# Checklist · an AI agent action

Copy this file, name it after the action (for example, “cancel-service”) and answer it before launch. If a question doesn’t apply, write “Not applicable” and why: that way it is clear it was considered.

---

## 1. The action

- **Action:** [what exactly the agent does, in one sentence]
- **Where it acts:** [system, data or account it touches]
- **Who requests it:** [the person, another tool, an event]

## 2. Reversibility and impact

- **Can it be undone?** [yes, with one gesture | yes, with help | no]
- **If it goes wrong, what does it cost?** [money, data, reputation, a person’s time]
- **Impact:** [low | high]
- **Starting level of autonomy:** [1 suggests | 2 asks permission | 3 acts and tells | 4 acts alone]
- **How and when it moves up a level:** [metric and threshold that justify it]

## 3. Intent preview

- [ ] It states the concrete action, not a vague intention.
- [ ] It separates what is reversible from what is not.
- [ ] It allows editing before approval (approving only part of it).
- [ ] It shows the expected effect.
- [ ] It shows the agent’s confidence and what it is based on.

## 4. Permissions

- **What it can read:** [list]
- **What it can write or run:** [list]
- **What is off limits:** [list]
- [ ] The person can see and change these limits.

## 5. During the action

- [ ] Concrete steps and their status, with no endless loading indicator.
- [ ] The person can stop the process midway.
- [ ] If there is an error, it says what happened and what was already done.

## 6. After the action

- [ ] Log: what was done, when and with what data, readable by the person.
- [ ] Undo available for [time] and tested.
- [ ] If it can’t be undone, the person was clearly warned beforehand.

## 7. When the agent is unsure

- **What signal detects it:** [low confidence, missing data, conflicting rules]
- **What it does:** [stops, explains what it is missing and hands the decision to a person]
- **Who it escalates to and with what context:** [person or team; summary and data]

## 8. Accountability

- **Who answers if something goes wrong:** [a person or team, not “the agent”]
- **How the affected person finds out:** [channel and timeframe]

## 9. What is measured

- [ ] Approvals without edits.
- [ ] Use of undo.
- [ ] Time to first approval.
- [ ] Escalations to a person.

## Decision

- **Status:** [draft | approved for pilot | approved | withdrawn]
- **Date and owner:** [date, person]