Hand over the task. Not the direction.
For each stage, name who is responsible. Set what AI may do, what needs approval and how to override or stop.
Four levels of delegation
Suggest
AI offers options. A person chooses among them, or none.
Risk: the menu narrows imagination.
Draft
AI makes a draft. A person edits, checks and decides.
Risk: review fatigue lets errors through.
Act after approval
AI takes a specific action only after a named person approves it.
Risk: approval becomes a rubber stamp.
Prohibited
AI is not used here.
Reason: risk, consent or keeping the skills to do the work.
Five questions for every stage
- Which human owns this stage, by name or role?
- What is the AI permitted to do here?
- What must be true before anything is approved?
- How can a person override or discard the output?
- What conditions trigger a stop or redirect?
A pause needs a next step
If a correction keeps missing the point, pause. Ask a qualified reviewer what to keep and change. Agree on a limited check, then check the result. The owner can be wrong too. Listen and revise the decision.
Keep the skills to do the work. If no one can work without the tool, refusing it is only a choice on paper.
Apply it in the Authority Map →