Give AI a scope note before a review

A report says customer onboarding is slowing down. You ask AI to review the spreadsheet you have, and it finds several delayed accounts. The conclusion sounds definitive. But the spreadsheet only covers one region and is missing the latest week.

The useful concept is a scope note: tell AI what it is reviewing, what is outside the review, and what conclusion the material can support. NIST's AI Risk Management Framework Playbook recommends defining the intended context of use and documenting knowledge limits. For a small review, a few plain sentences can make those boundaries easier to see.

Try it on one review

Choose a low-consequence task where you know what material you are providing, such as reviewing a draft monthly operations summary. Before asking for a conclusion, state the scope:

Review only the attached summary for changes in onboarding delays. It covers the Northeast region through October 24. It does not include the last week of October or other regions. First, restate what you can and cannot assess from these inputs. Then identify any conclusion that would require information outside this scope. Do not fill gaps with assumptions.

This is an illustrative prompt, not a tested recipe. Adjust it to match your actual material. The important part is to describe the boundary accurately, not to use special wording.

Illustrative example: AI replies that the summary can show whether delays changed in the Northeast through October 24, but it cannot establish the current company-wide trend because other regions and the final week are absent. That response does not prove the model understood the summary. It gives you a specific statement to check before relying on its analysis.

Compare the scope it restates with the files, dates, and sections you actually supplied. If it includes material you did not provide, says an omitted period is covered, or skips a limitation you named, correct the scope or stop. If the boundary is right, keep it beside the conclusion when sharing the result. Don’t let a local finding become a company-wide claim just because the answer reads smoothly.

Know what this check can’t do

A scope note helps describe the limits of the review. It does not make the data complete, establish that the analysis is accurate, or ensure the model noticed every omission. NIST notes that generative AI can produce false or internally inconsistent content, including in context-heavy tasks. Check important claims against the underlying material, especially when they could affect customers, money, or staff.

Use this for a review whose inputs are bounded and inspectable. If you cannot say what records were included, the first task is to find that out. If the decision depends on missing information, gather it before treating the AI's conclusion as a basis for action.

The practical check is simple: can you describe what the AI reviewed, and does its conclusion stay within that boundary? If either answer is unclear, the review is not ready to guide the decision.