"Everyone will use agents" is as false as "AI is useless"

In 2026, two equally exaggerated stories about AI agents are competing for attention. One says that within two years, an autonomous agent will carry out any office task without supervision. The other says the whole thing is a marketing bubble with no substance. Data from this year disproves both.

What people intend to do, and what is in production

According to Gartner’s 2026 survey of technology leaders, only 17% of organisations have deployed AI agents so far, while more than 60% expect to do so within the next two years. It is the most aggressive adoption curve Gartner has ever measured for an emerging technology, and one of the widest gaps between wanting something and having implemented it.

Deloitte found a similar pattern in another revealing figure: only 21% of organisations have a mature governance model for autonomous agents, even though 74% expect to use agentic AI to a significant degree within the next two years. Almost everyone is moving towards agents. Almost nobody has the structure to operate them safely yet.

In an independent analysis of more than 1,800 organisations, Cleanlab found that only about 5% have agents in production at a genuinely mature level, often with only basic capabilities. The gap between public talk and what is actually running in production is enormous.

Why total scepticism does not hold up either

The opposite response, “this does not work and is pure marketing”, does not stand up to the data either. Gartner places AI agents at the peak of inflated expectations in its hype cycle and expects them to enter the trough of disillusionment during 2026. But a hype cycle does not mean the technology is useless. It means expectations have run ahead of the ability to implement it well. In specific, well-bounded areas such as customer service, code generation for standard patterns and data processing, companies are reporting real efficiency gains of 20% to 40%.

Andrej Karpathy, one of the industry’s most frequently cited technical voices, summed up the moment with a line that avoids both extremes: the models are impressive, but they still need a lot of work. That describes 2026 better than either of the competing headlines.

Where agents are working, and what those cases have in common

A recent analysis of enterprise agents identifies a pattern among organisations making progress: they do not launch dozens of pilots at once. They focus on two or three high-value use cases, with a clear business owner, defined metrics and explicit limits on what the agent can decide on its own. Organisations that fail often do the opposite: they run pilots designed to impress in a demo, without a clear success measure from the outset.

The areas where agents are taking hold in 2026 share a common feature: they are bounded, well-governed environments, such as IT operations, financial reconciliation, internal support or employee onboarding. They allow for human oversight, have clear limits and deliver returns quickly. Despite the rhetoric, companies are not rolling out fully autonomous agents at scale across every function.

The conclusion that avoids both traps

AI agents are real and already produce measurable value in specific tasks. It is also true that governance, reliability and integration lag far behind the ambition in conference talks and headlines. Neither extreme (everyone will use them tomorrow, or they are useless) describes what is happening. A more accurate account of 2026 is that this is useful technology, still immature in some contexts, that makes faster progress when the problem is well defined than when it is sold as a universal solution.