A person approving every step is the right way to start. The trouble comes when it is also where things stay, because nobody wrote down what the agents would have to show before they could do more.
Leave that question open and two things happen. The approvals pile up on one desk, and the time the agents were meant to give back goes into reviewing their work. Then the reviewer, facing a long queue of items that are nearly always right, starts to skim. The approval stays on the process map, but the check has become a signature.
Earned autonomy is the way out: an agent does more only when its record shows it has earned it. This piece sets out a promotion review a COO can run.
Three steps, and one list that never moves
An agent's work on a given job sits at one of three steps.
- Asks first. The agent prepares the work, and a person approves each item before anything happens.
- Acts, then reports. The agent completes the work, and a person reviews a sample plus every exception the agent flags.
- Acts within limits. The agent completes the work inside written limits and raises only what falls outside them.
Beside the steps sits a list of decisions that never leave a person's desk, however good the record. Releasing a payment, changing a price or a commitment to a customer, and any decision about someone's job or pay are typical entries. Write that list first. It is what lets people trust the rest of the method.
The promotion review
Run it for one class of work at a time, on a date set in advance. Five questions decide it.
- Is the class of work narrow enough? "Invoices" is too broad to promote. "Supplier invoices that match the order and the delivery note within the agreed tolerance" is a class. Narrow classes move first; the awkward cases stay at the step below.
- Was the bar set before the record started? The job's owner writes down, at the start, how much work must pass through, over what period, and which mistakes would rule the agent out. A bar written after the results are in is a justification, not a test.
- What do the corrections show? Count what the person changed, not what they approved. A clean record means the reviewer checked the items and found nothing to fix. If nobody has corrected anything for a long time, make sure the reviewer was still reading: a sample re-checked by a second person tells you whether the record can be trusted.
- What does a miss cost at the next step? At the next step, some mistakes will travel further before anyone sees them. Ask where a miss would be caught, whether it can be undone and who would feel it. A class of work where one miss is expensive and hard to reverse may stay at "asks first" for good, and that is a sound result.
- What replaces the approval? Promotion changes the check; it does not remove it. Name the new check before anything moves: a weekly sample, every exception reviewed, a short daily summary for the owner. Name the signal that would move the agent back a step.
If all five are answered well, the class moves up one step, never two. If any one is answered badly, it stays where it is, and the review notes what would change the answer.
One review, worked through
Here is how the review might run in a mid-sized business whose finance team has AI agents preparing supplier invoice matches, with a person approving each one. It illustrates the method; it is not taken from a client.
The finance manager narrows the class to invoices that match the order and delivery within tolerance, from suppliers already on the books. Before the record started, they wrote the bar: a set number of matches across one full month-end cycle, and no wrong match approved through to the payment run. The corrections cluster on one supplier who changed their invoice layout mid-month, so that supplier comes out of the class. Taking a supplier out narrows the class; it never lowers the bar. For the rest, the manager's corrections are rare, and a second person's re-check of a sample agrees.
A miss at the next step would still be caught before money leaves, because releasing payments is on the list that never moves. The new check is a weekly sample and every exception, read by the finance manager.
The class moves to "acts, then reports". New suppliers, credit notes and partial deliveries stay at "asks first" until each has a record of its own.
Where the method breaks
The work changes underneath the record. A new system, a new supplier layout or a policy change means the record describes a job that no longer exists. Treat a material change as a reason to review again, and step back if in doubt.
Quiet periods thin the record. A seasonal slowdown can leave too little work to judge. The bar covers volume as well as time; reaching the date without reaching the volume does not meet it.
Pressure to promote early. The busiest month is when everyone wants the agent to do more. It is also the month in which a miss is hardest to catch. Hold the review date.
Stepping back is treated as failure. If moving an agent back a step feels like admitting defeat, people stop doing it. Make it routine: the record moved, so the step moved.
What stays with people
The job's owner sets the bar, runs the review and decides the step. The record informs that decision; it never makes it. The list that never moves belongs to the business, and changing it is a leadership decision, not a review outcome.
Autonomy is written down, or it is guessed
Sooner or later, someone will ask the agents to do more. That can happen by drift, as reviewers skim and approval becomes habit, or by a review with a bar set in advance and a record anyone can read. The second takes a little discipline each month. It is also why the people approving the work will trust what the agents do next.
At Futuvara, we build, implement and run AI platforms where AI agents do defined work under rules your business sets. If your team is approving every step and wants to know what would have to be true to let go of some of it, start a conversation.
Sources:
