AI Agent Autonomy Levels: How an Agent Earns the Right to Act

AI agent autonomy levels describe how much an agent may do without a person checking. At QAI, an agent starts by proposing while a person reviews everything. Next, it acts on its own where its work has consistently cleared your bar. Then it takes over specific kinds of changes by itself, with every action traceable and reversible. We call it earned autonomy.

This is part 3 of our series on how QAI builds AI agents. Part 2: why AI won't replace you explained why your team's expertise is what makes AI good. This part is about how much you let the AI do, and when.

Think of how you bring on a new employee. In the first weeks, someone checks everything they do. Once they have proven themselves on routine work, they sign off on small things alone. After a while, they own a desk, but they still don't sign the cheques.

An AI agent should earn trust the same way and follow the same rules as an employee: only the access its task needs, and a record of everything it does. The difference is that with an agent, you can measure exactly how often its work was right before you give it more room.

We use three stages.

1. It proposes, a person reviews everything. The agent prepares the work and a person approves or corrects every item. For example, it drafts the monthly bank reconciliation and the bookkeeper checks each line. Reviewing is faster than preparing, so this stage already saves time. It is also where the agent improves fastest. Each correction becomes a lesson it never needs twice: a rule added to its written procedure or, when there is one right answer, an automatic check. Over time, those corrections add up to a test built from your own work, what engineers call a benchmark. Passing that test consistently is what moves the agent to stage 2, and later to stage 3.
2. It acts where it has proven itself. For the kinds of items where its work has cleared your bar consistently, it acts on its own, and everything else still goes to a person. For example, it matches the routine lines itself and sends only the exceptions to the bookkeeper.
3. It acts on its own, within limits. It becomes autonomous one kind of change at a time, never all at once, and every action can be traced, checked and undone. For example, after each call it updates the customer's record in your customer database (CRM) by itself: what was discussed, their current priorities and objectives, who now makes the decisions on their side. It never touches the contract value.

An agent doesn't move up for every task at once. It can be at stage 3 for one kind of work and still at stage 1 for another.

Three conditions, all agreed before the work starts.

First, a bar set on your own work. Together, we decide what a good result looks like and what score the agent must reach. The test is the one built during stage 1, from your own examples and your team's corrections, not a generic public benchmark. The agent moves up only when it passes consistently.

Second, two questions about each action: can we check it, and can we undo it? Some decisions are like a door you can walk back through: a draft, a suggested match, a field you can change back. An agent can take those with light review. Others are like a door that locks behind you: a payment, a message sent to a client, a filing with the government. Those stay with a person.

Third, a named person decides. Moving an agent to the next stage is a decision someone makes and signs. It never happens by default.

Some lines we hold:

1. Money goes through a person. An agent never changes payments or bank details alone.
2. Anything final that leaves the company. A message to a client, a filing with the government, a signed document.
3. Decisions about a person. In Quebec, Law 25 already requires companies to tell people when a decision about them is made solely by automated processing. We keep a person in those decisions by design.
4. Anything you can't stop quickly. Every agent must be stoppable in one step, and every action is recorded with its proof.

Five questions. If you can't answer one, the agent isn't ready for the next stage.

1. The task. What exactly is it doing, and how will we know it was done right?
2. The information. Which sources does it trust, and what does it do when information is missing or contradictory?
3. The actions. What can it change, and which actions need a person's approval?
4. Stopping. How fast can we stop it, understand a failure and restart safely?
5. The decision. Who decides that it moves up, and on what evidence?

Less than you'd think, and it is what makes the gains last. Each stage pays for itself: at stage 1, your team reviews instead of preparing. At stage 2, they only see the exceptions. At stage 3, routine work runs on its own while your people handle what needs judgment.

The alternative, giving an agent broad access on day one, tends to end the same way: one visible mistake, and the whole project loses the team's trust. Earning autonomy in stages is slower for a week and faster for a year.

If you want to see how we split each task between the AI, its written procedure and automatic checks before any of this starts, part 1: skills vs agents covers it.