Review an AI result before you rely on it

For anyone who plans to send, publish or act on an AI-assisted draft, summary or answer

An AI result can read smoothly and still be wrong, incomplete or aimed at the wrong audience. You are the reviewer of anything you put your name to. This page gives you a four-part check: facts, omissions, tone and audience, and when to escalate. Running the check does not make a result reliable; it makes weak spots visible so you can fix them or hand them to the right person.

The examples use Northstar Services, a fictional company. A project coordinator asks an AI tool to summarize the weekly status of a billing portal migration for stakeholders. The draft below is practice material, not a real project or a company policy.

Worked example: the same update, before and after review

Both versions describe the same fictional migration. The reviewer compared the draft with the project tracker and the latest test notes.

AI draft (as generated)

“The billing portal migration is complete and customers experienced no disruption. All teams have signed off, and the portal now supports every payment method. We're thrilled with the incredible results!”

  • States completion, but only pilot accounts have moved
  • Says “all teams have signed off,” which no source supports
  • Claims every payment method works, yet one is untested
  • Celebratory tone for a status update read by cautious finance stakeholders

Reviewed version (ready to send)

“Pilot accounts moved to the new billing portal on the planned date. Remaining accounts move in the next phase. Finance Operations sign-off is pending until the reconciliation test is rerun. Card and bank transfer payments work in testing; invoice-by-post has not been tested yet. Next update: Friday.”

  • Scope stated: pilot accounts only
  • Pending sign-off named with its dependency
  • Untested payment method listed, not hidden
  • Neutral wording with a clear next update

What the reviewer actually checked

Each claim in the draft was traced to a source or removed.

Claim in the draft Check made Result
Migration is complete Compared with the project tracker phase list Changed to pilot accounts only
No customer disruption Searched the incident log for the cutover window; none found, but no customer feedback reviewed Removed; not verifiable
All teams signed off Read the sign-off record Corrected: Finance Operations pending
Every payment method supported Read the test notes Narrowed to two tested methods; one untested

The four-part check

Work through these in order. Allow time to check the claims and sources; pause the draft when a material claim cannot be verified.

  1. 1

    Check the facts

    Highlight every name, number, date, quotation and “has been done” statement. Trace each one to a source you trust, such as the tracker, the original document or the person who owns it. If you cannot find a source, treat the claim as unconfirmed and remove it or label it.

  2. 2

    Check for omissions

    Ask what the draft leaves out: open risks, dependencies, exceptions, people affected, or the view of a team that disagrees. Compare against your source material, not against the draft itself. Add the missing caveat in plain words.

  3. 3

    Check tone and audience

    Read it as the recipient would. Is the level of detail right for them? Does it sound more certain, more cheerful or more formal than the situation warrants? Replace hype and vague reassurance with specific, neutral wording.

  4. 4

    Decide: use, revise or pause

    Use it if every claim is supported and nothing material is missing. Revise it if you can fix the gaps yourself. Pause and ask a human reviewer if you cannot check a claim, if the topic is outside your expertise, or if the result affects people, money, data or commitments to customers.

When to pause and who to ask

Route the question to the person or process that owns the answer. Your organization's own review routes take precedence over this guide.

Ask a knowledgeable colleague or your manager

Use this for wording and judgment questions where the facts are already checked.

  • You are unsure whether the tone suits the audience
  • Two reasonable readers might interpret the summary differently
  • Say what you checked and what still worries you

Use the applicable reviewed process

Use this when the concern is about subject matter, data or sensitive impact rather than style.

  • The result mentions customer, employee or confidential information
  • It makes a legal, financial, security or contractual statement
  • It would affect a decision about a person
  • Follow the route your organization has approved for that topic

Pause and do not use the output

If you cannot adequately check a result, the safe choice is not to rely on it.

  • No source exists for a key claim
  • The output is long and you cannot review all of it before the deadline
  • Keep the original request and draft so a reviewer can see what was generated

Common questions

Does a clean review mean the result is correct?

No. Review lowers the chance of passing on an error, but it cannot guarantee accuracy. Keep claims tied to sources and stay open to corrections after you send.

What if the AI cites a source?

Open the source yourself. Confirm it exists, says what the draft says, and is current. Citations can be wrong or invented.

Should I tell readers that AI helped draft the text?

Follow your team's or organization's guidance on disclosure. If none applies to your situation, ask your manager before you send.

What if I disagree with the reviewer's change?

Explain the source you relied on. The aim is a supported statement, not a winner. If you still disagree, ask the owner of the subject area to decide.