Evaluate a task on four dimensions: repeatability, evidence, consequence and reversibility.
| Dimension | Lower risk | Higher risk |
|---|---|---|
| Repeatability | Same inputs and rubric | Every case needs new judgement |
| Evidence | Current sources are available | Important facts are missing |
| Consequence | Internal draft | Customer, money, rights or production system |
| Reversibility | Easy to undo | Difficult or impossible to undo |
Three operating modes
- Assist: AI drafts or organises. A person decides and acts.
- Controlled automation: AI performs a narrow step with tests, permissions, logging and fallback.
- Human-only: the decision requires authority, empathy, accountability or evidence the system cannot access safely.
A workflow can use more than one mode. For example, AI may summarise a complaint, a rule may calculate the eligible amount and a person may approve the refund.
Escalation is part of the design
"Ask a human" is not a failure when it follows a defined condition. It is the correct output when evidence or authority is insufficient.