How to Check Work You Did Not Watch

Your name goes on it

Delegating to a person and delegating to AI differ in one important way. A person who is unsure usually says so. They hedge, they ask, they flag the bit they guessed at. An assistant hands you the guess in the same tidy sentences as the facts, with no change in tone whatsoever.

So the review is yours, and it has to be a real one. The good news is that reviewing finished work takes about three minutes when you know where to look, against the forty it would have taken to produce it. That is still an extremely good trade.

Four passes, three minutes

Do these in order. Each one takes under a minute and they catch different things.

Pass one: does it answer the brief? Read your own instruction, then the output. Not "is this good" but "is this the thing I asked for, for the audience I named, in the format I specified". A lot of impressive output quietly answers an adjacent question.

Pass two: are the facts real? Every name, date, figure, statistic, quotation and citation. If you cannot trace it to something you supplied, treat it as unverified until you check it. This is where the damage lives.

Pass three: what is missing? Hardest and most valuable. Wrong sentences announce themselves. Absent ones do not. Compare against your brief and against what the reader will need.

Pass four: what is confidently wrong? Look specifically at the smoothest, most authoritative passages. Fluency is not evidence. The sentence that sounds most like it came from an expert is often the one that came from nowhere.

Make it audit its own work

You can make the review much faster by asking for the map before you read the territory. Do this in a fresh message once the work is done.

Before I read this properly, answer these about the work you just produced:

Which facts, figures, names or dates did you take from what I gave you?
Which ones came from your general knowledge, and how confident are you in each?
What did you assume because I had not told you?
What did you leave out that a careful reader might expect?
Which single part of this is most likely to be wrong?
Be blunt. Do not defend the work.

The parts you always read yourself

Some things do not get delegated, no matter how good the output looks.

Anything with your name signed to it goes through your eyes word by word. Anything going to a customer, a regulator or a court. Anything containing a commitment: a price, a date, a guarantee, a scope. Anything about a specific person, especially performance, pay or conduct. Anything a lawyer, an accountant or a doctor would normally be paid to say.

For everything else, spot check proportionally. A weekly internal summary needs the four-pass skim. A proposal worth six figures needs every line.

Here is my original brief and the work you produced.
Ignore what is there. Tell me only what is missing.
What did the brief ask for that the output does not deliver?
What would a sceptical reader immediately ask that this does not answer?
What context is a reader missing to act on this?
List them. Do not rewrite anything yet.

Read this as the person most likely to disagree with it.
That person is [a sceptical finance director / an annoyed customer / a rival team].
Give me the three strongest objections they would raise,
the weakest claim in the document and why,
and anything here that could be read in a way I did not intend.
Be harsh. Do not soften it.

List every number in this document as a table: the number, what it claims,
and where it came from (my source, your general knowledge, or calculated).
For anything calculated, show the arithmetic.
Mark anything you cannot trace with CHECK THIS.

Reviewing is a skill, not a chore

The instinct after a good handover is relief, and relief makes people skim. Build the three minutes into the job so it never feels optional. The habit you want is that finished work goes into a review, not into an email.

Do that consistently and you get the real prize: you can hand over more, because you trust your own checking rather than trusting the machine. Those are very different things, and only one of them is safe.