1. Can you say its job in one sentence?
An agent with a vague brief can't be judged, so it can never earn trust. "Helps the finance team" is not a job. "Matches supplier invoices to purchase orders and flags the ones that don't match" is.
What to look for: a written job, the rules it works under and the points where a person approves. If the job has grown since go-live without anyone writing it down, write it down before you widen it.
2. Can you see everything it has done?
Trust comes from a record, not a feeling. If you can't see what the agent was asked, what it read and what it did, you can't tell whether it has earned anything.
What to look for: a log of every action, in a form a manager can read without an engineer beside them.
3. How often does a person change its work?
This is the most honest measure you have. If people still correct much of what the agent produces, it isn't ready to work alone. If corrections have fallen and stayed low across weeks of real work, that is evidence.
What to look for: the share of the agent's work that a person edited or rejected, tracked over time, on live work rather than test cases. Agree before you start what level would justify the next step.
4. Do you know how it fails?
Every agent makes mistakes. What matters is knowing which ones, and whether they are the kind someone would catch. A mistake that shows up the next day is a different risk from one that sits unnoticed for months.
What to look for: a list of the errors it has made, sorted by how far each one would travel before anyone noticed. Widen what it may do first where mistakes are visible and easy to put right.
5. Can you take the step back?
More freedom for an agent should be a setting you can change, not a rebuild. If widening its job means rewiring the process, you will hesitate to pull it back when you should.
What to look for: a way to return the agent to approval-first in one decision, by a named person, on the same day.
6. Which decisions stay with a person, whatever the record says?
Some calls should not move however good the evidence gets: pricing, hiring, anything with legal exposure, anything that changes what you have promised a customer. Decide these first, so the debate about autonomy never reaches them.
What to look for: a short written list of decisions the agent may prepare but never make.
What the six have in common
None of them is about how clever the model is. They are about a clear job, a readable record, an honest measure and a way back. Autonomy given on the strength of a demo has nothing to stand on when the first bad week comes. Autonomy earned against evidence has a record behind it, and everyone can see why it was given.
If you answered "not sure" to two or more, start there. The agent may be fine. The evidence isn't there yet.
At Futuvara, we build, implement and run AI platforms where AI agents do defined work under rules your business sets.
