Using AI at work is not risky because the output is bad. It is risky because the output is fluent, and fluency reads as correctness. The fix is a short, boring review pass applied every time, targeted at the specific things that go wrong.
The five predictable failures
| Failure | What it looks like |
|---|---|
| Invented specifics | A confident citation, statistic, case number or quotation that does not exist anywhere. |
| Stale knowledge | Prices, versions, rules and people that were true at training time and are not now. |
| Silent arithmetic errors | A total that does not match the rows above it, stated with complete confidence. |
| Agreeing with your premise | You asked a leading question and it built a case for it rather than challenging it. |
| Smoothing away the caveat | The nuance that made the source correct is dropped in the summary. |
The two-minute pass
- Check every number. Recompute totals and percentages yourself. Never paste a figure you have not verified.
- Open every source. If a link, paper, section or clause is cited, open it. An unopened citation is not a citation.
- Test the claims you would be embarrassed to be wrong about. Names, dates, prices, legal and medical statements, anything about a real organisation.
- Read it once as the audience. Not for accuracy, for whether it says anything. Fluent text can be entirely empty.
- Ask what is missing. Models rarely volunteer the risk, the counter-argument or the thing that makes the answer harder.
"Which parts of this are you least confident about, and what would change the answer?" often surfaces the weak points faster than checking line by line.
Raise the odds before you check
- Give it the source material rather than relying on recall. Summarising a document you supplied is far more reliable than answering from memory.
- Ask for the reasoning alongside the answer, so you can see where it went wrong rather than only that it did.
- For anything factual, ask it to say plainly when it does not know.
- Run an important answer twice in separate sessions. Two very different answers mean neither is reliable.
What never leaves your hands
Some things are yours regardless of how the draft was produced: the final judgement, anything sent under your name, and any decision affecting a real person. Treat the model as a fast first-drafter with no accountability, because that is exactly what it is.
Also check what you put in. Client data, personal information and anything under a confidentiality agreement should not be pasted into a tool without knowing where it is stored and who can see it.
Key takeaways
- Recompute every number and open every citation.
- Supplying the source beats relying on the model's memory.
- Ask which parts it is least confident about.
- Check what you paste in as carefully as what comes out.
Ready to automate more than one task?
The AI Productivity & Career Accelerator builds the workflows, the judgement and the portfolio together.
Explore the course