AI agents are coming to your operations. Here's the catch.

For most of the last few years, "AI in the business" meant a tool you prompt and a model that answers. You ask, it responds, you decide what to do next. Useful — but bounded. The human is always the one moving the work forward.

Agents change that shape. An agentic system doesn't just answer; it pursues a goal across several steps, holds context as it goes, reaches into other systems to get things done, and makes decisions within boundaries you set — escalating to a person only when it needs to. The gap between a chatbot and an agent is the gap between a calculator and a colleague who owns a task.

The shift is real — and faster than the last one

Stanford's AI Index marks 2025–26 as the point agentic AI became genuinely capable on real-world tasks — even as most organisations still struggle to get it into production. The appetite is real: on McKinsey's mid-2025 survey of nearly 2,000 firms, around three in five were already experimenting with AI agents. And where they reach production, the results can be striking — Vodafone's AI assistant now handles a large share of routine customer contact at a fraction of the cost of live chat, with high first-contact resolution. For the organisations that successfully scale, reported returns can be high and payback periods short.

The technology works. That isn't the problem.

The catch: pilots are easy, production is not

That gap — almost everyone piloting, a fraction in production — is the whole story. Gartner has forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing governance, unclear ROI, and a lack of observability rather than model failure. Other research points to a wide governance gap sitting quietly behind the headline adoption numbers.

Agents don't usually fail because they can't do the work. They fail because the business around them wasn't ready to let something act on its own.

Why agents break the old playbook

Autonomy raises the stakes on everything around the technology. The moment a system can take actions, three questions stop being optional:

• Authority — what is this agent allowed to do on its own, and where must it stop and ask?
• Accountability — when an agent makes a call, who owns that decision?
• Observability — can you see what it did, why, and step in when it goes wrong?

None of these is a feature you can buy off the shelf. They're operating-model decisions — exactly the work that gets skipped in the rush from a promising demo to "let's roll it out."

What an agent looks like in practice

The abstraction is easy to nod along to and hard to picture. So here's a single, ordinary job — chasing an overdue invoice — done end to end by an agent we'll call Ivy. An invoice for £2,400 has gone 14 days past its due date.

Notice the three questions from the last section all show up in those six steps. Authority: Ivy can draft, log and schedule on its own, but not press send to an important customer. Accountability: the approval at step four means a named person owns the decision that reaches the customer. Observability: every action is logged against the invoice, so you can see what it did and why. The autonomy is real, but it stops exactly where the cost of a wrong action starts to climb — and that boundary is a choice you make, not a setting the model arrives with.

How to adopt agents without becoming a statistic

The firms that scale agents successfully tend to do the unglamorous things first:

• Start where the process is well-defined and bounded — clear inputs, a clear definition of "done," a low cost of error.
• Give the agent explicit authority and explicit limits, with a person in the loop wherever a wrong action is expensive.
• Instrument everything from day one, so you can see and audit what the agent actually does.
• Treat governance as part of the build, not a sign-off bolted on at the end.
• Measure against a real operational outcome — and be willing to pull an agent that isn't earning its autonomy.

Agents are the most capable tool the field has produced, and the upside for operations is real. But they reward the businesses that have done the engineering and the governance — and quietly punish the ones that mistook a good pilot for a finished system.

Sources: Stanford HAI, AI Index (2025–2026); Gartner, agentic AI forecast (June 2025); McKinsey, The State of AI (2025); IDC (2025). Adoption and production figures vary widely by source and by how "agentic" is defined.