AI for managers
How to review AI-assisted work
Written by AI Tools AcademyChecked against the sources listed below on 27 September 2026
Helpful first: Hallucination and trust · How to manage a team that uses AI
What should Priya check? Reading it again for typos won't catch the problems AI tends to create. AI output is usually fluent and well organised, which makes it easy to trust. The mistakes hide in the details: a price from the wrong quote, a delivery term that no supplier offered, a reasonable-sounding assumption nobody made.
This guide covers what to look for when reviewing work someone has drafted with AI, how much checking different work needs, and how to give feedback in a way that keeps people telling you when they've used it.
Why AI-assisted work needs a different kind of review
When a colleague writes something from scratch, their errors tend to be visible: a clumsy sentence, a missing section, a figure they flagged as uncertain. AI errors are different. AI tools generate text that sounds right, and they don't signal uncertainty the way a person does. A made-up figure reads exactly like a real one. The term for this is hallucination.
There's also a second risk. If the person who used the tool didn't read the result closely, they may not know what's in it. Your review then becomes the only real check, and you may not have the source material to hand.
How much checking does it need?
Match the depth of review to the risk. A quick internal note needs less than a customer contract summary.
- LOWER RISK Internal, low stakes, easily corrected (a team update, meeting notes for your own team): a read-through, plus a spot check of any names, dates and numbers.
- CAUTION Leaves the team, or informs a decision (a supplier comparison, a briefing for a director, a customer email): check every figure, date and claim against the source, and ask the author to walk you through it.
- HIGH RISK Money, legal terms, people decisions, anything published (quotes, contract terms, anything about an individual's performance or conduct): full check against the source, a second reviewer where the stakes justify it, and confirm AI was appropriate for the task at all.
Alex's supplier comparison is caution level. It will inform a buying decision, so every price and term needs checking against the three quotes.
What to check
Factual accuracy
Are the facts right? Check names, dates, figures, prices, product details and quotes against the original source. Pick the facts that matter most first: in a supplier comparison, the prices and delivery terms; in a policy summary, the rules and deadlines.
A quick check: choose three specific facts and find each one in the source. If any one is wrong, check them all.
Source grounding
Does every claim come from the material the person gave the AI, or has the tool added things from its general knowledge? Grounding means tying an answer to specific source material. AI tools can blend in general knowledge or invent plausible details, especially when the source doesn't answer the question directly.
A quick check: ask "where does this come from?" about any sentence that surprises you. If the answer is "I think the AI added it", it comes out or gets checked separately.
Completeness
Has anything important been left out? AI summaries can drop the awkward caveat, the exception, or the third option that didn't fit the pattern. They tend to smooth over disagreement in the source.
A quick check: skim the source for anything that looks like a condition, exception, risk or disagreement, and see whether it survived.
Tone
Does it sound right for the reader and for your organisation? AI drafts can be over-formal, over-enthusiastic, oddly apologetic or generic. For customer-facing work, a reply that sounds like a template can do damage on its own.
A quick check: read it aloud as the recipient. Would they believe a person at your organisation wrote this to them?
Bias and fairness
Does it describe people, groups or options fairly? AI tools can reproduce stereotypes and one-sided framing from the material they learned from. This matters most in anything about people: job adverts, feedback, descriptions of customers or communities.
A quick check: would the wording read the same if it were about a different person or group? Are some options described with warmer language than others for no stated reason?
Confidential information
Has anything gone into the tool, or come out in the draft, that shouldn't have? Check that the person used an approved tool for the kind of information involved. Check the output too: a draft email to one customer shouldn't mention another customer's pricing.
A quick check: ask which tool and account they used, and what they pasted in. If the answer worries you, see the shadow AI guide. Your organisation's policy and approved tools still take priority.
Assumptions
Has the AI filled a gap with a guess? If the source didn't say something, AI may assume it: that prices include VAT, that delivery is free, that a meeting was agreed. These assumptions often appear as plain statements.
A quick check: look for sentences stating things that would normally need confirming. Ask "did someone actually say that?"
Missing context
Does the work take account of things the AI couldn't know? The tool only knows what it was given. It won't know that one supplier let you down last year, that the MD has already ruled out a particular option, or that a customer is in a dispute.
A quick check: ask yourself what you know about this that isn't in the source documents, and whether the conclusion still holds.
Whether the person understands it
This is the check managers most often skip. If someone can't explain the work in their own words, they can't defend it, correct it or spot when it's wrong.
A quick check: ask them to explain the key point, and how they reached it, without reading from the document. This is a normal part of reviewing anyone's work. It isn't a test of whether they used AI.
The review checklist
Use this for any work that leaves the team or informs a decision. Tick it as you go.
Prefer a checklist of your own? An approved AI tool can adapt this one for a specific kind of work.
I manage a team that sometimes uses AI to draft [type of work, for example supplier comparisons, customer replies or monthly reports]. Write a review checklist of no more than 10 items that a manager can use before this work is sent or used. For each item, name the specific kind of error it catches and a quick way to check it against the source material. Include a check on whether the author can explain the work in their own words. Keep each item to one line. Use UK English.
Why this works: It asks for checks tied to specific errors, so the list is about this work and not generic advice.
The AI can suggest checks. It can't do them for you. An AI asked to check its own work will often confirm it, because it is predicting a likely answer and not looking anything up. The Prove It course explains why, and how to check properly.
Example review conversations
These show the difference between a review that finds problems and helps, and one that finds problems and shuts people down.
Priya reviews Alex's supplier comparison
Priya picks three figures from the comparison and finds them in the quotes. Two match. The third, a delivery charge for Supplier B, isn't in Supplier B's quote at all.
Priya: Thanks for flagging that you used Copilot, that helps me know what to look at. Talk me through how you'd rank these three.
Alex: B comes out cheapest overall, mainly because of the free delivery.
Priya: Where did the delivery charge for B come from? I can't find it in their quote.
Alex: Let me look. You're right, it's not there. I think Copilot filled it in. I checked the unit prices but not the delivery lines.
Priya: Easily done, that's why we check. Can you get the real delivery terms from B tomorrow morning and rerun the comparison? And for anything going to a buying meeting, let's check every line against the quotes, not only the prices.
Priya thanked Alex for disclosing, asked him to explain the work, found the problem by checking the source, and turned it into a specific habit for next time. She didn't criticise him for using AI.
Maya reviews a draft customer email from Leah
Maya Roberts, the Office Manager, reviews Leah's reply to a customer who complained about a late delivery. Leah says she drafted it with AI.
Maya: The facts are right, thanks for checking the order dates. The tone feels a bit off to me. "We sincerely apologise for any inconvenience this may have caused" doesn't sound like us, and the customer was very specific about what went wrong. How would you say it if they rang you?
Leah: Probably just that we're sorry the order was two days late, and that we've changed the courier booking so it won't happen again.
Maya: That's much better. Use that. Keep the structure the AI gave you, it's clear, but put it in your words.
A review that goes wrong, and a better version
What happened: Dan Okafor finds an error in Grace's AI-drafted budget summary. He says: "This is why I don't like people using AI. Just do it properly next time."
The effect: Grace hears that using AI is the problem. Next time she uses it, she doesn't mention it. Dan loses the information he needs to review her work properly.
A better version: "The total in section 2 doesn't match the spreadsheet: it says 48,200 and the sheet says 42,800. Can you check where that came from? When you use AI on anything with figures, check every total against the source before it comes to me. Thanks for telling me you'd used it, that's why I knew to check the numbers closely."
The better version names the specific error, gives a clear rule, and thanks her for disclosing. Grace hears that the checking needs to change. She doesn't hear that using AI was wrong.
Giving feedback without discouraging disclosure
You want people to tell you when they've used AI, because it tells you where to look. Your reaction the first few times decides whether they keep telling you.
Thank people for disclosing. Every time, especially early on. It costs nothing and it shapes behaviour.
Criticise the checking, not the tool use. If AI was allowed for the task, the problem is a missed check. Say that. "You should have checked the delivery lines" is fair. "You shouldn't have used AI" teaches people to hide it.
Be specific. Name the error, show where the right answer is, and give one habit that would have caught it.
Separate a rule breach from a quality issue. If someone put customer data into an unapproved tool, that is a different conversation from a draft with a wrong figure. Handle it through your normal process, calmly, and don't mix it into routine feedback.
Share what you catch. With names removed, near misses make good team learning. "Last week a comparison had a delivery charge that wasn't in the quote" teaches the whole team what to look for.
Don't only reward speed. If the quickest person always gets the praise, people will skip checks. Notice and thank people who catch errors.
Common mistakes
Trusting polish. Well-formatted work isn't necessarily accurate work. AI makes formatting cheap, so it tells you less than it used to.
Checking only what's there. Missing caveats and dropped options are as important as wrong facts, and harder to spot.
Asking the AI whether it's right. It will usually say yes. Check against the source.
Reviewing when the author should have. Your review is a second line. The person who used the tool checks first, and you confirm.
Forgetting decisions about people. Anything AI-assisted that affects someone's job, pay or conduct needs a named human decision-maker who can explain the reasons. The ICO's guidance on explaining decisions made with AI sets out what organisations should consider. Your organisation's policy, HR team and data protection contact decide how it applies to you.
Sources and further reading
This page explains good practice in plain English. It is not legal advice. Your organisation's policy and approved tools take priority.