Every company that puts AI to work falls into the same trap. Either it demands human approval for everything, and the time saved disappears into the approver's queue. Or it lets the system act alone, and the first embarrassing case kills the whole project. Both mistakes come from the same idea: treating autonomy as an on/off switch.
Autonomy is a limit, like a manager's signing authority. It starts low and rises with a track record.
Anthropic, which builds one of the most widely used models in the world, published research in February 2026 reaching the same conclusion: autonomy is not a fixed property of a model or a system, but something that emerges from a deployment — shaped by the model's behaviour, by how the user supervises it and by how the product is designed. In other words: how much the system decides on its own is set by you, in the design, and not by the model vendor.
Three tiers that work in any process
Goes out on its own. A cheap, reversible, high-volume action: classifying, routing, filling in a field, drafting a reply. The error costs one click to correct.
Goes out after an "ok". An action that reaches the customer or moves money below a threshold. The person reviews a finished screen and approves in seconds instead of building it from scratch.
Never goes out alone. An expensive or irreversible decision: credit above the limit, cancelling a contract, anything with a legal effect. Here the AI prepares and the person decides, always.
The rule that moves the limit up
Write the promotion condition before you switch anything on. One that works: a type of case moves up a tier after two hundred human-reviewed occurrences with a correction rate below what the manual process already produced. Publish the number, review it monthly, and take the decision out of the realm of opinion.
It changes the conversation with the team too. Nobody is asking for blind faith in a machine. They are asking for the same thing you ask of a new employee: start with a small mandate, show results, earn a bigger one.
What has to exist for this to work
A record of everything the system did, with the date, the input and the reason. With no record there is no track record, with no track record there is no promotion, and the company is stuck at the first tier forever. Ask for it in the proposal, in writing, before you sign.
Source: Anthropic, "Measuring AI agent autonomy in practice" (February 18, 2026). The sentence above summarizes the finding rather than quoting it.
Want the three tiers drawn for your own process?
It is part of what comes out in writing in the diagnostic, with an owner named for each tier.