Machines are good at a lot. People are good at the rest. We use language models every week, and we've become quite firm about where they belong.
Where they save time
- First drafts. Interview guides, workshop plans and test plans. A draft that's a bit wrong is faster to fix than a blank page.
- Variations. Twenty ways to word a button, or ten angles on a concept. We choose, the model generates.
- Structure in large amounts of text. A first sort of open survey answers, or suggested themes across notes.
- Prototypes. It has never been faster to make a working prototype to test with.
Where they don't belong
- Instead of your users. A model can tell you what people in general might say. It knows nothing about your users, in your situation. Synthetic users are a supplement at best.
- Judgement. What matters most, what we leave out and what we stand for is still our job.
- The final version of the text. Models write fluently but generically. We write the final text ourselves, especially in Norwegian.
Three rules we follow
No personal data in without an agreement. Interviews and user data don't belong in a tool that isn't approved for it. Check the agreements before you paste.
Check everything that looks like a fact. Models invent sources, numbers and quotes with great confidence. Anything that goes further gets checked against the original.
Show where AI was used. When an analysis had help from a model, we say so. That way the people using it can judge it properly.
The main thing
AI makes us faster at what used to be tedious. We spend the time we save on what actually moves things: talking to people, watching them use the solution and making hard choices together with our clients.