How to Review AI-Assisted Work Properly

The trap waiting for you

Your team gets faster. Three times as much draft material arrives at your desk. You are conscientious, so you read all of it carefully, and within a fortnight the entire productivity gain has relocated into your evenings.

This is the standard failure of AI adoption at team level, and it is entirely predictable. Generation speeds up. Review does not. If review stays with one person reading everything line by line, that person becomes the constraint, and the pilot you were so pleased with now looks like a personal workload problem.

The way out is not reading faster. It is reading differently: check the things that actually go wrong, sample the rest, and push the first check back to the person who produced the work.

Always verify four things

Whatever else you skim, these four get checked every time, because they are where AI is confidently and specifically wrong.

Numbers. Any figure, date, percentage or total. AI will produce a number that is the right shape and the wrong value, and it never sounds uncertain. Check them against the source, not against whether they look reasonable.

Names. People, companies, products, job titles. Getting a client's name or role wrong in a document that took five minutes to write is the exact outcome that turns senior enthusiasm into a ban.

Claims. Anything stated as fact: a rule, a policy, a precedent, a quote, a citation. If a sentence would need a source to defend it, it needs a source.

Commitments. Dates, deliverables, prices, guarantees, anything beginning with we will. AI writes commitments cheerfully. Your organisation has to keep them.

Sample the rest

For everything outside those four categories, review a proportion, not the whole. A workable pattern for a new use case: review everything for the first week, then a third, then roughly one in five once the error rate has settled. If you find a serious problem, go back up a level for a fortnight.

Choose the sample deliberately. The most useful items to pull are the longest, the ones going to the most senior or most external audience, and the ones produced by whoever is newest to the tool. Random sampling is fair. Targeted sampling finds more.

Below is a draft, then the source material it was written from.
Extract every number, date, name, factual claim and commitment in the draft.
For each, mark it as one of:

MATCHES the source
CONTRADICTS the source (say what the source actually says)
NOT IN the source
Return a plain list, most serious problems first.
Say nothing about tone, structure or writing style.
DRAFT:
[paste it]
SOURCE:
[paste it]

Push the first check back

Your team should not send you anything they have not already checked themselves. That sounds obvious, and it stops happening the moment output becomes cheap, because a five-minute draft does not feel like it warrants a five-minute check.

Make the expectation explicit and boring. Nothing reaches you unchecked. If a figure in a draft is wrong, the person who sent it is the person who missed it, not the tool. This is not harshness, it is the only version of the arrangement that scales past you.

Give them the checking prompt above. A person with a good checking routine is worth far more to you than a person with a good drafting routine.

I am about to send this draft to my manager. Before I do, act as a hostile reviewer.
Find:

Any number, date or figure I have not evidenced
Any name or title that might be wrong
Any sentence that promises something on behalf of my organisation
Any claim that would need a source if challenged
List only problems. Do not praise anything. Do not rewrite it.
DRAFT:
[paste it]

Feedback on work a person part-generated

This is the genuinely awkward bit. When someone hands you a document they wrote, your feedback is about their thinking. When they hand you something they prompted, edited and approved, feedback aimed at the writing can feel oddly misdirected on both sides.

The move that works is to shift the subject from the prose to the process. Not this paragraph is vague, but what did you check before sending this. Not the AI got this wrong, but the figure in paragraph three did not match the source, so what would catch that next time.

That framing does two useful things. It keeps accountability with the person, where your ground rules already put it, and it improves the checking habit rather than just fixing one document.

A team member sent me AI-assisted work containing this problem:
[what was wrong, and what the source actually said]
Draft what I say to them in under 80 words. Rules:

Focus on the checking process, not on their writing ability
Ask what they checked before sending
End with one specific habit that would catch this next time
Do not be passive aggressive, and do not soften it into meaninglessness

Keep a short private tally of what you actually catch over the first month. If the same category keeps appearing, that is a training gap or a prompt problem, not a character flaw, and it usually takes one conversation to fix. If nothing serious appears for several weeks, cut your sampling rate and take the evening back.