Checking AI work when you are not the expert

Reading well is not being right

The uncomfortable part of workplace AI is not the obvious mistake. It is the answer that arrives in clean prose, in the right register, with the right amount of hedging, about something you have no way to check. A contract clause. A tax treatment. A dosage. A regulation number. A line of SQL that will run against production.

Fluency and accuracy are produced by the same machinery, so they arrive together whether or not the second one is real. That is the whole problem in one sentence. A person who cannot audit the content is left auditing the tone, and tone is exactly the signal that stopped carrying information.

You do not need domain expertise to do better than that. You need a small set of checks that operate on the shape of the answer rather than its substance, and the discipline to run them before the answer leaves your hands.

Check the shape before the substance

Ask what kind of claim this is before asking whether it is true. Claims come in a few families, and each family has a different failure mode. Confusing them is how a plausible paragraph becomes a commitment nobody meant to make.

• ✓A quoted fact from a document. Checkable by anyone. Open the document. If the assistant cannot name the document, you have a guess wearing a citation.
• ✓A number. Ask where it came from, what period it covers, and what it is a share of. A number without a denominator is a mood.
• ✓A rule or requirement. These are jurisdictional and dated. A rule stated without a source and an as-of date should never be repeated in your own voice.
• ✓A recommendation. Not true or false, so do not check it for accuracy. Check it against the constraint it was supposed to satisfy, which the model may never have been told.
• ✓A summary. The failure is not error, it is omission. Compare the length of what went in with what came out and ask what was dropped.

Three questions that need no expertise

These work in any field, including ones where you could not spot a wrong answer if it were printed in bold.

First: what would make this wrong? A useful answer can name its own failure condition. If the assistant cannot say what fact would overturn its conclusion, it has not reasoned about the conclusion, it has produced text that resembles one.

Second: what is this based on? Not a link, which can be invented, but a document you can open, a system you have access to, or a person who owns the answer. If the chain ends inside the model, you are the only evidence.

Third: who pays if this is wrong, and when would they find out? A wrong internal draft costs an hour. A wrong figure in a board deck costs a quarter of trust. A wrong compliance answer costs something you cannot price. Effort should scale with that, not with how interesting the question was.

Ask for the parts you can check

You can change what the assistant hands you so that more of it is checkable by a non-expert. This is a prompting skill, but not the kind that gets sold in courses. It is asking for a shape you can inspect.

1. 01Ask for the answer and the source separately, so a missing source is visible rather than dissolved into the prose.
2. 02Ask it to mark what it is confident about and what it inferred. Then treat the second list as unwritten.
3. 03Ask for the question you should have asked. In unfamiliar territory that is often worth more than the answer.
4. 04Ask what a person who knows this field would check first, then take that to a person who knows this field.
5. 05Ask it to argue the opposite case. A conclusion that survives that is worth more than one that was never pushed.

Saying you do not know is a professional move

The pressure that produces most bad AI-assisted work is not laziness. It is the feeling that you are now expected to have an answer, quickly, on something you would previously have escalated. The assistant removed the excuse for not knowing, so people stop saying they do not know.

Restore the sentence. It is short and it works: this is what the draft says, here is the part I could verify, here is the part I could not, and here is who should look at it. That is a complete piece of professional work. A confident paragraph with an unverified rule in the middle of it is not.

The version of this that gets people in trouble is quieter than a lie. It is repeating the model's framing in your own voice, in your own email, until nobody downstream can tell which claims were checked. Keep the seam visible.

Where Datum fits

All of this works with the tools already in your job and no product at all. Datum exists for the weeks when you want reps instead of resolutions: short workplace situations where the plausible answer is wrong, a stop you have to place yourself, and a written call you can reopen later. Core is free, nothing expires if you skip a day, and the vault Datum stores your files in is one it cannot read.