Make a small AI change reviewable.

In this synthetic scenario, a support team spends part of each shift drafting answers about account setup. Its approved help article says that an administrator must enable exports. An older team note suggests that anyone can enable them. The proposed pilot drafts responses from the approved article and surfaces the conflict for a reviewer; it does not send messages or change account settings.

A useful first step is to inspect a small set of representative questions with the people who answer them. Include a question the article cannot resolve. Record the source, the expected answer and the reason a reviewer might reject a draft. Attendance at a training session is not evidence that the workflow helps; review effort and unresolved errors belong in the decision as well.

• Scope one task and one approved source.
• Keep sending and account changes outside the exercise.
• Retain unsupported questions as test cases.

The same pilot has three responsibilities: maintaining the source, checking a draft and deciding whether the pilot may expand. These responsibilities may belong to different people. A source owner can correct an outdated article without being the person who approves every outgoing response. A reviewer can reject a draft without silently rewriting the underlying policy.

Write down what happens when the sources disagree. For this exercise, stop the affected draft, record the conflict and ask the source owner to resolve it. Do not let repeated copying turn an unapproved note into a policy. If a responsible person is unavailable, keep the decision pending instead of treating silence as authorization.

• Name a responsible role for each decision.
• Separate source corrections from draft edits.
• Specify a pause and escalation path.

Imagine that three synthetic drafts were reviewed: two matched the article and one claimed exports were available to everyone. The instruction was revised and those same examples then passed review. That is useful evidence about a narrow revision. It says nothing yet about unfamiliar questions, different access levels or the effort required during a busy shift.

Before expanding, retain the original failure and try a new conflicting-source example. Record who reviewed it and how much correction was needed, without inventing a performance result. A decision note should separate what was observed, what remains unknown and the next bounded test. If a failure would affect a person or account, the plan needs a human decision before action.

• Keep the denominator and failed cases visible.
• Test a new case after changing the instruction.
• Pause when a draft exceeds its evidence or authority.