How to review AI output before you use it

Keep the speed of a useful draft. Add the judgment that makes it fit for the work.

Start with the free AI output review worksheet, or use the six checks below on paper. This method is for everyday business drafts. It is not a certification or a substitute for specialist review where the consequences require it.

1. Define good before judging good

A pleasant paragraph is easy to approve when the job is vague. Write down the audience, intended use, required details, constraints, and next action first. For a landing-page introduction, that might mean: explain who the service is for, describe two verified deliverables, avoid unsupported performance claims, and invite a fit call.

For a product brief, the finish line could be a testable acceptance criterion and the edge cases a prototype must handle. For a sales note, it might be three verified account facts and three discovery questions that do not pretend to know the prospect's needs.

2. Separate statements from evidence

Highlight numbers, dates, named customers, quotes, promises, and sources. For each, ask what supports it. Open the original page, check its date and context, and confirm that it actually supports the sentence. Reproduce calculations using the original inputs.

NIST's Generative AI Profile describes how generated answers can confidently present false information and misleading citations. That is why a convincing explanation alone should not settle a factual question.

A useful note is specific: “The service notes include monthly reconciliation; they do not include a measured time-saving result.” If there is no supporting evidence, remove the claim, label it as a hypothesis where appropriate, or stop to verify it.

3. Check what was lost

Summaries can sound accurate while omitting the exception that matters. Compare the output with the source notes and constraints. Did a tentative idea become a commitment? Did a customer concern disappear? Did a one-person experiment become a claim about every customer?

4. Read it as the audience

Check whether the intended reader can understand the offer and the next step. Replace vague promises with concrete deliverables. Keep the tone appropriate to the relationship. In sales preparation, an unverified inference should become a question, not fabricated personalization.

5. Review permissions and commitments

Look for details that should not be shared, material you do not have permission to reuse, and promises the business has not approved. Identify the person who can authorize the output. A checked box in a worksheet does not grant permission.

6. Try the next action

Open the call-to-action link. Test the prototype path. Recalculate the total. If the draft proposes a process, walk through a small example with the person responsible for it. Keep the outcome of this check beside the draft so the next reviewer can see what was actually tested.

A worked example: repair a confident marketing draft

Illustrative scenario: a hypothetical bookkeeping service supports solo consultants. Its approved service notes include monthly reconciliation and a monthly summary. No customer-count, award, time-saving, or tax-outcome evidence has been supplied.

The prose is short and confident. It also introduces four unsupported factual or performance claims and a guarantee. Polishing the tone would not fix the underlying problem.

• Purpose: needs revision. The actual monthly deliverables are missing.
• Evidence: needs revision. The award, hours saved, percentage, and customer count have no source in the approved notes.
• Context: needs revision. The audience is solo consultants, but the draft drifts into a general founder claim.
• Audience: needs revision. The reader needs a plain explanation of what the service does.
• Permissions: not checked. Confirm the final claims and invitation with the service owner.
• Next step: not checked. Add the correct fit-call link and test it.

The revised version uses the supplied deliverables and removes unsupported performance claims. It still needs the service owner's review and a tested call-to-action link. This is an editing example, not a real customer result or a claim that the revision increased conversions.

What if the second draft is still wrong?

Narrow the task. Ask for one paragraph or one decision at a time. Give a specific correction tied to a source. If the same error returns, rewrite that part yourself or change the process. Repeated prompting without new evidence is not a verification method.

How much review is enough?

Match the review to the consequences and the task. An internal brainstorming note and a public pricing claim need different checks. Include review and correction time in your workflow estimate. After trying several real examples, record where the process repeatedly fails and improve the brief or handoff.

Turn this into a repeatable habit

Keep three things together: the original inputs, the reviewed output, and the unresolved questions. Use the customer-notes-to-campaign workflow for a marketing example, or the business decision workflow to separate evidence from assumptions.

Source checked October 7, 2026: NIST AI 600-1, Generative AI Profile, section 2.2. The six-check worksheet is AI Cofounder's editorial method, not a NIST certification.