Within the next four years, agentic AI will resolve 80% of common customer service issues without the need for human intervention, resulting in a 30% reduction in operational costs, per a 2025 Gartner report. And elsewhere, as you might expect, the companies with the most riding on the success of agentic AI have also been bullish through 2025.
Swami Sivasubramanian, VP of Agentic AI at AWS, has argued that the technology has already delivered a “tectonic change”. And Microsoft has heralded 2025 as the year of the ‘Frontier Firm’, with agentic AI providing “intelligence on tap”, claiming that the number of its AI agents will exceed 1.3 billion by 2028. However, not everyone is convinced that the hype is justified.
“An AI agent can’t deliver intelligence on tap if it’s not producing practical, relevant insights, or if it spends most of its time wrestling with inaccessible data,” Aaron Harris, Global chief technology officer (CTO) at Sage, tells ITPro. “To be effective, agents need reliable, relevant data across every business function. The goal shouldn’t be to bolt AI onto existing workflows and call it innovation; it’s to redesign work so agents become genuine teammates, not just tools.”
A reticence to move too quickly with agentic AI is supported by EY’s recent AI Pulse Survey, which found that while 34% of organizations have already started to implement agentic AI, just 14% of senior business leaders said that the technology had been fully implemented in their company.
The research also found that 89% of senior leaders, while being generally optimistic about agentic AI’s benefits, still believe that built-in human intervention will be crucial. And it appears that, despite the increasing attention being focussed on agentic AI, there are a number of issues which are still holding companies back from full integration.
“These statistics seem to contradict each other,” says Dan Diasio, global consulting AI leader at EY, in the survey. “Our respondents believe agents will manage business units, yet also that they will never be able to do it without human assistance. This speaks to how we are both impressed with and intimidated by what we’re seeing from AI.”
Huge expectations around agentic AI
Peter van der Putten is an assistant professor in AI at Leiden University, one of Europe’s leading research institutions. He agrees that the claims and expectations surrounding agentic AI are leading to unrealistic and unhelpful consequences for enterprises.
“The expectations around AI and agents are huge. And vendors are making statements that all you need to solve your enterprise problems is to unleash an army of agents,” van der Putten tells ITPro. “But if not properly controlled and governed, this army is more likely to go and wreak havoc than bring peace and prosperity in the enterprise. And enterprises know this.”
According to van der Putten, today’s AI agents are unable to take the real-world complexity into account, which the majority of enterprises need to deal with. And the thing that makes them appealing — their apparent autonomy — is also their biggest weakness.
“Enterprises want to innovate, but they are held back by legacy,” van der Putten explains. “They can’t use agents without ensuring they are complying with corporate and government rules and regulations. And even if they would be governed and connected, it is hard to judge if they are doing the right thing if you can’t interpret what they are doing - and why.”
Other experts agree with van der Putten, stating that while the technology behind agentic AI is more than capable of making certain decisions, it may be moving too fast for enterprises to keep pace, especially in areas such as compliance and architecture.
"Today there isn’t yet a robust architecture pattern for implementing agentic AI at the enterprise level,” explains Silvia Lehnis, consulting director of Data and AI at UBDS Digital. “AI is best thought of as infrastructure, and it depends on the existing infrastructure and applications evolving to be able to integrate it.”
A step in the right direction
According to Lehnis, the model context protocol is a step in the right direction but communication patterns within agentic AI may still struggle to scale well enough to handle large levels of traffic and data. This is similar to the challenges faced in early years of the internet, she explains. Just like those early days, enterprise culture is also as much of a barrier as technological concerns.
"Many organizations aren’t ready,” Harris tells ITPro. "The sticking point isn’t the technology – it’s trust. Agents can already reconcile accounts, flag anomalies, even anticipate compliance risks, but adoption will only scale once businesses have confidence in how they operate, explain their reasoning, and can be audited.”
Nowhere is the issue of trust more apparent than in the world of commerce, where AI agents are being used as assistants and autonomous actors, capable of initiating and completing purchases independently of the shopper.
“Platforms like PayPal and Amazon are already enabling agent-led purchases, which leads to more transactions taking place outside the traditional storefront, and often without the retailer having clear visibility into how the purchase was made,” says Xavier Sheikrojan, director of risk intelligence atSignifyd.
Although agentic commerce promises to streamline the path to purchase for businesses, Sheikrojan says that it’s a path paved with “blind spots”. This is because when an AI agent takes over the transaction many of today’s retail processes, rooted in context and behavioral signals such as fraud prevention, disappear.
“What’s still missing is a verification layer designed for non-human buyers: a way to confirm that the agent is trusted, the shopper has authorized the purchase and the information presented was accurate,” Sheikrojan says. “Until those safeguards are in place, agentic AI risks exposing companies to vulnerabilities with fewer tools to manage them.”
With 31% of respondents in a recent Gartner poll saying that they planned to wait until the technology had developed before investing further, and with Gartner also predicting that 40% of current agentic AI projects would be cancelled by 2027, many companies are approaching agentic AI with increasing caution. And, until they can be certain that areas such as trust, compliance, and integration have been fully addressed, many enterprises will continue to put their plans for agentic AI on hold.
With 31% of respondents in a recent Gartner poll saying that they planned to wait until the technology had developed before investing further and Gartner itself predicting that 40% of current agentic AI projects will be cancelled by 2027, many companies are approaching agentic AI with increasing caution. Until they can be certain that areas such as trust, compliance, and integration have been fully addressed, many enterprises will continue to put their plans for agentic AI on hold.
