How far should AI be allowed to decide on its own?
Action by action, according to what an error costs and whether it can be put right. Between ‘AI suggests’ and ‘AI acts alone’ there are several settings. Choosing the right one is often the real strategic decision.
Updated on 24 October 2026
What is going on.
The question is often put as all or nothing: either AI assists and a person decides, or it decides alone. Put that way, it leads to excessive caution or to reckless enthusiasm.
In between, the autonomy dial has four settings. AI can suggest, shedding light on a decision; prepare an action that a person approves before it goes out; execute on its own an operation whose scenario is written in advance; or orchestrate, choosing for itself how several authorised operations follow one another. The right setting depends on the action: what an error costs, whether it can be put right, how often decisions are taken, which rules apply. Setting it action by action, not tool by tool, is often the real strategic decision.
What we observe.
- The same tool can warrant several settings.Answering a routine question and granting a commercial gesture do not commit the company in the same way.
- The chosen setting is rarely written down.It results from technical choices made during the project, without an explicit decision.
- The right setting changes over time.Trust is earned on measured results, and can also be lost.
What you can do right now.
- For every action entrusted to AI, write down what it is allowed to do on its own, and what must go through a person.
- Have that choice approved by the business owner, not just by the project team.
- Decide in advance under what conditions the dial can be turned up or down.