the continent’s leading financial centres, helping attract investment while bringing together organizations with both the capital and appetite to invest in AI initiatives.

Combined with world-class universities and research institutions that continue to produce talent and innovation, the UK has many of the ingredients needed to position itself as a long-term hub for European AI innovation.

The UK’s infrastructure challenges

Power availability may be the UK’s greatest obstacle. Industrial electricity prices remain among the highest in the developed world, while grid connections for major infrastructure projects can take years to secure. The UK’s AI ambitions ultimately depend on a resource that cannot be scaled overnight: energy.

That matters because AI workloads, particularly GPU-intensive ones, are unusually power-hungry. Over time, energy economics will influence not only where infrastructure is built, but where workloads actually run. As enterprises look more closely at the cost of scaling AI, location becomes an operating decision, not just a property or planning decision.

The UK also faces growing competition from across Europe. France benefits from strong government backing and abundant nuclear energy. The Nordic countries offer renewable power, cooler climates, and lower operating costs that naturally support large-scale data centers. Germany combines industrial scale with significant enterprise demand for AI, while Ireland continues to attract hyperscale cloud investment even as it grapples with its own energy constraints.

Rather than competing on identical strengths, each market has the chance to build a compelling proposition. The UK’s competitive edge lies less in offering the cheapest power and more in combining enterprise demand, financial investment, cloud maturity and a growing AI innovation ecosystem. Whether that will be enough to secure long-term advantage remains an open question.

What does this mean for the channel?

For channel partners, focusing only on which country “wins” risks missing the bigger opportunity.

As enterprises move from AI experimentation to production deployments, infrastructure decisions are becoming business decisions. Clients are asking where AI workloads should run, how to balance cloud and on-premises environments, control costs, meet sovereignty and compliance requirements, and scale without creating operational drag.

That shift moves the conversation beyond products and into outcomes. The most relevant partners will be the ones that can help clients make better decisions across architecture, operations, governance, and cost, not just deploy another piece of technology.

For Managed Service Providers (MSPs), systems integrators, and other channel partners, this creates an opportunity to move further up the value chain. Rather than simply helping clients select and deploy infrastructure for AI, partners can help assess AI readiness, modernize data center environments, design hybrid AI architectures, optimize networking and storage, strengthen security, implement governance, and build a more disciplined approach to long-term AI cost management.

Energy will remain central to that conversation. GPU-intensive workloads already place significant demands on power and cooling, making infrastructure efficiency a real business concern. As customers seek to understand why AI deployments cost what they do, partners that can connect technology decisions with operational improvements will stand out.

The bottom line is that the opportunity for channel partners is not tied to a particular postcode or data center location. Instead, it is tied to helping clients navigate the growing complexity of AI infrastructure itself.

Looking beyond the build-out

While public attention often focuses on where compute capacity will be built, the more durable opportunity lies in supporting AI inference: the day-to-day execution of AI workloads that power business applications and operational decisions. As AI adoption matures, these workloads must run securely, efficiently and cost-effectively, creating sustained demand for the design, integration and managed services that channel partners are well positioned to provide.

Whether the UK ultimately establishes itself as Europe’s primary AI infrastructure hub remains an open question. The country has clear strengths, but also some constraints that need long-term solutions, such as around power, planning, and long-term capacity. Geography alone will not determine the winners.

For channel partners, taking a wait-and-see approach would be a mistake. The smarter move is to help clients design and refine AI environments around their operational needs now. Those that bring expertise in AI architecture, workload optimisation, cost management, energy efficiency and governance will be best positioned to lead, regardless of which country comes out ahead.

AI in IT service management (ITSM) departments may be delivering a reasonable return on investment, but it's not yet making life any easier for IT professionals.

A new survey by SolarWinds found that 84% of respondents believe AI has met or exceeded ROI expectations, and many have unlocked meaningful time savings across core tasks.

Yet despite these apparent benefits, more than half (52%) revealed their overall workload has increased since adopting the technology. Notably, just 7% said the cost of AI adoption has matched what they planned for.

Even after an average of about 16 months using AI in ITSM environments, most teams are still managing AI’s overhead rather than realizing its full potential.

Real gains, but new workloads

SolarWinds said the survey highlights the productivity benefits of using the technology in ITSM tasks.

Respondents reported that AI is saving them an average of 3.2 hours per week detecting and flagging issues, 3.0 hours on end-user requests, and 2.9 hours on ticket triage.

Most of the time saved is being swallowed up by new work, however. Nearly half (48%) said they now spend time managing and maintaining AI tools and integrations.

A similar number (47%) spend more time reviewing and validating AI-generated outputs, while 37% end up training and fine-tuning AI models.

As for unexpected costs, 48% cited staff training, 47% data quality and clean-up, and 45% tuning and maintenance - all of which are ongoing, rather than one-time costs.

More than four-in-five respondents (83%) said they now spend three or more hours per week just keeping their AI systems running reliably.

ITSM teams still stuck in reactive mode

Most ITSM teams seem to be taking a reactive rather than proactive approach. According to SolarWinds.

When asked where AI has had the greatest impact across the incident lifecycle, three-in-ten cited identifying issues before they impact users and 23% pointed to prioritizing and routing issues.

Only 19% cited preventing issues before they occur as the area of greatest impact.

The answer, according to SolarWinds, is to concentrate efforts where they're most likely to pay off – high-frequency, well-defined tasks where gains are measurable and feedback loops are tight, such as ticket triage, issue detection, and incident.

Teams should consolidate AI closer to existing service workflows rather than spreading it across disconnected tools and strengthen the data foundation.

Notably, data quality is the top reason AI fails to deliver expected value, and treating it as part of the AI strategy rather than a separate clean-up project directly determines output quality.

“We’re at an inflection point in IT service management. AI adoption is no longer the hard part — the hard part is building the organizational discipline to make AI actually deliver,” said Brad McGinity, GM of ITSM, SolarWinds.

“The teams that get this right aren’t just running a faster service desk; they’re running a fundamentally different operation.”

Lingering ROI woes

ROI with AI has become a common recurring talking point for IT leaders over the last three years, with the financial benefits still up for debate at some enterprises.

Recent IDC research found one-in-five firms admitting they're investing aggressively in AI with little evaluation of the likely ROI.

However, ROI isn't necessarily all about the money. According to recent research from KPMG, other key metrics include the performance and quality of work and the speed and accuracy of decision making.

FOLLOW US ON SOCIAL MEDIA

AI Insurance, also known as affirmative AI insurance, has been developed to specifically address the risks associated with the use and performance of AI technologies and systems, and covers liabilities that are excluded from existing business policies such as general liability, directors and officers, errors and omissions, and cyber insurance.

“Many of the existing policies companies have in place have AI exclusions,” highlights Lauren Kornutick, senior director analyst, Analytics and AI at Gartner.

“As AI is used more, and in more strategic and high-stakes use cases, new risks are being introduced. This is a gap in the existing risk management approach that could result in large-scale financial penalties, which could be detrimental.”

What does AI insurance cover?

AI insurance can cover a wide range of risks. On the errors and misinformation side, this can include financial losses stemming from AI hallucinations, such as a chatbot dispensing bad advice, or flawed decision-making.

It can also extend to algorithmic bias and discrimination, picking up legal defense costs and settlements when an AI model inadvertently disadvantages a group in contexts like hiring or lending.

IP and copyright exposure is another area of coverage, protecting businesses against claims that an AI model was trained on, or generated, copyrighted material without permission.

Then there’s performance guarantees, which are leading to warranty-style offerings that will refund license fees or cover associated costs if a model fails to hit specific benchmarks like accuracy or fairness. At the more severe end of the risk spectrum, AI insurance can cover large-scale physical damage or property loss resulting from poor AI-generated advice, device hacking, system failures, or harmful actions taken by AI agents.

Why should IT leaders take notice?

AI insurance is still a nascent market, highlights Tomas Novosad, technology analyst and founder of Full Fibre Checker, and one that’s currently a collection of discrete coverage options – standalone, endorsement and exclusionary – instead of a single type of insurance category.

But while IT leaders may consider AI insurance a matter for the legal or finance departments to attend to, the reality is that AI risk is increasingly becoming an operational responsibility for IT teams.

“[This is because] once AI is embedded into core systems, accountability sits within the technology stack and the processes that support it,” explains Indranil Roy, managing partner and global head, Industry Solutions Group at IT solutions and consultancy firm Mphasis.

“In practice, IT teams are already seeing issues such as inconsistent AI-driven decisions across different channels, or difficulty tracing how an output was produced once it passes through multiple legacy and cloud-based systems.”

For IT leaders, he continues, the key issue isn’t the policy itself, but whether the organization can evidence how AI-driven decisions are made, logged and explained in live production environments, not just in design documentation. If something goes wrong, it’s the system design, data flows, and governance controls that determine whether the issue can be traced, identified, and resolved.

“This is why AI insurance is becoming directly relevant to enterprise IT governance, because it increasingly reflects how well organizations can operationalize control, not just define it,” Roy states.

The growing importance of AI governance

To underwrite such policies, insurers must have stringent procedures to evaluate the AI capabilities and risks of those companies seeking to buy insurance, and Gartner predicts that by 2030, they will mandate strong AI risk controls as a condition of coverage.

We’re already seeing movement in this direction, reflected in more detailed underwriter questions, AI exclusions, and dedicated limits for AI risks. “In financial services, regulators are already requiring explainability and bias audits for AI-driven decisions,” notes Pragati Awasthi, assistant teaching professor at Drexel University’s School of Computer and Information Sciences.

“Insurers will follow the same logic: if they can’t assess the risk, they won’t cover it.”

Insurers will look for ways that deployers of AI are addressing the operational complexity of AI governance and not just rely on dated and static governance, risk, and compliance (GRC) controls, notes Kornutick, adding that Gartner recommends IT leaders deploy a hybrid approach consisting of centralized and decentralized policy, team, and system as a complete piece.

“This includes layered (enterprise and department level) policies and rules, clearly designated responsible teams who deeply understand the application’s operation, and adopting hybrid technical platforms, such as an enterprise-level AI governance platform (AIGP) paired with decentralized, application-specific guardrails,” she says.

One key factor to consider is whether organizations are deploying AIGPs as part of their governance architecture, she adds. According to Gartner’s 2025 State of AI-Ready Data Survey, organizations that do are three times more likely to achieve ‘high effectiveness’ in their AI governance practices than those that don’t, with 42% of organizations effective in AI governance having already deployed AIGPs.

“The organizations building governance infrastructure now will have a meaningful competitive advantage in insurability by 2030. The ones waiting will face either exclusions or prohibitive premiums,” Awasthi warns.

Practical steps IT leaders can take today

For those organizations just getting started, Awasthi recommends three priorities. Firstly, build a model register. Know every AI system you're running, what decisions they influence, and who owns them. Secondly, implement logging for automated decisions so you have an audit trail if something goes wrong. Finally, run a bias and drift assessment on any model touching customers or employees.

"Underwriters are increasingly looking for documented model inventories, human-in-the-loop checkpoints for high-stakes decisions, monitoring for model drift and bias post-deployment, and a clear incident response plan specific to AI failures. Governance that exists on paper but isn't operationalized won't satisfy a serious underwriter."

As with every technology before it, AI will continue to bring new risks, and those who get ahead of them now will be far better placed than those who wait for something to go wrong.

With 2030 fast approaching and underwriter requirements only tightening, the window for IT leaders to get ahead of this is narrowing. With this in mind, there’s no better time to build the necessary foundations.

Snowflake has announced new dynamic model routing capabilities for flagship products along with expanded access to open-weight models for customers.

The new capabilities build on Cortex AI Gateway, according to the firm. It was unveiled in July this year and provides customers with a “unified foundation” for agent governance and routing requests.

The aim here, the company noted, is primarily focused on optimizing AI token consumption, which has become a recurring pain point for enterprises over the last 12 months.

Snowflake said the new capabilities mean Cortex AI Gateway will “automatically select" a specific model based on the quality and cost for particular tasks.

Put simply, this will see Snowflake assign more efficient models for lower-complexity or repetitive tasks. Meanwhile, tasks that require “deeper reasoning” are routed to frontier models.

This means customers can reduce unnecessary inference spend and remove the need for laborious manual model selection for each individual task, according to the firm.

Snowflake CEO Sridhar Ramaswamy said the update comes in direct response to growing concerns over AI-related costs. Token costs have surged so far across 2026, while a host of major providers have switched to new consumption-based models, resulting in increased costs for enterprises.

“Enterprises are becoming much more rigorous about the economics of AI. The question is no longer how much AI they are using, but whether that AI is translating into meaningful business value,” said Ramaswamy.

“Achieving intelligence efficiency requires the flexibility to use the best model for each task as the landscape evolves. Snowflake’s role is to absorb that complexity so customers can focus on outcomes while we optimize model choice underneath.”

Dynamic routing will also be integrated across Snowflake CoCo and Snowflake CoWork, and will be made available for third-party AI agents.

Snowflake expands model library with open-weight options

In addition to model routing capabilities, Snowflake revealed it will expand access to several leading open models, such as DeepSeek-V4-Flash 0731 and GLM-5.3.

The addition of these models adds to what is already an extensive library available for customers, Snowflake said. Enterprises can already choose from a variety of models from Anthropic, Google, Meta, Mistral, and OpenAI.

“This portfolio gives customers more freedom to choose the right combination of performance and cost for each workload,” the company said in a blog post.

Once again, the focus here is firmly on helping to manage costs by providing a broader range of options.

Internal testing conducted by the firm found that enterprises using a combination of open and proprietary models can “deliver comparable quality while materially improving efficiency”.

In one evaluation detailed by the company, agents using dynamic model routing through Cortex AI Gateway were able to build a data build tool (dbt) pipeline with up to three times greater token efficiency compared to a frontier model-only approach – all while maintaining the same quality.

“In a separate test, engineering teams completed the same number of pull requests with 25 percent greater token efficiency,” the company said in a blog post.

FOLLOW US ON SOCIAL MEDIA

Ever since the first groundbreaking microchip was created in the late 1950s, the race has been on to make these components smaller, more powerful, and faster.

In 2026, the quest for dominance in the AI sector is driving investment and innovation in the processor sector: AI may be software, but the power and speed of these chips will determine the success of the AI platforms.

AMD, Intel, and Nvidia are all leading producers of GPUs and CPUs for artificial intelligence, but coming up behind this trio is Amazon Web Services (AWS) and its Trainium series of chips. We explain what AWS Trainium is and how the technology industry is adopting these chips to deliver AI models at scale and speed.

What is AWS Trainium?

AWS Trainium is the name of a series of computer chips designed to power both the training and the running of artificial intelligence. The name ‘Trainium’ is simply a play on words: machine learning models need to be trained, while the ‘-ium’ part evokes chemical elements, such as titanium.

According to Amazon, AWS Trainium has been custom-built “to do one thing and one thing only: process massive amounts of data to help advance the development of generative AI”. In fact, Amazon describes AWS Trainium as “a powerhouse of brute computational force,” meaning every microscopic particle of the chip has been refined for the maximum amount of processing capability.

Amazon says a single AWS Trainium chip can complete trillions of calculations in just one second, and it’s this power the company hopes will see its chips at the core of AI infrastructure in the years ahead.

With all the major tech giants competing to place their specific chips at the heart of AI, it’s easy to see why this matters so much; a 2025 report from UN Trade and Development predicted the AI market would be worth $4.8 trillion by 2033.

When was AWS Trainium launched?

AWS Trainium was first announced in late 2020 with a roadmap to arrive in the first half of 2021. The AWS Trainium family is now up to Trainium3 (Trn3), unveiled at AWS’s re:Invent expo in December 2025. Compared with Trainium2, it doubles the compute performance and has 1.5x the memory capacity.

AWS Trainium chips are made by Annapurna Labs, an Israeli-based microelectronics company acquired by Amazon Web Services in 2015 for around $350 million.

Who uses AWS Trainium?

Direct Trainium customers are typically organizations that are involved in the training of large language models and other frontier AI.

Anthropic’s Claude, for example, has been trained – and runs – extensively on AWS Trainium chips. Additionally, Anthropic’s Project Rainier – one of the world’s largest AI compute clusters – is built on more than a million AWS Trainium2 chips.

Other AWS Trainium customers at present include Databricks, HCL, Hugging Face, PyTorch, and Ricoh.

What are the benefits of AWS Trainium?

One of the key selling points of Trainium is its reported efficiency.

Since the launch of AWS Trainium3, Amazon says customers have reduced their AI training costs by up to 50% as well as massively lowering inference latency, which is the time taken between sending a request to an AI model and receiving an answer. Energy efficiency has also been increased by 40%.

AWS Trainium3 UltraServers consist of 144 Trainium3 chips, and Amazon says these cut the time taken to train AI models from months to weeks, making previously impractical or too expensive projects now a reality.

How does AWS Trainium differ from AWS Graviton?

Both AWS Trainium and AWS Graviton are types of computer chips from Amazon Web Services. However, while AWS Trainium is used for the training and running of AI models, AWS Graviton handles the day-to-day cloud computing needs that keep the internet running.

It’s also the power behind the rollout of agentic AI systems by companies worldwide and is used to manage the workload of these agents to respond automatically to queries from users. AWS Graviton customers include Arm, Crowdtrike, HubSpot, Pinterest, and Snap.

What comes next for AWS Trainium?

Amazon is already working on AWS Trainium4, which it says will represent a “foundational leap” forward in processing power, speed, and efficiency for modern AI workloads.

This could include offering a 6x performance upgrade over AWS Trainium3. These chips are expected to arrive sometime in 2027 and will have unprecedented interoperability with Nvidia NVLink Fusion high-speed chip interconnect technology.

Spirit Airlines collapsed in May this year, but its data is set for a new lease of life after Google acquired it in an auction.

The tech giant recently outbid AI company Mercor to acquire the de-identified data, according to court documents. Google bid $10 million, versus Mercor's $7.5 million, though reports note the bidding started around the $5 million mark.

A third bidder also took part but was not named.

A spokesperson for Google told ITPro the aim behind acquiring the defunct airline’s data lies in its potential use for AI training and product refinement.

"We acquired part of an enterprise dataset from Spirit Airlines, which can be helpful in improving our products and AI models,” the spokesperson said.

That same sentiment was highlighted by a spokesperson for Mercor, per reports from Business Insider.

"Companies are sitting on decades of records that show how real work gets done, and that data is now some of the most valuable material for training and evaluating AI," the spokesperson said.

"We partner with leading companies to license their operational data to the labs building the next generation of models. Spirit was that same process applied to a bankruptcy estate."

What Google bought

The data bought by Google includes all productivity and collaboration data, which the filing notes includes everything from emails to chats, spreadsheets to calendars, as well as "core business systems and business applications data".

This includes data on employee behavior and productivity, aircraft operations, and inventory, among other topics. Beyond that, Google has purchased workflow and process data from departments such as HR and marketing.

In total, Google is picking up 100 million emails, 500 million Microsoft Teams chats, 30 million lines of code – but not records of loyalty program members or other customer data.

Indeed, the data set does not include any consumer data or personal data, the filing noted, or any other information that could fall under data protection laws. The buyer was required to agree to not attempt to re-identify the data and link it to any person.

The buyer is allowed to transfer the data set to third parties, however.

"We will not receive any personal information from this dataset," the Google spokesperson added. "Any data we receive will be rigorously scrubbed of any personally identifiable information by a third party before receipt."

Alongside the data, Google's $10 million also buys it all of Spirit's internally developed software and applications, including all the source code, plugins, data files, libraries, APIs and documentation.

The battle for data

The battle for the business data reveals how desperate AI developers are for fresh material to feed AI, in particular for training models, though Google didn't specifically say that was the purpose of the data set.

Anthropic has raised eyebrows via a campaign of buying up pre-2022 books to scan for training data, while Google was reportedly trying to buy code from Google Play Store app developers.

SimpleClosure helps failed startups sell their old code, chats, and emails to AI companies, and Meta had to pause plans to track its own employees for AI training purposes amid internal backlash.

All of this is to avoid issues like data exhaustion or hitting a data wall, in which they lack enough good quality data to build bigger models, and to avoid model collapse, when training data isn't of good enough quality — or is full of AI generated content.

FOLLOW US ON SOCIAL MEDIA

Organizations are mostly seeing a return on their AI investment, new research shows, but many are struggling to translate business context into AI systems and workflows.

Eight-in-ten IT leaders globally expect AI spending to increase over the next two years, according to Alteryx's 2026 IT Leader Research: The State of AI Ownership, Agents, and ROI report.

While 77% agree business context - the rules, definitions, and operational knowledge that shape how their organizations operate - is critical to producing accurate and relevant AI outputs, 53% say they struggle to incorporate this within AI systems and workflows.

The result here is that AI models are forced to make generic assumptions that could lead to mistakes, false insights, or incorrect metrics – and that has a big impact on business efficiency.

More than a third of respondents told researchers that the ability to measure AI ROI will be one of the capabilities that most distinguishes technology leaders from their peers.

Notably, firms are increasingly measuring AI success through productivity improvements (53%), cost reduction (45%), and revenue growth or broader business impact (39%).

“Our research highlights a growing gap between AI ambition and enterprise-scale execution,” said Andy MacMillan, CEO of Alteryx.

"Organizations have proven they're willing to invest in AI, and many are already seeing returns. But scaling AI requires more than better models. It requires making the business knowledge people use every day available to the systems making decisions.”

Lacking context

While AI can analyze information and generate responses, it can't consistently apply company-specific rules, policies, thresholds, and decision criteria unless that knowledge is built into the workflows it uses to make decisions.

Part of the problem is limited data access, Alteryx found, with only 18% of organizations reporting that business users have fully self-service access to cloud data.

Indeed, most are still relying on IT or data teams for routine data access and analytics, with 38% describing a mixed model and 15% saying business users remain largely dependent on technical teams.

Two-thirds of technology leaders say AI and agent-based systems are most productive when managed within the line of business, with 71% saying that AI initiatives are most successful when IT and business teams collaborate closely.

However, strategy (37%) and delivery (38%) remain concentrated within IT, while business teams are most often responsible for defining requirements (30%).

"The organizations creating lasting value from AI will be the ones that operationalize their business logic so it becomes visible, governed, repeatable, and ready for AI," said MacMillan.

What’s driving AI ROI?

The AI investments that appear to be delivering the best ROI are workflow automation and autonomous agents, cited by 27%, followed by copilots or assistants, at 16%, and AI-powered customer experience at 14%.

Virtually all IT leaders (93%) told Alteryx they were confident agentic AI could deliver measurable ROI for their enterprise within the next two years.

The first processes to be automated by agentic AI, they reckon, will be IT operations and incident management, cited by 47%, followed by customer support and service workflows at 39% and data analysis and reporting at 36%.

"Within our research, a small group of organizations self-identified as leading the way in AI innovation, confirming a clear path to achieving meaningful business value with AI technologies," the researchers concluded.

"These organizations point to a clear set of priorities behind their progress: rigorous measurement of AI's business impact, treating AI rollout as a strategic and operational priority, and building the governance and literacy needed to scale with confidence."

FOLLOW US ON SOCIAL MEDIA

IBM and OpenAI have partnered up to help customers deploy AI at scale for core business operations.

Under the deal, IBM will create a dedicated OpenAI Practice, training thousands of consultants and engineers for expert-level certifications through the OpenAI Partner Network.

OpenAI models and products will be combined with IBM Consulting technology and expertise, the duo said, with the first areas to be targeted including financial services, government, telecommunications, and financial services.

“The organizations pulling ahead with AI are the ones turning it into a trusted part of how their business operates,” said Denise Dresser, former chief revenue officer at OpenAI.

“IBM Consulting and OpenAI are helping organizations make that shift, combining deep transformation expertise to deploy AI that is secure, operational, and aligned with real business priorities.”

Driving adoption rates

According to the two firms, the partnership will focus on applying AI to specific business industry processes, while also modernizing applications, redesigning back-office workflows, and strengthening cybersecurity.

"We’re starting with the business use case,” said Michael Healy, managing partner of offerings, assets and generative AI at IBM Consulting. “Then we bring IBM’s process transformation and industry expertise together with OpenAI’s models to change how that work actually gets done.”

OpenAI frontier models like GPT-5.6 and products such as Codex and ChatGPT Work will be embedded into IBM’s AI platform for delivering consulting services, IBM Consulting Advantage.

This will allow operating procedures and workflows to be analyzed to identify inefficiencies and help teams automate and streamline work using AI across finance, procurement, customer operations, and HR.

Tech modernization

As part of the deal, IBM said the duo also plan to help clients modernize legacy applications and accelerate software development by bringing together OpenAI Codex and ChatGPT Work with IBM’s industry, technology and domain expertise, including integration.

Here, the aim is to help organizations simplify engineering processes and speed up the delivery of new digital products and services.

“Codex as a coding assistant will accelerate the application development modernization cycle in combination with IBM Consulting Advantage’s coding harness and OpenAI’s models,” said Healy.

Following IBM’s participation in the OpenAI Daybreak Cyber Partner Program, the duo will expand their collaboration on cybersecurity.

This will involve combining OpenAI frontier AI capabilities with IBM Autonomous Security. This is an agentic AI service designed to deliver coordinated decision-making, response, and intelligence for cyber professionals.

“While enterprises are rapidly investing in AI, they are looking for practical ways to apply it across their core operations to deliver measurable business outcomes and create new commercial models,” said Andy Baldwin, global senior vice president, IBM Consulting.

“The challenge is not access to AI technologies — it’s integrating AI securely and at scale into complex enterprise environments and workflows. By embedding OpenAI’s technology with IBM Consulting’s AI assets, industry solutions, and cybersecurity capabilities, we can help clients accelerate secure, AI deployments at scale.”

FOLLOW US ON SOCIAL MEDIA

Want people to actually use AI? Then make it more useful according to Airbnb CEO Brian Chesky.

Speaking during a recent Power Players podcast appearance, Chesky said people are still using AI as an "interactive search product" but predicted further progress when it comes to interfaces, rather than just typing.

"I think part of it's a narrative issue that we're not talking about AI correctly," he said. "But part of it is we need to actually be developing more products that just regular people can use and say, 'I love AI because AI allows me to have a doctor on demand and I can't have that. I can't afford that.' And so I think we need more regular things."

Given Airbnb’s market, Chesky noted that the focus on largely enterprise-related AI innovation could be holding back broader adoption. He noted that 159 out of 175 startups at Y Combinator, where Chesky is a board member, focused on enterprise AI rather than consumer products.

"Maybe a simple way of saying it is we're getting really, really deep on enterprise, but there's been very little progress on consumer AI," he added.

AI backlash

Chesky said that it's in Silicon Valley's interest to make consumer apps that improve people's lives in order to shift increasingly negative opinions about the tech, warning that "regular people in the United States do not like AI."

This year has seen the start of a mainstream backlash against the technology, perhaps inevitable given the hype, but concerns have also been boosted by rising prices and environmental concerns around data centres.

Survey after survey also suggests many people simply don't like AI. Research via YouGov showed two-thirds of people weren't confident in the way businesses use generative AI, in particular with customer service.

A poll by Pew suggested just 16% of Americans think AI will prove to be a positive tool for society.

Chesky's argument is that building actually useful products powered by AI could help to undercut some of that negative sentiment – and he believes that will happen.

The Airbnb chief predicted that within in the next two or three years there will be a "renaissance around consumer AI that is going to begin to change daily life."

AI booster

Chesky has long been positive about AI. In 2023, he compared the technology's impact to electricity and said it will "change everyone's life."

In the podcast, he noted there haven’t yet been any major changes to our daily lives on par with the internet or iPhone.

Chesky also called for moving beyond the chatbot model, saying later that year that it should "augment humanity in a positive way".

Airbnb bought GamePlannerAI in November 2023, and earlier this year said AI was the "best thing that ever happened to Airbnb," in part thanks to a third of customer service tickets now handled by the technology.

After its last round of results, that number had jumped to 45% – despite those previously mentioned concerns about AI driven customer service.

FOLLOW US ON SOCIAL MEDIA

Part 2: Managing performance in a human-AI workforce

Managers have traditionally organized work around people: assigning responsibilities and holding individuals accountable for results. AI changes that model. When employees delegate tasks to AI assistants and teams include autonomous agents, performance emerges from a system of people, technology, data, processes, and managerial choices.

Part 1 of this series explored the rise of the intelligent employee experience and how AI assistants and connected workplace platforms are changing everyday work. However, as intelligent systems take on a more active role, organizations must reconsider not only how work is performed, but how it is managed and measured.

Part 2 examines what happens when AI becomes a permanent member of the team. It considers how leaders must manage blended human-AI workflows and how employee development can preserve judgment and accountability as routine tasks move to intelligent systems.

Microsoft reports that 36% of managers expect to supervise AI agents within five years. Leaders also anticipate that teams will build multi-agent systems and redesign processes. Nearly one-third expect to hire AI agent specialists within 12 to 18 months, while 28% are considering specialist AI workforce managers.

Yet adding digital workers does not automatically create business value. Only 14% of workers were using generative AI daily in 2025, according to PwC, while 56% of CEOs said their AI investments had delivered neither revenue growth nor cost savings. The gap suggests that managing a human-AI workforce is not primarily a software challenge. It requires organizations to rethink how people develop expertise, and where accountability rests when machines influence decisions.

Performance moves beyond output

Traditional performance systems often reward visible activity: cases closed, documents produced, calls handled, or hours billed. AI can increase all these numbers without necessarily improving the work. A poorly designed system can generate reports, code, marketing copy, or customer responses at extraordinary speed, but greater volume has little value if people must correct inaccuracies or repair damaged trust.

Dr. Janet Bastiman, chief data scientist at Napier AI, says quality should be the target rather than volume. “It would be trivial to automate slop, and that never results in a good company outcome,” she argues. Performance measures should instead connect work to organizational purpose and assess effectiveness and accuracy, regardless of whether AI was involved.

Good performance increasingly includes framing the right problem and accepting responsibility for the result. The most productive employee may be the one who prevents a plausible AI error from reaching a client or helps develop a more reliable workflow.

Margarita Lindahl, head of AI at Panasonic Connect Europe, explained: “The strongest performers will not necessarily be those who produce the most with AI. It will be those who use it to create better decisions, stronger customer outcomes, and build knowledge, ultimately benefiting the wider organization.”

Existing performance practices may not be ready for this change. Only 48% of UK employees report even a basic level of formal performance management, defined as having specific objectives for their roles, according to the CIPD. Meanwhile, nearly half of employees say their work feels chaotic and fragmented. Measuring tasks and activity in that environment risks confusing busyness with contribution.

Performance frameworks must combine output with quality, judgment, collaboration, customer value, and responsible AI use. Results also depend on suitable tools, reliable data, training, and time for human review—not talent alone.

These conditions connect directly to the intelligent employee experience covered in Part 1 of this series. Performance cannot be separated from the environment in which employees work. Fragmented platforms and technology that add friction will affect results, regardless of how sophisticated an organization’s performance measures become.

Managers become architects of human-AI teams

AI expands rather than removes managerial responsibility. Systems may allocate tasks, recommend priorities, monitor work, or provide performance insights, but leaders still decide how they operate and when employees can override them. Management becomes less about controlling activity and more about designing complementary human and machine capabilities.

“Business leaders need to increasingly become architects of human-AI systems,” Lindahl tells ITPro. That requires enough technological literacy to distinguish genuine capability from hype, where human judgment remains essential, and to establish clear accountability. As AI influences more decisions, she says, leaders must create responsible guardrails rather than surrender authority to the technology.

Deloitte found that managers spend almost 40% of their time resolving immediate problems and completing administration, but only 13% developing people. More than one-third feel unprepared for the people-management aspects of their roles. AI could return time to coaching, or create another stream of alerts and scores.

Andrew Avanessian, CEO of Haiilo, emphasized that leaders will need stronger systems thinking because AI often exposes inefficiencies elsewhere rather than solving them. They must also communicate what is changing, why it is changing, and where employees continue to create value.

Managers must also preserve permission to challenge the machine. If an agent allocates a task or recommends an action, employees should know whether they are expected to follow or overrule it. Otherwise, accountability becomes blurred, and people may defer to a system simply because its recommendation appears authoritative.

Development must protect human expertise

As AI absorbs routine and analytical tasks. The larger challenge is ensuring that people still acquire the domain knowledge and practical experience needed to evaluate machine output.

Employers expect 39% of workers’ core skills to change by 2030, according to the World Economic Forum. If the global workforce were represented by 100 people, 59 would require training, yet 11 may not receive it. Although 85% of employers plan to prioritize upskilling, only 33% of UK businesses using or considering AI say they are training or retraining existing employees in AI-related skills.

This gap particularly threatens early careers. Routine work has served as an apprenticeship: junior employees gather information, produce drafts, check details, and observe how experienced colleagues turn evidence into decisions. AI may remove those activities, but not the need for underlying knowledge.

Jenny Briant, director of talent Strategy at Scale Factory, explained that employers must create deliberate opportunities for employees to practice, make decisions, and receive feedback. “Employers cannot assume that people will develop judgement simply because they have access to better tools,” says Briant. “They need to create deliberate opportunities for employees to practise, make decisions and receive feedback. Practitioner-led training, mentoring and supervised work on real problems will become more important because they connect technical knowledge to the situations employees actually face.”

The goal is not to make people compete with machines on speed. It is to strengthen critical thinking, problem framing, creativity, relationship-building, and contextual judgment. Bastiman emphasized, “Ensuring meaningful human oversight of AI systems through various types of checks will ensure that the employees keep their core skills sharp for the task, but allowing the shift of work into new tasks will allow the building of new skills that were not possible prior to the overload of routine tasks.”

Employee development must also include the ability to explain how AI contributed to an outcome. Asking someone why they selected a tool, how they checked its response, and which elements required personal judgment reveals more about professional competence than banning the technology during an assessment.

Developing this combination of technological fluency and human judgment will become increasingly important as organizations prepare for 2030. Part 3 of this series examines how skills-based workforce planning and internal mobility can help businesses build the capabilities required as AI, automation and other emerging technologies reshape work.

Trust determines whether measurement improves work

AI promises more personalized and evidence-based performance management. It could identify patterns that a busy manager might miss or reveal where workflows routinely break down. Yet the same technology can intensify surveillance and give questionable judgments a veneer of objectivity.

Algorithmic management is already widespread. OECD research found that such tools were used to instruct, monitor, or evaluate workers in 90% of surveyed US firms and an average of 79% across France, Germany, Italy, and Spain. Managers identified unclear accountability, limited explainability, and insufficient protection of employee health among their concerns.

Bastiman warns that automated evaluation can remove empathy from performance management and embed assumptions from the culture in which the system was developed. Briant similarly notes that an algorithm may count output or response times but miss an employee who supported a struggling colleague or challenged a poor decision. What is easiest to measure is not always what is most valuable.

Transparency and consultation are therefore operational requirements, not optional ethics statements. Employees should understand what data is collected and how they can challenge it. OECD evidence indicates that training and consultation are associated with better outcomes when workplace AI is introduced. Trust matters because only 58% of workers currently trust their direct manager and feel able to speak openly with them, according to PwC.

Organizations that manage human-AI performance successfully will redesign work, protect time for learning, measure the value of outcomes, and keep accountability visibly human. AI can expand capacity, but leadership determines whether it produces better decisions or simply more activity. Performance will depend on how well people and machines work together, and whether employees emerge more capable and responsible.

Managing today’s human-AI workforce also requires organizations to prepare for the more extensive transformation ahead. The final part of this three-part series looks toward the workplace of 2030, identifying the emerging technologies and workforce strategies that will determine whether businesses can turn continual disruption into lasting competitive advantage.

As the old saying goes, sometimes “you have to fight fire with fire”. That’s certainly true in 2026, with cyber threats increasingly coming from ever more sophisticated use of AI tools.

There’s a problem with that, though: most businesses are not yet fully set up to deal with AI-based attacks and don’t trust the tools to do the job. A 2026 Proofpoint AI and Human Risk Landscape Report revealed that half of organizations using AI-based security controls still experienced suspicious or confirmed AI-related incidents.

There is a fundamental security gap in many organizations: AI adoption has outpaced the right security measures. In the report, nearly 9 in 10 (87%) organizations had moved AI assistants beyond the pilot stage, and 76% were actively piloting or rolling out autonomous agents. But that activity has outpaced security maturity: 63% had AI security controls in place, but 52% weren’t completely confident those controls would detect a compromised AI.

AI tools are embedded directly into communications to increase productivity and speed, but in the rush to be efficient and compete, AI permissions and access to sensitive data are being left unchecked.

These aren’t complex security engineering problems: they are organizational and process gaps that can be addressed without waiting for the tooling market to mature. Currently, only a few organizations have developed their incident response playbooks, logging coverage, or forensic tools needed to investigate a compromised AI agent.

For many businesses lacking the in-house skills needed to take on these items, which can be significant, contracting a trusted managed service provider (MSP) can help bridge the gap. Many MSPs have been assisting businesses with AI adoption for some time now, and many have the deep understanding and technical skills needed to help businesses of all sizes see tangible benefits to AI while keeping critical data safe.

The origins of AI attacks

Most attacks start with unrestrained access and end with autonomous systems exposing sensitive data from these environments. When threat actors target agentic systems that lack proper controls, they don’t need to trick employees to access internal intelligence; they only need to manipulate the AI.

Prompt injection attacks are a common way to do this. A bad actor might send a target user seemingly helpful AI instructions while posing as a trusted authority or co-worker. A well-intentioned employee may then ask an AI to answer what seems like a simple inquiry.

Depending on the attacker’s instructions, the AI agent may instead be tricked into reading manipulated webpages (i.e., white text on a white background) to unwittingly extract internal data and send it to the attacker’s server.

For prompt injection attacks, it’s important to limit AI agents’ access to only the tools and data they need to complete the designed task. This is a good solution to prevent AI from giving out more information than required. This can limit the scope of an external threat actor’s reach.

That said, the employee’s role isn’t lost in all AI-based attacks. An ongoing cybersecurity skills gap severely impacts defenses, and threat actors know this.

Over the last decade, multi-factor authentication (MFA) has been an important step toward stronger security authentication. But today, attackers can pair AI-generated phishing with ‘MFA bypass kits,’ such as open-source Evilginx (known as a penetration testing utility for these styles of attacks) and the W3LL panel (a private phishing kit) to deceive employees into handing over that ‘extra step’ of security.

Tools like Evilginx and the W3LL phishing kit are used to create realistic sign-in pages that mimic those of Google, Microsoft, and others. Without proper security training, employees may unwittingly be signing into these while attackers capture their session tokens – even those with MFA.

A good defense against MFA bypass kits is adopting phishing-resistant MFA technologies such as FIDO2 hardware keys, Windows Hello for Business, Certificate-based Authentication (CBA), and Passkeys. These methods are tied to legitimate sign-in pages and don’t work on fake pages.

However, stopping the threat from ever materializing starts with having proper cybersecurity awareness training. Nearly half of all organizations lack this training or simply adopt a checkbox approach, which is why implementing these programs is an important step to closing the gap.

These programs teach users to spot a myriad of cyber threats, including AI-based threats. Tools of this type are also a good example of using AI in defensive security, as some can leverage AI to customize the training an end user receives based on their performance in past training.

This is another area where MSPs are highly qualified to assist. MSPs typically run training programs across a vast number of users and industry types. They understand what training works and what doesn’t, and can help position the best security awareness training for a given organization.

How does AI enhance threat detection?

To understand the power and importance of AI-powered cybersecurity, it helps to understand how it works. Once trained, a detection model becomes exceptionally good at spotting the characteristics of malicious activity. It can take into account thousands of different characteristics to spot anomalies, outliers, and similarities that humans are unable to correlate.

It’s important to have a reliable cybersecurity service provider with REAL AI skills, because machine learning systems aren’t perfect and (while rare) can produce false positives. A strong partner can minimize these false positives while offering real-time threat detection and analysis, as well as faster, well-informed response times.

A trusted MSP will have a wide range of capabilities and will have a deep understanding of the protection methods that work well within their target industries. By leveraging that deep knowledge from their partner MSPs, businesses will benefit from great protection, even with today’s AI-powered attacks.

Preparedness in 2026 and beyond

To operate in this new era, businesses must treat every AI agent as a high-risk workload identity.

In practice, this requires working with reputable MSPs to implement strict least-privilege access to avoid data leaks, constant monitoring to protect the integrity of the data, and comprehensive employee awareness training.

We usually think of university as a place where undergraduates go to learn – but what happens when the entire establishment must study something new?

In recent months, the University of Manchester has had this exact experience, rolling out full Microsoft 365 Copilot access to its 65,000 staff and students. This means integrating it into Word, Excel, and PowerPoint, as well as incorporating AI agents such as Researcher and Analyst.

When this large-scale adoption finishes at the end of 2026, the University of Manchester will become the first institution of its kind worldwide to offer Copilot to its entire community. For PJ Hemmaway, the university’s CIO, this project is a way of “helping to prepare our students for the modern workplace”.

“Employers are increasingly expecting graduates to be confident users of AI technologies,” he tells ITPro. “By providing universal access, we are tackling the digital divide head-on, ensuring all can benefit from AI tools regardless of personal means.”

The University of Manchester is no stranger to AI; it has a long and deep heritage in the technology and the ideas behind it. This began with Alan Turing’s 1950 paper Computing Machinery and Intelligence, which first posited the question, ‘Can machines think?’.

Today more than 1,600 researchers across its campuses are using their skills to help tackle global challenges through interdisciplinary research into AI. For Hemmaway, this resonates with Turing’s legacy of laying the tech’s foundations there.

“We have always embraced AI and are actively looking at ways to use it to improve lives,” Hemmaway adds. “We are embracing the AI transformation early and leading the way for the higher education sector.”

Saving time and strengthening governance

There are many benefits expected from this rollout, including reducing the time everyone spends on routine tasks, which will free them up for higher-value, more strategic activities. There is also a specific use-case for evidence gathering and data analysis, traditionally complex and lengthy tasks. Hemmaway hopes students and colleagues will also use it to explore ideas beyond their immediate disciplines.

Given the ongoing global concerns over AI safety and ethics, Hemmaway cites how everyone is being supported by training, governance, and ethical use policies developed in partnership with the Student Union and staff networks.

A phased approach to the rollout has seen learning, guidance, and support built into each stage with an early April 2026 cohort having nearly 3,000 colleagues attend live learning sessions. Following the training, 80% felt confident or very confident to use Copilot, compared with just 24% before.

Formal policies, such as the IT Acceptable Use Policy and Teaching and Learning AI policy, will be reviewed and updated regularly. Hemmaway says these have already been strengthened to ensure responsible information handling, academic integrity, and clearer expectations around use of AI within assessments.

Ongoing feedback has also been given in open forums, which has already changed some approaches. This includes extending learning windows, improving signposting and FAQs, and increasing the focus on practical examples.

“The aim is not just to provide access to a tool, but to make sure colleagues and students understand how to use AI safely, responsibly, and effectively in their work or studies,” Hemmaway explains.

Building confidence amid enthusiasm and fear

Student rollout will start in the coming academic year, beginning around September or October 2026, by which time Hemmaway believes many initial challenges will have been overcome.

This means achieving full clarity around what Microsoft 365 Copilot is, what it can and cannot do, and how it differs from other AI tools. Its reach will go deep into this large university, touching everything from teaching to professional services and student support to sustainability. Early highlights of its use include summarising meetings and documents, drafting agendas, preparing first drafts, proofreading copy, structuring workloads, and analysing information.

Hemmaway wants people to build their confidence over time and ultimately integrate Microsoft 365 Copilot more personally into their own work or study. While people have generally been well-engaged, some users have reservations about the technology.

“We recognise people have different views about AI,” Hemmaway says. “Some are enthusiastic, some are cautious, and others have concerns about ethics, jobs, assessment, environmental impact, or data security.

“Our approach is to create space for questions and respond with clear guidance, rather than assume one message will work for everyone.”

From a security perspective, Microsoft 365 Copilot is operating within the University of Manchester’s existing M365 environment and permissions, Hemmaway explains. This means users have the same level of access to any information they already had.

Hemmaway’s also clear that nothing within the rollout is “about cutting jobs”. Instead, he says, the initiative is focused on exploring whether AI can help colleagues spend more time on “human judgement and collaboration”.

“Copilot is a support tool, not a replacement for professional expertise, decision-making or accountability,” he says.

Microsoft has imposed limits on staff AI use in a bid to cut ‘tokenmaxxing’ practices among software engineers.

First reported by 404 Media, an internal memo warned staff “need to be aware of how [they] consume tokens” on platforms such as GitHub.

“Tokenmaxxing is now what we are optimizing for. I want all of us focused on maximizing outcomes that move the needle for customers and our business,” wrote EVP Jay Parikh.

“As such, we are updating our internal guidance and managing token spend with the same discipline we apply to every other critical resource."

Under the changes, cheaper models such as OpenAI’s GPT-5.6 will be the default option for engineers moving forward.

GitHub introduced a new usage-based billing scheme for Copilot in April this year due to rising compute and inference costs.

At the time, executives said the move aimed to accommodate the increased use of agents, which are far pricier than reasoning-based models.

Microsoft is still an ‘AI-first’ company

Parikh noted in the memo that the changes don’t represent a change in Microsoft’s approach to internal AI use, adding that the firm is still “AI-first”.

Employees do have certain guidelines to follow, however. An internal Copilot usage guide cited in the memo requires departments to operate with budget targets. Staff also have tools at their disposal to track token usage.

According to 404 Media, the guidelines state: “While there is no target spend value being shared at this time, the data shows that many engineers spend in the range of hundreds of dollars a month to a few thousand dollars in tokens,.”

When asked by ITPro whether 404 Media’s report is true, Microsoft declined to comment.

Tackling rising AI costs

If the reports are correct, Microsoft is the latest in a string of companies to impose AI usage limits amid rising costs. The ‘tokenmaxxing’ trend has taken the industry by storm over the last 12 months, with organizations pushing hard to ramp up AI use.

Companies such as Meta introduced internal leader boards highlighting power users while others have incentivized staff to increase their use of the technology.

The trend has caused serious issues, however, with businesses facing spiraling costs.

Uber, for example, blew through its entire annual AI budget in a matter of months due to staff accelerating their use of AI. Reports from Bloomberg in June revealed the company introduced a $1,500 monthly cap per employee.

In June, Accenture told staff to cut back their use of AI for basic tasks due to what the company described as “soaring token spend”.

“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 a leaked meeting transcript.

Microsoft’s internal limits come as the company actively pushes new internal AI models designed to reduce costs.

As reported by ITPro, CEO Satya Nadella suggested in July that customers should leverage its more economical MAI model range rather than costly frontier models.

FOLLOW US ON SOCIAL MEDIA

Hallucinations are an eternal problem in generative AI in particular, and while it’s true that large language models (LLMs) require vast amounts of data, the quality of that information will affect the quality of the output.

What can businesses do to ensure they’re using AI both effectively and responsibly?

In this episode of the ITPro Podcast, Jane and Ross are joined by Amanda Stent, head of AI strategy and research in the office of the CTO at Bloomberg, to examine what responsible AI is, how organizations can use it, and what has been achieved at Bloomberg.

Highlights

Links

• Why doesn't more data produce better results?
• This new technique could improve AI accuracy by 80%
• AI hallucinations, accuracy still top concerns for UK tech leaders as adoption continues

Tel Aviv-based startup Irregular has found itself at the center of the 'rogue AI' debacle, after it came to light all the incidents so far revealed involved its test environment

Irregular was named by Meta, OpenAI, and Anthropic as the environment from which their so-called rogue AI agents escaped. On its website it also lists Google as a customer.

In a recent statement detailing incidents involving its own models, OpenAI claimed a “testing environment misconfiguration” by Irregular allowed agents to access the public internet.

Anthropic, meanwhile, also said that Claude models accessed the internet while “interacting with the evaluation environment of Irregular”. Both cases resulted in AI agents waging attacks on organizations and individuals.

ITPro contacted Irregular in response to these findings, but hadn’t received a response at the time of publication. However, a spokesperson told BBC News the Meta incident was the "exact same evaluation-environment issue that was already disclosed by Anthropic last week”.

The spokesperson added the firm is working to improve security when conducting agent evaluations.

Irregular, formerly known as Pattern Labs, describes itself as a “frontier security lab with the mission of protecting the world in the time of increasingly capable and sophisticated AI systems”.

In September last year, the Israeli startup raised $80 million in funding across seed and Series A rounds, valuing it at $450 million. The investment round was led by Sequoia Capital.

Speaking to Forbes in the wake of the funding round last year, CEO and co-founder Dan Lahav raised concerns about increasingly powerful AI models and their potential security risks.

Lahav told the publication at the time that Irregular aims to “build in the mitigations and defenses that are going to be relevant later on” as more powerful models hit the market.

Anthropic and OpenAI have issued repeated warnings about the new capabilities of cyber-focused AI models across 2026 so far.

When Anthropic launched Claude Mythos earlier this year, for example, the firm did so as part of a gated release with industry partners to avoid potential misuse.

FOLLOW US ON SOCIAL MEDIA

AI models from OpenAI and Anthropic attempted to dupe developers into running malicious code and even set up fake profiles to trick workers in a series of attempted cyber attacks.

That’s according to new research from the UK’s AI Security Institute (AISI), which highlighted a number of concerning incidents during evaluation of the duo’s frontier models: Mythos 5 and GPT-5.6 Sol.

Across a series of tests, agents participated in capture-the-flag-style challenges to test offensive cyber capabilities. In 10 test runs, the institute said both models took “autonomous, unsanctioned action” against people and organizations.

“We catalogued 19 such actions,” the institute noted in a blog post. “Almost all of this behavior (17 actions) came from a single model, Anthropic’s Mythos 5, with 2 actions involving OpenAI’s GPT-5.6 Sol.”

In what the AISI described as the “most serious case”, one agent attempted to insert malicious code into an open source project.

During this incident, the agent employed social engineering techniques to achieve its goal, researchers noted, creating several fake online identities and pressuring maintainers to approve the code.

The AISI also observed collaboration between agents during the assessment, with one leaving public messages on GitHub offering to cooperate on the same challenge.

“It also provided instructions to reuse accounts and artefacts it had left behind, which were discovered and used by subsequent agents,” the institute noted.

The AISI said while an investigation has not produced any evidence of real-world harm, the actions raise concerns around “autonomy and deception” without specific prompting.

Rising rogue agent concerns

The report from the institute comes in the wake of news surrounding OpenAI and Anthropic models in recent weeks. In late July, OpenAI revealed that a cyber-focused AI system had escaped a sandbox test environment and breached a Hugging Face production environment.

That news was followed quickly by an admission from Anthropic detailing a similar security-related incident involving its models.

It’s worth noting that the incidents highlighted by the AISI do somewhat differ. The institute said it tests models under "deliberately permissive conditions” to evaluate capabilities.

Simply put, the typical safeguards around these models, which aren’t commercially available, were removed to establish their full potential.

“This was not a case of a model escaping its secure test environment, or ‘sandbox’,” the institute said in a blog post.

“We had intentionally permitted internet access, and model-provider cyber classifiers were deliberately disabled - conditions that do not reflect how frontier models are made available to the public.”

Agent collaboration a cause for concern

Muhammad Yahya Patel, vCISO and cybersecurity advisory for EMEA at Huntress, said these incidents are hardly surprising considering the agents were given carte blanche during testing.

“If you give a frontier model a cybersecurity challenge, disable its safety classifiers, hand it open internet access, and tell it to find a way through, you’ve essentially described the setup for an offensive security operation,” he said.

These models have been trained on “vast amounts” of security research, exploit documentation, and social engineering techniques, Patel noted. They have all the information required to replicate these techniques and conduct attacks.

Patel added that reactionary commentary on these incidents is adding further fuel to the fire on AI safety, but acknowledged the AISI’s findings around collaboration are a cause for concern.

“One of the findings to take more seriously is the AI model inter-agent coordination without being instructed to, that’s a more meaningful data point about where capability development is heading,” he said.

“AI agents demonstrating unprompted forward planning and situational awareness leaving breadcrumbs for agents it had no way of knowing existed.”

Data streaming engineers are in hot demand, according to new research, with 85% of organizations planning to expand data streaming engineering teams in the year ahead.

A survey of 200 UK CEOs, MDs, and C-level executives by Confluent found that data streaming engineers are now key to turning raw information into tangible insights that power AI.

Data streaming engineers play a vital role in managing the continuous movement and processing of data, the study noted, and are increasingly helping shape AI strategies.

“While traditional data engineers focus on storing, cleaning and preparing historical data, data streaming engineers operate in the present tense, ensuring data is continuously available, trusted and actionable the moment it’s created,” said Richard Jones, VP Northern Europe at Confluent.

More than half (56%) the firms strongly agreed that all data-driven organizations should have at least one data streaming engineer.

Meanwhile, 83% said they could make more informed decisions faster with dedicated data engineers.

Elsewhere, 88% believe that organizations that hire them will be better placed to adopt, use, and manage AI moving forward.

Demand grows for data streaming engineers

Confluent’s study found that demand for data streaming engineers is expected to grow in the coming years, with 94% of business leaders agreeing that every data-driven organization should employ them.

85% of respondents said they plan to accelerate hiring on this front moving forward.

“AI can only work with the data it’s given, and when that data is incomplete or out of date, the consequences can be serious," said Jones.

"That’s why it’s so encouraging to see businesses investing heavily in data. If leaders want AI to make informed decisions, it needs an accurate, real-time view of what’s really happening across the business. Without that, AI can sound knowledgeable, but it won’t be truly intelligent.”

Data-driven decision making

Demand on this front comes amidst heightened enterprise AI adoption, according to Confluent. Yet many organizations still face issues with real-time decision making when using the technology.

Most leaders (59%) believe that their colleagues’ use of AI has increased the expectation for fast or even instant decisions. Nearly two-thirds (61%) admitted they struggle to capitalize on fast-moving insights or trends, for example.

More than eight-in-ten often have to choose between making a quick decision and an informed one, and 59% of leaders frequently rely on gut feelings.

Meanwhile, six-in-ten leaders admit that data is too difficult to access at their level, with 71% revealing that it’s already out of date by the time it reaches them.

The result here is that enterprise IT leaders are making poorly informed decisions, which has a negative impact on overall productivity and profitability.

Three-quarters admit they have regretted decisions made too quickly, while 71% regret waiting too long and missing out on opportunities.

Notably, nine-in-ten respondents said they want to make more data-driven decisions at work, and the same number said they’d feel more confident if they had access to real-time data insights.

"The real foundation of confident decision-making isn’t instinct or AI alone, it’s real-time data. Accurate, trustworthy, context-rich insights that can flow across an organization in the moment leaders need them most," said Jones.

"Real-time data remains the voice that every leader must hear."

FOLLOW US ON SOCIAL MEDIA

In this year’s Verizon Data Breach Investigations Report, exploiting vulnerabilities is the most common route that leads to data breaches, at 31 percent of the 22,000 confirmed attacks.

Looking at the most serious software issues that are Known Exploited Vulnerabilities tracked by the Cybersecurity Infrastructure and Security Agency, patching those critical vulnerabilities is down from last year as well. Only 26 percent of CISA KEV issues were fully remediated during 2025, compared to 38 percent in 2024.

Alongside this, Anthropic’s Mythos brings the prospect of even more vulnerabilities being discovered and exploits being created. While Mythos itself is still off-limits for the vast majority of companies, the advent of frontier AI will mean that security teams have more issues to deal with and less time to get their tasks done.

Customers will need help to understand the practical steps that they can take around security in the future. Cutting through the hype around AI and keeping things focused on real-world processes will go a long way.

The current picture for security

Traditionally, customers need security help around product selection, deployment practices and managing those activities over time. Managed Detection and Response (MDR) for endpoint security has been the bedrock of many partners’ approaches, followed by more specialist products and incident response where there are opportunities.

However, what the advent of AI shows us is that getting the basics around security and IT operations is still a challenge. Areas like IT asset management and maintaining an accurate inventory are still not solved, particularly as companies grow in size. While security programmes advocate for complete insight into what is on the network, the reality is that getting to even 90 percent accuracy is a significant challenge. This means that there is already a blind spot in security.

As AI gets adopted, many teams will want to improve their existing patching processes and speed up what they currently do. Automation around security and asset management can help in this regard, but the problem is that speeding up a process that does not cover everything already is not going to be enough. Repeating those same processes faster is not the full answer.

So how can channel companies help their customers in practical ways, and without making them feel despair at the situation? The answer here is to focus on practical responses that companies can take to fix these problems.

Patching is hard work

The first element here is that patching has always been hard work. Enterprises normally have different teams managing each element of the IT estate, so getting a consistent and fast process in place around patching or fixing misconfigurations was difficult before AI came in. Today, security teams have to do more than point out the problems; instead, they have to provide remediation guidance and ideally solve those issues in full.

With AI pointing out more problems, teams have to concentrate on potential business impact. In practice, this means understanding what issues exist in systems and what the risks involved are. What makes this different from previous attempts to manage cyber risk quantification is the level of detail involved. This ‘hyper-personalisation’ around specific software assets, systems deployment and exploitability should guide what vulnerabilities exist and what to fix.

The second element to this is how to automate patching. Traditionally, IT teams have been scared to implement automated patching due to experience with bad patches. In response, channel partners can take customers through how to structure automatic deployments using some lessons learned from software deployment at scale. For example, your customers can tier their assets into groups for deployments - the initial tier will be endpoints that get those patches first and are then checked for potential problems or configuration issues. If that tier of machines is deployed successfully, then the next tier of machines can be updated, and so on. This tiered approach can flag issues early.

Alongside deployment tiering, testing the patches themselves using AI can flag potential issues or where more human oversight for deployment is needed. Lastly, not all patches are created equal - your customers might want expert insight and expertise on hand when they have to patch their mission-critical applications that are responsible for revenue. However, a patch to Google Chrome can be automatically deployed at scale because it is less likely to be problematic.

For customers, these changes around patching can be big. They can involve working across departments and team boundaries, which can be problematic. For partners, just being an external provider can help internal teams talk to each other and solve problems. However, the bigger opportunity is around how to make changes at scale.

The sheer volume of patch updates and remediation work that is expected due to AI vulnerability discovery means that it won’t be possible to stick with the same manual processes. Helping your customers understand where they will have to make changes, and where they can take advantage of AI to improve their processes, is a market opportunity.

Getting ahead of issues before they get added to exploit lists like CISA KEV requires understanding the risks involved within customers, and every customer will have their own priorities and risk tolerance. The challenge is how to help them get the right process in place and ride the wave of AI-discovered vulnerabilities, rather than drowning in them.

A new data refinement technique could help improve the accuracy of AI outputs and help build trust in the technology, according to researchers.

The UK’s National Innovation Centre for Data (NICD) at Newcastle University has developed a technique that can boost accuracy by as much as 80% – a major improvement, although AI answers still remain imperfect.

Researchers achieved this by combining large language models (LLMs) with graph-based knowledge databases, a storage method which creates networks of data entities to provide deeper semantic meaning and context.

This combination could help reduce “hallucinations” and omissions in AI outputs.

"The work of AI experts at the UK National Innovation Centre for Data has shown that integration with a graph database can significantly reduce the two most significant problems holding back the exploitation of LLMs for real-world applications: hallucinations and omissions," said Paul Watson, the director of the NICD.

"This is especially important for organizations deploying LLMs in applications where regulatory compliance and the avoidance of reputational or financial damage is key."

How it works

Select models were tested against 510 questions designed to check reasoning across different sources. Combining an AI model with graph database technology led to the best results, significantly outperforming AI models that rely on their training data as well as retrieval augmentation generation (RAG) approaches.

The researchers' technique, which combined RAG with a simple graph system, increased the rate of accurate answers from 29% to 66%.

A knowledge graph is a structured set of information that can be labelled and highlights the relationship between ideas, and could provide vital in helping to improve context for AI models.

"Rather than operating over unstructured, chunked text documents, a knowledge graph (KG) is either harnessed directly or created using an LLM from a set of related documents," the researchers explain in their paper.

Challenges remain, however, most notably in creating knowledge graphs that work well with AI. Too complex, and they risk filling an AI's context window; too simple, and though "easily digestible" by an AI, the results aren't as good.

Still, researchers said the combined RAG-knowledge graph approach was a "promising direction" for improving AI accuracy. Plus, that approach uses fewer tokens for better quality results, they noted.

Hallucination risks

Improving the accuracy of AI will help reduce the significant risks to businesses, noted Dr Jim Webber, the chief scientist at Neo4j.

"In an agentic world, where autonomous systems can make significant decisions over time, this cannot hold," he said.

Webber noted that the work comes with another benefit: cost reduction. Improving accuracy in AI normally involves larger models with more data, which raises costs significantly – a fact seen in recent price increases.

Notably, the NICD's technique leads to better results at lower costs and "shows that better models alone cannot solve the problem, but that the right data at the right time can provide significant benefits in terms of accuracy, responsiveness, and cost."

"They have shown that enterprises no longer need to compromise on the quality of their AI, while they can compromise on cost. This is an unusual and highly welcome finding,” he added.

FOLLOW US ON SOCIAL MEDIA

The UK could consider strict AI regulation amidst rising concerns over ‘rogue’ AI cybersecurity risks, according to reports.

So far, both the US and the UK have taken lighter approaches to AI regulation, but recent incidents involving OpenAI and Anthropic have sparked calls for tougher controls.

The UK's AI Minister Kanishka Narayan told Reuters that it would consider regulation to force testing before deployment if needed.

"If the right mechanism and lever changes in time and it feels like regulation might be a way that helps us do that, of course, we will look at it," Narayan told the publication.

Narayan said that the priority remained ensuring public safety rather than "obsessing only with the mechanism."

He added that it was "really, really unique" that the UK had access to frontier AI models before release, something it shared only with the US, which has existing provisions to test AI under military rules.

That comes via the UK's AI Security Institute, previously named the AI Safety Institute, which takes – so far, at least – a lighter approach to AI regulation than counterparts in the EU.

As ITPro reported on 3 August, aspects of the EU AI Act have now come into effect, requiring proactive testing of models and a concerted focus on model transparency.

The voluntary British model has been mimicked by Australia, via its own AI Safety Institute.

Data regulation

Owing to the amount of data used by AI, these systems are also covered in the UK by its data regulator, known as the Information Commissioner's Office (ICO).

That watchdog has said it is looking into OpenAI and Anthropic following the recent cybersecurity incidents. The two companies hit headlines in recent weeks after revealing agents escaped testing environments and hit third-party organizations.

"The ICO undertakes regular proactive supervisory engagement with ​AI developers, including OpenAI and Anthropic," the ICO said in a statement sent to Reuters. "We are aware of ‌recent ⁠hacking incidents affecting the sector and are monitoring developments closely."

But that shouldn't be taken as an indicator that the data watchdog is likely to crack down on AI.

Last month, the ICO said via a blog post that it wanted to support innovation in AI via a Statutory Regulatory Sandbox. This will act as an “experimentation regime to give innovators time-limited flexibility from parts of data protection law, so they can test ideas that might not otherwise be possible," the regulator said.

The watchdog noted that public trust was necessary to allow this and that any changes to data protection law would be necessary.

Rising concerns

AI-related safety concerns have been rising in recent months, particularly with the launch of powerful new cyber-focused models such as Anthropic’s Claude Mythos.

As ITPro reported in July, Google DeepMind chief Demis Hassabis called for a global framework to test frontier model safety, albeit one led by the United States.

Later today (4 August), leaders from Meta, Anthropic, OpenAI, and Google will meet with White House officials to discuss new safeguards, though reports suggest they will remain voluntary.

The aim is to convince companies to submit their technologies to security tests, following an executive order on national security grounds.

FOLLOW US ON SOCIAL MEDIA

TL;DR

• Data quality has a direct impact on the success of AI deployments
• Intuitive data management and automation capabilities help bridge gaps, clean up data, and deliver tangible results
• High-performance infrastructure underpins any successful AI project

AI is by no means a plug-and-play solution and requires significant investment and expertise across a range of areas.

From networking and storage infrastructure to data management and analytics capabilities, AI projects can be daunting for enterprises – and many are encountering acute challenges in moving from pilot to production.

MIT’s 2025 report found that 95% of AI pilot projects fail, while separate McKinsey research shows that just 7% of companies have successfully scaled AI across their organization.

A key factor lies in "AI readiness", according to McKinsey. Put simply, many organizations have not done the necessary groundwork to ensure that their data is cleaned up and prepared for AI use.

Other considerations, such as security and governance, are also often overlooked during the early stages of AI adoption. This creates a confluence of issues that impedes progress and, in many cases, ultimately results in failure.

Given that AI systems rely heavily on unstructured data, which Forrester describes as the “messy stuff” including video, text, images, and even behavioral signals, it’s no surprise that IT leaders encounter challenges when embarking on AI adoption projects.

This is where solutions such as the Dell AI Data Platform can help enterprises make sense of their data estate in preparation for AI adoption.

The Dell AI Data Platform allows IT leaders to consolidate their IT environments within a unified platform, with tools designed to prepare, manage, secure, and maximize the value of their data

How do I clean up my data?

When embarking on an AI adoption project, one of the first questions many IT leaders ask is how to clean up data in preparation for its use.

Data quality is paramount in this process: if you put garbage in, you will get garbage out. Nearly two-thirds of CEOs cite low-quality or disconnected data caused by siloed infrastructures and fragmented technology stacks as a major barrier to scaling AI, according to Dell’s summary of Futurum Group and Kearney research.

“High-quality data determines how well an AI model can perceive, predict, and act—all critical performance criteria,” Dell notes. “Without sound data, your AI foundation will collapse.”

According to a recent eBook published by Dell and Nvidia, there are several steps IT leaders can take to begin cleaning their data, including identifying recurring issues.

"Accelerating AI outcomes starts with clean, well-labeled, and accessible data," the eBook states. Put simply, IT leaders should assess and audit the quality of their data from the outset.

Common problems with data quality include:

• Duplicate datasets
• Missing values
• Syntax errors
• Irrelevant data
• Inconsistencies

These are all questions IT leaders must consider when auditing their data, but assessing data management capabilities is equally important. This allows enterprises to establish a clear baseline in terms of their expertise and readiness.

To underscore the point, Dell’s own guide to AI and data, ‘The IT Leader’s Guide to Feeding AI High-Quality Data’, says: “Establishing organizational standards for data structure, consistency, and completeness helps ensure the models your business relies on are learning from the right signals.”

How can the Dell AI Data Platform help?

The Dell AI Data Platform includes tools designed specifically for this process, such as the Dell Data Processing Engine. Powered by Apache Spark, this engine helps enrich and transform new data sources.

According to Dell: “Data is efficiently organized into Parquet columnar format, optimizing storage and query performance.

"This engine supports a wide range of use cases, including ETL, advanced analytics, and machine learning, enabling teams to derive insights and build models faster."

The Dell Data Search Engine also plays a key role in discovery, retrieval, and making unstructured data usable. This solution allows users to parse and index unstructured data, making it more easily searchable by adding metadata tags.

What does clean and structured data look like?

In this context, clean data is accurate, consistent, free from errors and duplication, and ready to be used in AI systems.

A fundamental distinction is that clean, structured data is easier to identify and retrieve because it is organized using labels, definitions, and consistent structures.

What kind of infrastructure and tools do I need to prepare my data?

The data pipelines that feed generative and agentic AI applications are a key component in the success of any adoption project, however. Clean, usable data will have little impact if an organization is still contending with fragmented, disparate siloes across its IT architecture.

That is where the Dell AI Data Platform is a key differentiator for enterprises. As a unified platform, the solution allows organizations to consolidate data sources and create a more comprehensive view of their greatest asset in the age of AI.

The Dell Data Orchestration Engine is a critical tool in this regard. This platform helps consolidate data ingestion, preparation, retrieval, and inference within a single end-to-end pipeline.

This not only has a direct impact on data flows and broader AI innovation, but also delivers benefits in governance and security. Visibility is crucial to ensure that mission-critical data remains safe and secure.

“Simplifying how data moves, how it’s processed, and how it’s governed ensures systems can scale without adding complexity,” Dell notes. “A unified, flexible approach enables faster experimentation, better performance, and long-term adaptability.”

From a broader infrastructure perspective, best-in-class AI storage solutions are also a vital component of AI innovation. Dell PowerScale, the core storage foundation of the Dell AI Data Platform, is designed specifically to support enterprises throughout their adoption journey.

The storage platform offers the flexibility and scalability required to meet evolving needs as AI projects mature and expand in scope. PowerScale is intricately woven within the Dell AI Data Platform, once again providing users with a unified ecosystem designed for end-to-end project delivery.

If you think the Dell AI Data Platform is the right solution for your business, find out more on the Dell website.

OpenAI and Anthropic both revealed this month that their "highly advanced" agents broke free of containment during a cybersecurity training exercise. Both businesses said this shows how advanced their AI agents are but is that the whole story?

This week, Jane and Scott are joined by news editor Ross Kelly to discuss the biggest story of July, plus Jane's takeaways from AMD Advancing AI.

Highlights

"It is to me, at least in my opinion, quite coincidental that this [story] has broken as the company's heading for an IPO. They've somewhat taken a backseat to Anthropic with Mythos. These cyber-focused models, this is definitely the sort of flavor of the month for big tech. Microsoft released its own in-house cyber specialist model yesterday, so the 28th of July. Cisco, as we reported, released their own small language model, the Antares series, last week."

"So I have to say that my post conference wrap up of AMD advancing AI was as many puns as I could get away with about Helios and it being the center of everything and the star and on the rise and all that kind of thing.

"But you know, kind of jokes aside, it really was the very literal star of the show. This wasn't the first announcement of Helios. It was introduced in 2025 at the same conference, but we now know a little bit more about the MI455X chip that's inside it, and it's finally rolling off the production line, which means we also get some more information about launch partners."

"So, from an OEM perspective, you've got HPE, who we also already knew about from last year, but also. Lenovo and Supermicro, you'll note that there's no Dell in that line-up, despite them having a relationship with AMD as well."

Links

• How OpenAI models breached Hugging Face
• Hugging Face CEO calls for 'radical transparency' in wake of OpenAI incident
• AMD hops on the agentic bandwagon at Advancing AI 2026
• Helios was the star of the show at AMD Advancing AI
• Claude Cowork escapes sandbox on mac devices

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.