Levels of autonomy for AI agents: a simple way to plan your next step with AI

The idea

Their point is that how much an AI agent does on its own is a design decision, separate from how capable it is. A very capable agent can still be set up to check with you at every step, and a simple one can be trusted to run a narrow job by itself. So rather than grading the AI, they grade the role left for the person using it.

The five levels

1. You’re the operator. You’re in charge and make the decisions. The AI helps when you ask, and suggests things, but doesn’t act on them without you.
2. You’re a collaborator. You and the AI work on the job together. You can edit what it’s done or take over at any point.
3. You’re a consultant. The AI plans and does most of the work, and comes to you for your expertise and preferences. You steer it with feedback rather than doing the work yourself.
4. You’re the approver. The AI works on its own and only comes to you when it’s stuck, needs access to something, or wants sign-off on something that matters. You decide in advance which actions need your approval.
5. You’re an observer. The AI runs fully on its own. You can watch what it’s doing and switch it off, but you’re not steering it.

How it helps you plan

• Pick a level for each task, not for the whole business. Drafting a reply and paying a supplier deserve very different levels.
• Most businesses get real value at levels one and two first, and that’s a good place to build confidence.
• Each level up needs more trust, more checking and somebody who owns it. Before moving a task up, ask what happens if it gets it wrong, who checks, whether you can see what it did, and how you’d stop it.
• Something being capable doesn’t mean it should run on its own. That’s a choice you get to make.

If you’re just starting out, our AI Capability Matrix is a simpler first step.