The first week of a project used to look the same every time. Read everything the client sent. Look at ten competitors. Sit through a stack of interview recordings, write sticky notes, group them, and on day five present a problem statement that was mostly right.
That week now takes about two days. Not because I think less, but because an AI tool does the reading, sorting and first rough grouping while I do the parts that need judgement.
What I hand over
I give AI three specific jobs.
- Desk research. I share the brief, the product’s help centre and a list of competitors. I ask for a side-by-side comparison: how each one handles the journey we’re working on, where they ask people to commit, and what they hide. It’s a starting map, not a finding.
- Making sense of interviews. Transcripts go in with the same instructions every time: pull out every pain point, quote it word for word, and note who said it and when. Then group the pain points into themes.
- Competitor walkthroughs. Screenshots of ten checkout or sign-up journeys, annotated step by step. Tedious for a person, quick for AI, and easy to check.
On a recent fintech onboarding project, twelve interviews went from raw transcripts to a board of themes in an afternoon.
- 2 days for what used to fill week one
- 12 interviews sorted into themes in an afternoon
The checks I now run
The catch is that AI is confident in the same way a junior researcher is on their first project. It fills gaps. Early on, a summary gave me a neat quote about “not trusting the identity check” that nobody had actually said.
It was a believable mix of two different people, stitched together and put in quotation marks. So I changed my process, not the tool.
- Every quote needs a timestamp If I can't jump to it in the recording, it doesn't go on the board.
- I read three transcripts myself Picked at random, before I look at the AI's themes. If we disagree, I trust my read and ask again.
- Competitor claims need a screenshot Or they get deleted. AI describes features from two redesigns ago, or ones that never existed.
- Counts get recounted "Most participants" means nothing until I know if it's seven of twelve or three.
This sounds like it eats the time saved. It doesn’t. Checking is faster than doing, and the checking is where I actually form a view.
What changed in me
I spend less time as a filing clerk and more time arguing with a draft. The problem statement I present at the end of week one is sharper. I’ve had three extra days to poke holes in it rather than piece it together.
The risk is getting lazy with the checks once the results look good for a few projects in a row. They always look good. That’s the problem.
The takeaway
Use the time AI gives back to think harder, not to move faster.
Let the AI do the reading. Don’t let it do the believing.