AI agents: from pilot to production

Ask any leadership team whether they use AI agents and most will say yes. Ask how many agents run in production, on real work, without someone babysitting them, and the answer gets much smaller.

Lots of pilots, fewer systems

Adoption surveys this year vary widely because they measure different things. Some find that most companies “run” AI agents, often counting a single trial. Others find that only around a quarter of organisations have scaled agentic AI anywhere, and fewer than one in ten in any single function. Gartner expects 40% of enterprise apps to include task-specific agents by the end of 2026, but that is a forecast, not a measurement.

The gap between “we tried it” and “it runs our process” is where most of the value, and most of the effort, sits.

Why agents stall after the pilot

This week’s CNBC Technology Executive Council summit kept coming back to the same theme: governance. Security leaders pointed out how many decisions an agent has to get right to behave as intended, and argued for unified security and observability data so teams can see what agents are doing.

Vendors are moving the same way. Recent launches include identity and access controls built for agents, and governance features in the major business platforms. The tools are catching up to a simple fact: an agent that can act also needs limits, logs and an owner.

In our experience pilots stall for a few predictable reasons:

• The task was too vague. “Help with operations” can’t be measured. “Match invoices to purchase orders” can.
• No one measured it. Without tests on real examples, nobody can say whether it is good enough to trust.
• It had too much access, or too little. Either it was a security risk, or it couldn’t reach the systems it needed.
• No one owned it. When the demo ended, so did the project.

Five things that get an agent into production

1. One task, clearly defined. Pick a job your team does many times a week, with a clear “done”. Write down what good looks like.
2. Tests on real examples. Build a set of real cases, including the awkward ones, and score the agent on them before launch and after every change.
3. Least-privilege access. Give the agent its own identity and only the permissions the task needs. Read-only first, then widen.
4. Approval on anything that matters. Let the agent draft, prepare and propose. A person approves payments, customer messages and anything hard to undo, at least until the numbers say otherwise.
5. Full logs and an owner. Every action is recorded and reviewable, and a named person is responsible for how the agent performs.

None of this is new or exotic. It is ordinary software engineering applied to a new kind of system, and it’s the difference between an impressive demo and something your team relies on every day.