Why AI 'makes things up' at work, and how to stop it
Most people who've given up on AI at work have the same story. They asked it something about their own business, it answered confidently, and part of the answer was invented. A figure that didn't exist, a policy that wasn't theirs, a customer name from nowhere. After that, why trust anything it says?
That reaction is reasonable. But the invented answer usually tells you more about the request than about the tool.
Why it guesses
A chat assistant is built to give you an answer. If you ask about something it hasn't been shown, like your sales last week, your refund policy or what "the usual format" means in your office, it has two options. It can say it doesn't know, or it can produce something that sounds right. Unless you've told it otherwise, it often does the second.
So "it made something up" usually means "it was asked to fill a gap, and it filled it." A trainee in their first week would do the same if you said "just do the Monday report" and walked off.
The fix: a context pack
A context pack is one page you give the AI before you ask it to do the job. You write it once and reuse it every time the job comes round. It has four parts.
1. The facts it needs. Paste in or attach the actual material: the sales export, last month's version of the report, the policy document. If it isn't in the pack, assume the AI doesn't know it.
2. What "good" looks like. One example of a finished report you were happy with, plus a line or two on format: "Three headings. Figures in pounds, no pence. Under one page."
3. The rules. Things a new starter would get wrong. "Exclude internal transfers." "Region names are North, South and Online, nothing else." "Never estimate a figure."
4. Permission to say "I don't know". This is the line most people leave out, and it matters most:
If a figure or fact isn't in the material above, don't guess. Write MISSING and say what you'd need.
That one instruction turns a confident wrong answer into a visible gap you can fill.
What a context pack looks like
For the weekly sales summary, the whole pack might be:
Job: draft the weekly sales summary for Monday's managers' meeting. Material: this week's sales export (attached) and last week's summary (attached) as the format to follow. Rules: use only figures from the export. Compare each region with last week. Flag any region that moved by more than 10%. Exclude internal transfers. If unsure: don't guess. Write MISSING and say what you need. Output: under one page, three headings: Headlines, By region, Things to check.
That's it. It takes ten minutes to write the first time, and you only change the attachments after that.
You still check it
A context pack doesn't make the AI right every time. It makes the draft checkable. Because every figure has to come from the material you gave it, you can check the draft against that material in minutes, rather than wondering where a number came from.
Treat it like a trainee's first draft: read it, fix what's wrong, and add a rule to the pack each time you spot a repeat mistake. The pack gets better every week, and so does the draft.
Where are you now?
Not trusting the output is a common reason people stop using AI at work, and the context pack is one part of the fix. The readiness check below takes about three minutes and shows which part of handing work to AI is most likely to trip you up.