# Responsible AI Pilot Log

Use this log after completing Method Learn’s **Decide what AI should—and should not—do** path and the Practical AI Workflow Review.

The purpose is to record a small, reversible test of one assistive step. It is not legal, privacy, security, employment, or professional advice, and completing it is not certification. It does not prove that a model, vendor, prompt, or automation is safe or suitable for consequential use.

Do not use private, confidential, financial, health, access, customer, or client information unless there is a legitimate basis, an approved environment, and an appropriately reviewed handling process. Redacted or invented examples are sufficient for an initial learning exercise.

## Pilot identity

- Workflow:
- One step being assisted:
- Named human owner:
- Named reviewer:
- Pilot start:
- Pilot end:
- Approved tool or environment:
- Manual fallback:
- Person who can stop the pilot:

## Why this step is being tested

- Repeated burden being examined:
- Observable finished output:
- Evidence required for a correct result:
- Why assistance may be appropriate:
- Why the decision is not being delegated:

## Decisions AI must not own

- Truthful claim or promise:
- Consequential decision:
- Price, scope, or timing:
- Access or permission:
- Moral, fairness, legal, or professional judgment:
- Exception handling:
- Final publication or communication:

## Acceptance checks

Write checks that can be verified against named evidence.

- [ ] Every stated fact matches a source.
- [ ] Missing information is marked rather than invented.
- [ ] No unapproved promise, price, timing, scope, or outcome appears.
- [ ] Required fields are present.
- [ ] The named owner and next action are correct.
- [ ] Private information is absent unless approved and necessary.
- [ ] A person reviews the result before it affects anyone else.
- [ ] The output can be corrected without hiding the original evidence.

Workflow-specific checks:

- [ ]
- [ ]
- [ ]

## Stop conditions

Stop the test and return the work to a person when:

-
-
-

Examples may include missing evidence, sensitive data, inconsistent output, an unsupported claim, an exception, an unavailable tool, or review work that costs as much as doing the task correctly.

## Safe test log

| Test | Case | Evidence used | Checks passed | Corrections needed | Review time | Decision |
| --- | --- | --- | --- | --- | --- | --- |
| 1 | Ordinary case |  |  |  |  |  |
| 2 | Missing information |  |  |  |  |  |
| 3 | Exception or edge case |  |  |  |  |  |
| 4 | Tool unavailable / manual fallback |  |  |  |  |  |

Fluent language is not a pass. Compare the result with the evidence and acceptance checks.

## Corrections ledger

| Date | Incorrect addition or omission | How review found it | Correction | Process change |
| --- | --- | --- | --- | --- |
|  |  |  |  |  |
|  |  |  |  |  |
|  |  |  |  |  |

## Privacy, access, and failure review

- Minimum data required:
- Data excluded from the test:
- Minimum access required:
- Where test outputs are stored:
- Who can view them:
- When they will be removed:
- How an error becomes visible:
- What happens when the tool is unavailable:
- Manual fallback tested on:

## End-of-pilot decision

Choose one:

- [ ] **KEEP ONE BOUNDED ASSISTANCE STEP.**
- [ ] **REVISE AND TEST AGAIN.**
- [ ] **RETURN THE WORK TO A MANUAL PROCESS.**
- [ ] **STOP UNTIL THE EVIDENCE, ACCESS, OR CONTROLS IMPROVE.**

Reason:


Evidence supporting the decision:


Maintenance owner:

Review date:

## Accountability sign-off

- Person accountable for the workflow:
- Decision date:
- What remains human-owned:
- Condition that will trigger another review:

AI does not approve its own role. The responsible person owns the continued-use decision, truthful claims, quality, access, privacy, commitments, exceptions, and consequences.