Set an agent's freedom too tight and people approve every item until approving becomes the job. Set it too loose and the first serious mistake ends the project.
Autonomy works better as a promotion. A job starts with a person approving every step, and it moves up only when the record shows it has earned the next level. That is earned autonomy: the freedom follows the evidence, and the evidence is written down.
This piece is for the chief operating officer (COO) who owns a process that AI agents now run, or soon will, and has to decide how much they may do alone.
Four levels for one job
Autonomy belongs to a job, not to an agent. An agent can be trusted to match invoices on its own and still need approval for every email to a supplier. Place each job on one of four levels.
- Prepare. The agent drafts; a person does the work and uses the draft if it helps.
- Propose. The agent prepares each item, and a person approves, changes or rejects it. Every outcome is recorded.
- Act and report. The agent acts within written limits. A person reviews a sample afterwards, plus everything the agent flagged.
- Act within limits. The agent acts alone inside its limits and raises only the exceptions. A person reads the record on a set day.
Above all four sits a line: the decisions that never move. More on those below.
Five tests before a job moves up
The owner writes the bar for each test before the first case runs, not after the results arrive. A bar set afterwards tends to fit whatever happened.
- Enough cases. The job has run often enough at its current level to mean something: a full month of invoices, say, not one good Tuesday.
- People agreed with it. At the Propose level, count how often a person approved an item without changing it. The changes are the signal: each one shows where the agent's judgement and the business's differ.
- The mistakes were cheap. Every mistake was caught before money left, before a customer read it, or before a wrong record mattered. One costly mistake outweighs a long run of good ones.
- It asked when it was unsure. The agent raised the cases it could not settle instead of guessing. An agent that never raises anything is not always right; it may be hiding its doubt.
- The owner signs the move. The job's owner approves the promotion in writing, names the new limits and sets a date to review them.
A job moves one level at a time. Jumping from Propose to Act within limits skips the stage where you learn what the agent gets wrong when nobody checks first.
What sends a job back down
Promotion only works if demotion is normal. Agree these triggers when you set the bar.
- A serious mistake. One mistake of the kind test 3 exists to prevent sends the job down a level until the cause is understood.
- A change underneath. A new process, a new system or a new AI model under the agent sends the job back to Propose. The record was earned on the old setup.
- The exceptions dry up. If the agent suddenly raises nothing, check why before you celebrate.
An example: supplier invoices
Here is how one job might move. It illustrates the method; it is not taken from a client.
The finance manager owns invoice capture and matching. In the first month, the agents propose a match for every invoice and a clerk approves each one. The clerk's changes cluster around one supplier whose invoices list items differently. That rule is fixed, and the job runs another month at Propose with few changes left.
The finance manager signs the move to Act and report. The agents now match invoices that agree with the order and the delivery note, and flag the rest. A clerk reviews a sample each week. Releasing payment stays exactly where it was, with a person.
Then the business moves to a new accounting system. The job goes back to Propose for a month, by prior agreement, not as a punishment.
Where the method breaks
Approval turns into a rubber stamp. When a person approves item after item all day, they stop reading. Their agreement then measures their fatigue, not the agent's quality. Keep the volume at Propose small enough to read properly, or check a sample of the approvals as well.
Rare jobs never earn a record. A job that runs a few times a year may never collect enough cases. That can be the right answer: it stays at Propose.
The bar moves with the mood. After a good month, people want to promote early. After a bad week, they want to switch everything off. The bar written before the first case protects you from both.
Nobody reads the record. If no one looks at what the agents did, no level above Propose is safe. Name who reads it, and when.
What stays with people
Some decisions never move up, however good the record. Set them before the first job goes live: releasing payments, changing a commitment to a customer or supplier, anything about a person's job or pay, and anything with legal weight. The agents can prepare these. A person decides them.
Autonomy you can explain
The test of this method is simple. When the board asks why an agent was allowed to act on its own, the owner can show the record, the bar and the signature. Freedom you can explain can be widened. Freedom you cannot explain is taken away after the first mistake.
Futuvara is a creative technology agency. We build, implement and run AI platforms for our partners. If you own a process that AI agents run, or soon will, and want a second view on where its lines should sit, start a conversation.
