Test AI on one real task before investing in a bigger system.

Complete this with the person who owns the result and someone who performs the work. If those are the same person, answer from both perspectives.

1. Responsibility

1. Work we are trying to improve: Name one recurring responsibility, not “use AI.”
2. Current friction: What repeats, slows down, gets missed or requires avoidable rework?
3. Human decision: Which judgment, promise or relationship remains accountable to a person?

2. Boundaries

4. AI-supported portion: What small gathering, organizing, comparing, routing or drafting task may receive assistance?
5. Permitted inputs: What information is approved for this system?
6. Prohibited inputs: What private, regulated, sensitive or unverified information must stay out?
7. Source of truth: Which approved records, policies or examples should the result follow?

3. Review

8. Expected output: What should the system return, in what format, and for whom?
9. Named reviewer: Who checks the result before anyone relies on it?
10. Quality standard: What must be accurate, supported, complete, on-voice and appropriate?
11. Escalation: When must the system stop, mark uncertainty or hand the work to a person?

4. 14-day test

12. Baseline: How does the work happen now, and where does the friction appear?
13. Small measure: Compare repeated steps, preparation time, missed contacts, preventable revision or another observable condition.
14. Stopping condition: What failure, risk or extra supervision means the test should pause?
15. Final decision: Keep, revise, stop or investigate a different problem?

Reusable instruction frame

You are supporting [bounded task] for [person or team]. Use only [approved sources]. Do not use or infer [prohibited information]. Produce [specific output]. Mark uncertainty and missing information instead of guessing. Stop and request human review when [escalation condition]. The final decision and approval belong to [named role].