Go further · Advanced pathway
Delegating an action to AI.
For users who already understand dialogue, verification and taking back control.
Before continuing
This is not an introductory pathway.
Delegation begins when AI does more than answer and starts using tools or modifying an environment. The user must already know how to:
- distinguish a proposal from an action that has actually been executed;
- verify information and an action trace;
- interrupt the work and take control again;
- recognise a sensitive, irreversible or unauthorised action.
In training: begin with a simulation, then a simple, reversible and supervised action.
The delegation contract
What is not authorised is forbidden by default.
The contract does not give the machine general freedom. It defines a limited mandate, expected evidence and the moments when the human must take back the decision.
Mandate: define what is entrusted
The user writes what the machine may do alone, what is forbidden and what requires approval. The authorised tools, files, data, recipients and budgets are also specified.
A task or permission absent from the contract is forbidden by default.
Question
What are you willing to let the machine do without asking you?Deliverable: describe the result before seeing it
Format, length, recipient, required content and quality criteria are written before execution. The machine cannot move the target afterwards.
Question
How will you recognise that the work is good before you have seen it?Stop points: name the decisions the machine cannot make
The machine stops before deleting, sending, publishing, spending, transmitting data, exceeding a budget or choosing between two important directions. Whenever possible, it first presents a simulation or reversible draft.
Question
At what point must it stop and ask you?
Forecast: write what you expect
During execution, the user records the plan they would have followed and the result they expect. This personal version prevents the machine’s output from becoming the only reference.
Question
What would you have done first?
Trace: audit actions, not only the result
The user checks a verifiable trace: tools called, files touched, sources used, messages sent, costs, errors and the order of actions. The user does not request the machine’s private chain of thought.
Question
Can you describe what the machine actually did, step by step?Gap: compare mandate, result and forecast
Any unauthorised action is a contract violation, even if it appears useful. An additional suggestion is acceptable only when clearly separated and not executed.
Question
What did it do that you had not authorised?Takeover: decide what the human resumes
The user observes, verifies, appropriates and decides. They also choose what to redo personally, what to undo and what the next contract must correct.
Question
What will you redo yourself, and what will you change in the next contract?
Safeguard
A written contract is not a technical permission.
Real access must remain minimal. A backup, an undo option and human approval are required for any sensitive or difficult-to-reverse action.
What remains human
Intention, limits and the final decision.
The machine may help format the contract or perform a preliminary check. It must neither define the mandate for the user nor validate its own execution.
Return to the main method