When may an AI agent act on its own?

Short answer. An AI agent may carry out a step on its own when that step has proved reliable: its inputs are stable, its method is known, its exceptions are understood, and someone owns the result. Until then, the step asks a person first. A few kinds of step always go to a person, however reliable the agent is.

This guide shows how to sort each step of a recurring job into one of three sign-off zones, how a step earns the move to acting alone, and how it moves back when something changes.

Three ways a step can run

Every step an AI agent takes sits in one of three sign-off zones. The zone says how much the operator, the person in charge of the agent, is involved.

• Acts alone. The agent carries out the step and writes it to the logbook. The operator reviews the log, not each step.
• Asks first. The agent prepares the step and waits. The operator approves, corrects or rejects it before anything happens.
• The operator signs off. The operator makes the decision. The agent prepares what the operator needs, and never takes the step itself.

A job is rarely in one zone. Reading an invoice can act alone while paying it waits for the operator.

What always goes to the operator

Four kinds of step stay with the operator, whatever the agent’s record:

• Money leaves the company: a payment, a refund, a credit.
• Something goes to an outside party: an email to a customer, a file to an authority, an order to a supplier.
• A commitment is made: a price, a delivery date, a contract term.
• A rule is unclear: the agent cannot tell which rule applies, or the rules disagree.

A person must be able to answer for each of these decisions, so they stay with a person even when the agent has never made a mistake.

Four tests before a step acts alone

Test	The question to ask	What fails it
Stable inputs	Does the step receive the same kind of input in the same form each time?	New suppliers, changing formats, documents from many sources
A known method	Is the way to do the step written down in rules the agent follows?	Steps that rely on someone’s judgement no one has written down
Understood exceptions	Do we know the ways the step goes wrong, and does each one have a path?	Surprises in the last month’s corrections
An owner	Is one named person responsible for the result of the step?	“The team” owns it, or no one does

All four must hold. How long a step must run cleanly before it moves is the operator’s call, set for each step: a step that posts small amounts can earn the move sooner than one that changes customer data.

How a step earns the move

1. Every step starts in “asks first”. The agent prepares, the operator checks.
2. Corrections are written down. When the operator corrects a step, the correction goes into the agent’s rules and the company dictionary, a written list of what the company’s words and numbers mean. The model itself does not change, so what the agent learns stays with the company.
3. Evals show when it is ready. Evals are repeatable tests that run the step on known cases and check the result. They show whether the step now passes the four tests, and that a change of model or rules is safe before it goes live.
4. The operator decides the move. Nothing moves to “acts alone” on its own. The operator moves it, and the move is written to the logbook.

The way back stays open

A step that acts alone is still watched. Its results go to the logbook, exceptions still go to a person, and the step moves back to “asks first” when something changes: a new supplier, a new system, a new rule, or a correction the operator did not expect.

Moving a step back is part of the method: it keeps the company in control while more of the work runs on its own.

An example: posting a supplier invoice

Step	Zone	Why
Read the invoice and pull out the lines	Acts alone, once proved	Stable input from known suppliers, checked against the order
Propose the posting (account, VAT, cost centre)	Asks first, then acts alone for known suppliers	Corrections go into the rules until the proposals stop needing them
Post an invoice from a new supplier	Asks first	The input is new, so the first test fails
Approve the payment	The operator signs off	Money leaves the company

The example is general. Every company’s steps and zones are its own.

You stay accountable

When a step acts alone, the company is still responsible for it. A machine cannot be held accountable; the operator can. Four things carry that responsibility: access the company controls, guardrails that hold, people who understand the system, and a logbook of what was done. Read more in Who is accountable when an AI agent makes an entry?