When building is cheap, research decides what gets built

AI collapsed the cost of building. The new risk is shipping the wrong thing at speed. The question holding teams back is no longer whether they can build something. It's whether they should, and that's a research question.

So research moves to the centre of the job. Its purpose sharpens into two things: deciding what to build, and validating that it worked. Everything else is decoration.

Running research without a research department

Small teams rarely get dedicated researchers, and that's fine. Research works as part of every designer's job when the mechanical layer is handled for them.

AI takes the mechanics: transcribing sessions, clustering feedback, first-pass synthesis, drafting discussion guides. People keep the parts that need judgment: choosing the question, sitting with users, and deciding what a finding actually means for the product. Split it that way and a team of three sustains a research practice that used to need a department.

The most expensive research is the research you repeat

Ask around any company that's a few years old and you'll find the same study run twice. Same question, a year apart, paid for in full both times. Nobody planned that. The first study's findings were in a deck, the deck was in someone's drive, and that someone changed teams.

Research that can't be found doesn't exist. The fix is structural: findings need a permanent home, and the home has to be built for retrieval, by people and by AI.

Build the repository like you'll query it

Four rules make a research repository actually usable.

Store findings, not decks. Break every study into individual findings. One claim, its evidence, the date, and how confident you are. Decks are where findings go to die.

Keep it in plain text with fixed fields. The same structure for every entry, in a format machines read as easily as people do. Fancy tools come and go. Text is forever.

Summarise at every level. One line per finding, one paragraph per study, and a master index listing everything. Whoever is looking, human or AI, navigates the index first and drills down second.

The index also keeps the cost of asking down. AI is billed by the token, on what it reads as much as what it writes, so pointing it at a whole repository means paying for it to read studies that were never relevant. With an index of one-liners it scans a page, picks the right study, and pulls the full detail only when the question needs it. Queries cost what the question needs, however big the archive grows. Answers sharpen too, because models get worse at finding things the more they have to read.

Date everything and let findings expire. What was true about your users two years ago informs. What was true last quarter decides.

Ask before you commission

Then one standing rule makes it compound: no new study starts before querying what's already known. With a repository built like this, that query is a plain question asked of an AI, answered in seconds with sources attached.

Past research becomes the first participant in every decision. Questions that used to trigger a six-week study now start from an answer you already paid for. Each study makes the next one cheaper, which is the closest thing to free velocity a lean team can get.