Research Did Not Get Automated. It Got Easier to Skip.

The most methodical research plan I ever wrote was for a university website. Thirteen to fifteen participants recruited by segment, prospective students, alumni, business, government. Ten task scenarios across the four things people actually came to the site to do. Moderated sessions, recorded. Card sorting to settle the information architecture. A medium fidelity prototype refined between rounds. Then the same measures repeated once the site was live, so we could say whether it had worked rather than whether it had launched.

Nothing in that plan has been made unnecessary by AI. Several parts of it have been made faster. The temptation the tools introduce is to skip the parts that are still slow, because the fast parts produce something that looks like research.

What got faster

Synthesis got faster, and I am glad of it. I can put twelve interview transcripts in front of a model and get a first pass at the themes in minutes, which used to be two days with sticky notes on a wall. I still do the wall, because reading the transcripts myself is where the understanding comes from, but the model is a good second reader. It catches the quote I forgot and it does not get tired.

Personas got better, not because the model writes them but because the data behind them is easier to join. At Betfred we built personas from transactional, demographic and behavioural data together, so a persona was not a poster but a definition the platform could act on, and the homepage rendered differently for James the sports bettor than for a casino player. That took a data team months. The joining is now the easy part. The judgment about which behaviours matter is not.

Competitive audits got faster. Feature by feature comparison of a staffing site against its two largest rivals, sixty criteria across registration, profile, search and apply, was a week of work in 2017. It is an afternoon now, and the afternoon is better spent on what the audit means.

What did not

The people you did not talk to are still the people you did not talk to. When a client asks whether the model can just generate the users, the honest answer is that it will, fluently, and the result will tell you what a model thinks people like that say, which is the average of everything ever written about them. It will not tell you that the tradespeople ordering from a plumbing catalogue wanted filters by pack size before anything else, because that came from watching one of them order on a Tuesday morning for a job that afternoon. It will not tell you that a broker's real pain was not the system but the client who would not wait for the quote. Those came from the room.

Usability testing did not get automated either, and for AI products it got harder. When the product is an assistant, the test is not whether a person can find the button. It is whether the answer was right, whether it was right for them, and whether they trusted it enough to act. The research data is the transcript. I read hundreds of them for the insurance assistant, and I built an evaluation corpus from them, questions with known good and known bad answers, scored for false positives, run again every time the prompt changed. That is user research by another name. It is the most valuable research I did on that product, and no tool did it for me.

The two questions

I ask two questions of any research plan now. First, what do we know because we watched someone do it, as opposed to because we read about people like them? Second, when the product is live, how will we measure the same thing again? The second question is the one that separates research from decoration. The university plan repeated its task measures after launch. Most plans I see do not, and AI has done nothing to change that; it has only made the first round look finished sooner.

The tools are good. I use them daily and I would not go back. But research is the part of design where the cost was never really the hours. It was the willingness to find out you were wrong, and that has not been automated.