In this guide 8 sections
The project update reads beautifully. It is concise, confident and ready to send. It also says the launch is on track, although two approvals are missing.
You asked AI to tidy rough notes. Somewhere between the notes and the polished paragraph, an uncertainty became a commitment.
That is the review problem worth solving. You do not need to prove that every sentence was written by a human. You need to know which parts are supported, which parts are interpretation and which parts could cause somebody to act on a false assumption.
Check permission before improving the prompt
First, make sure the tool and the information you plan to use are permitted by your organisation. A convenient consumer account is not automatically an approved place for customer details, internal strategy, personnel records or confidential documents.
If the rules are unclear, ask the appropriate owner. Removing a name may not make a document non-sensitive if its details still identify a person or reveal company information. Use genuinely fictional material to practise a workflow while access is unresolved.
Approval to draft is also not approval to send. A system that can generate a customer email should not silently acquire permission to deliver it, change a record or make a promise on your behalf.
Mark the sentences that carry consequences
NIST's generative AI risk profile[1] describes confabulation: outputs can present incorrect information confidently. It also recommends reviewing and verifying sources and citations. A fluent explanation is therefore not evidence that its underlying claim is true.
For an ordinary work draft, start by marking names, dates, figures, claims about completed work, promises, recommendations and statements attributed to other people. These are useful review targets because they can change what a reader believes or does.
“We discussed a possible Friday launch” is different from “We agreed to launch on Friday”. “The sample contained two complaints” is different from “Customer satisfaction is declining”. “I suggest checking with finance” is different from “Finance has approved”.
You are looking for changes in meaning, not just spelling mistakes.
A worked example: the update that became too certain
Imagine these fictional source notes: “Pilot planned for Friday. Security review outstanding. Three people tested the draft workflow. One could not complete setup. Priya to confirm whether the help team can cover launch.”
An unreliable polished version could read: “Our successful pilot confirms we are ready for Friday's launch. Security and support are aligned, with only minor setup issues remaining.”
The problem is not its tone. The pilot has not happened. The security review is incomplete. Support capacity is not confirmed. Three testers do not establish broad readiness, and the notes do not justify calling the setup failure minor.
A supported rewrite would be: “We are aiming for a Friday pilot, subject to security review and confirmation of help-team coverage. Three people tested the draft workflow; one could not complete setup. We need to resolve or assess that issue before deciding whether to proceed.”
The better version is less impressive and more useful. It preserves the decision somebody still has to make.
Match the check to the claim
For a number, go back to the relevant source and check the calculation, units, date range and denominator. A correct sum can still answer the wrong question if it covers a different period.
For a quotation, find the original wording and context. If you cannot verify it, remove the quotation marks and the attribution or omit the statement. Do not ask another model to invent a source that would make it convenient to keep.
For a project status, check the actual record and owner. For a recommendation, identify the assumptions and trade-offs. “Best option” has little meaning without saying best for which objective, with which constraints.
If your knowledge is not sufficient to judge a technical or high-consequence claim, get a qualified reviewer. Legal, medical, financial, safety-critical and employment decisions require more than a quick writing check. A colleague glancing at a document is not automatically the right review process.
Give the model a narrower job
Instead of asking for a persuasive update, try an instruction such as: “Rewrite these approved notes for clarity. Preserve unresolved questions and conditional dates. Do not add facts, approvals, outcomes or commitments. List any ambiguity separately for me to resolve.”
That can make review easier. It does not guarantee that the output follows every instruction. Read the result against the notes, particularly where the language becomes more certain or more favourable.
You can also ask the tool to identify its factual claims before you review them. Treat the resulting list as a convenience, not a complete audit. It may miss the same unsupported assumption that produced the draft.
Another model can offer a useful challenge, but agreement between models is not independent confirmation of the underlying facts. Evidence comes from the relevant record, source or qualified judgement.
Make the final decision visible
Before sharing, decide whether the draft is ready, needs changes or should not be used. Keep unresolved claims visibly unresolved. If another person must approve a commitment, do not replace that approval with careful wording.
Where company policy or the audience requires disclosure of AI assistance, follow it. Do not claim personal research, interviews or analysis you did not perform.
The point of AI assistance is to help you produce useful work. If checking an unfamiliar output takes longer than doing a small task yourself, changing the workflow may be the sensible improvement.
Your next move
Take one non-sensitive AI-assisted draft and highlight every fact, promise and recommendation. For each, ask: “What would I point to if somebody challenged this?”
Anything without an answer needs verification, clearer uncertainty or removal. The send button is the end of the review, not the start of somebody else's.
If the draft concerns competing commitments, ask what should move when your manager adds more work. For applications, turn responsibilities into CV evidence without inventing numbers.
About this guide
NIST: Artificial Intelligence Risk Management Framework, Generative Artificial Intelligence Profile, AI 600-1, section 2.2 and action MS-2.5-003. Supports the risks of confidently incorrect output and the need to verify sources. The project-update example and review routine are CareerFood's practical illustration, not a certification or a complete AI compliance process.
Sources and further reading
Optional source checks. External sources open in a new tab.