When a study runs late, almost nobody blames synthesis. They blame recruitment. The participants were hard to find, the screener was too tight, legal took a week to approve the incentive. All of that is real, and all of it is visible, because recruitment happens in a shared calendar where everyone can watch it not happening.
Then the last session ends, and the study disappears for a fortnight.
That fortnight is usually the longest single stretch in the whole project, and it is the one nobody schedules. Sessions have participants waiting, so they hold their slot. Synthesis has nothing external forcing it, so it absorbs every delay upstream of it and then quietly adds its own.
What the fortnight is actually made of
If you break it down, it is rarely one big task. It is a pile of small ones that each seem too minor to mention in a status update.
- Rewatching sessions two, five and nine, because something was said that now seems to matter.
- Fixing transcripts that got names, product terms and numbers wrong.
- First-pass coding, which is genuinely slow and genuinely dull.
- Discovering that two researchers coded the same behaviour under different names, and reconciling them.
- Working out which of eleven candidate themes are actually findings and which are just things people said.
- Building a deck, then rebuilding it once someone senior asks a question the structure cannot answer.
Nobody signs off on a fortnight for that list. It just happens, one afternoon at a time, and by the end of it the decision the study was meant to inform has usually been made without it.
Where the time can genuinely come out
This is the part of the process where AI has changed the arithmetic, and it is worth being specific about which part, because the vendor pitch is usually much broader than the reality.
Transcription and cleanup stop being a task at all. That is the easy one, and most teams already have it.
First-pass coding across every session at once is the significant one. A researcher coding twenty sessions works sequentially and gets tired; the themes found on day one and day three are not coded to the same standard. A first pass across the whole set, generated in minutes, is imperfect but consistent — and consistent-but-imperfect is a much better starting point for a human than sequential-and-uneven.
Retrieval across old studies is the one nobody expects. Most enterprises are sitting on three or four years of research that no one has time to reread. A new study's themes can be checked against that archive in an afternoon, which sometimes reveals you already answered the question two years ago, in a different team, under a different name.
Consistent but imperfect is a better starting point for a human than sequential and uneven.
The failure mode, and the check that prevents it
The risk is obvious and worth naming plainly: a fluent set of themes that nobody has traced back to what people actually said. It reads well, it survives the readout, and it falls apart the first time a sceptical stakeholder asks which sessions a claim came from.
The rule I work to is simple. Every theme carries the session identifiers it came from and at least one verbatim quote. If a theme cannot produce those, it is not a theme yet — it is a hypothesis, and it gets labelled as one. This takes almost no extra time when done as you go, and it is impossible to reconstruct afterwards.
It also changes the conversation in the readout. A stakeholder who can be shown the three sessions behind a finding argues with the finding. One who cannot argues with the method, and that argument is unwinnable.
A check to run on your last study
Find the date of your last session and the date the findings were shared with the people who needed them. If the gap is more than two weeks, that gap is the most improvable number in your research practice — more than sample size, more than tooling, more than headcount.
It is also the number your product team is quietly routing around when they decide not to wait for you.
If your team is working through this, tell me how it's going.
← All notes