An AI pilot can work technically and still change nothing. It produces promising answers, and those answers never become part of a decision that matters.
With agents, the stakes go up. The pilot no longer just answers. It acts. So before expanding it, I want to know who owns the decision it touches, what would have to change for the system to support it, and how much it may do without a person reviewing it.
Who owns the operating result?
A committee can sponsor a project. Someone still has to be accountable for the decision or workflow the project is meant to improve.
That person should help define success and settle the trade-offs that show up during implementation. The project owner who tracks milestones is often not the operating owner who lives with the result.
Ask who will maintain the capability, review how it performs, and decide what happens when it fails after the project team has moved on.
What context is missing?
A model doesn’t know which source your team trusts, what a supplier term means, or why a simple-looking case needs an exception.
Some of that knowledge sits in systems. Some sits with the people doing the work. Making it explicit can mean changing data, documentation or the workflow itself.
For an agent that uses tools, context also means boundaries: what it may read, what it may change, and when it must ask. A plausible answer is not permission to act.
How will people use the result?
A useful output can still arrive at the wrong moment or outside the tools people use. It can also create review work nobody planned for.
Walk through the whole task with the people who will use the system. Look at what happens before it is called, how someone checks its work, and what happens next. Include the uncertain cases and the failures, not just the demo that goes well.
A practical first step
Pick a workflow that matters but has clear edges. Name its owner and users, document the context it needs, and decide how you’ll compare the result with today’s way of working.
Then build the smallest useful test. Keep enough visibility to see what the system did, and make it clear who can approve, correct or stop it. Don’t widen its permissions until that evidence is in.
The result should tell you whether the capability deserves a bigger role, and what the organization has to change before it gets one.