When AI answers are frustrating, is it usually a context problem?
Usually, yes. But context is really three different things – what the tool can see, what you told it, and what it was asked to do – and each one fails differently. Fixing one can make another worse.

Three kinds of context
| Kind | What it sounds like | What's going on | The fix |
|---|---|---|---|
| Material | "It made up a price." "It doesn't know our process." | The tool can't see the document that holds the answer, or it's seeing an old one. | Give it the material, in one place, kept current. The library. |
| Instruction | "That's not what I meant." "It's too generic." | The ask was shorter than you'd give a person. | Write the ask as you would to a new starter: goal, material, format, an example. |
| Task | "It got the totals wrong." "It forgot what I said yesterday." | The task is one the tool's weak at, or it needs memory it was never given. | Change the task: break it up, give it a tool for the arithmetic, paste in yesterday. |
| Model | "It's just not very good at this." | Genuinely the model, once the other three are fixed. | Try another model, or a stronger one, for that task only. |
Most people reach for the fourth row first. It's usually the last one that applies.
The new-starter test
Before blaming the tool, imagine handing the same request, with exactly the same material, to a bright person on their second day. "Draft a reply to this" with no price list, no history and no idea what you want to say would get a generic reply from them as well. If the new starter would need to ask three questions first, the tool needed those three answers too and didn't get them. In our experience this settles most arguments about whether the technology's any good.
Why more context doesn't always help
Giving the tool everything is usually the next mistake. A drive full of old versions, half-finished drafts and three price lists that contradict each other gives you answers that are confidently wrong, because the tool found something relevant and it was the wrong something. Context works when it's the right material, current, and not drowned out. We find that removing documents improves answers more often than adding them does.
The habit that fixes most of it
Write the ask the way you'd brief a person. It has four parts, a sentence or two each:
- What you're trying to achieve, and who it's for.
- The material it should use, attached or named.
- The format you want back, and how long.
- One example of a good answer, if you've got one.
It takes a minute longer than typing "summarise this", and it pretty much removes the second row of the table. Teams that pick up this habit stop complaining about the tool within a fortnight, in our experience – which tells you where most of the frustration was coming from.
When it genuinely is the model
Long chains of arithmetic, counting things in a big document, anything that needs careful logic over many steps, and very recent events are where the model itself falls short. For those, break the task into pieces, give it a calculator or a spreadsheet rather than asking it to do sums in its head, or switch to a model built for reasoning for that one task. Don't switch everything because one task is hard.