There’s a question every research plan is quietly built to avoid, and it’s the simplest one: how many interviews is enough? Five? Twelve? Thirty? Ask a room of researchers and you’ll get a shrug dressed as a rule of thumb — “five gets you most of the usability problems,” “twelve is where themes stabilize” — numbers that sound like findings but started life as budgets. The honest answer to how many has been known for fifty years, and it isn’t a number at all. You keep going until you stop hearing new things.
That idea has a name in the trade — saturation — and it’s less clever than it sounds. Interview enough people about the same thing and the new material dries up. The first conversation is all new. The second repeats some of the first and adds a little. By the tenth you’re mostly nodding along to things you’ve already written down. Plot the count of genuinely new themes against the number of interviews and you get a curve that everybody in qualitative research can draw from memory: steep at the start, bending, and then flat. Saturation is that flat part. It’s the place where the next interview costs you the same twenty minutes and hands back nothing you didn’t have.
The old answer was a budget wearing a lab coat
Here’s the uncomfortable thing about saturation: for most teams it was never actually reached. It was declared. You ran five interviews — not because the fifth taught you nothing new, but because the fifth was where the calendar ran out. Then, because “we stopped at five” sounds arbitrary and “we reached saturation” sounds rigorous, the budget quietly borrowed the lab coat. The stopping point was set by recruiting and scheduling and the moderator’s week, and the research language was draped over it afterward.
We’ve made this argument before from the cost side: the cap on research was never curiosity, it was the calendar. Saturation is where that cap does its quietest damage. When each interview is expensive, you can’t run past the point of diminishing returns to confirm you’ve reached it — the whole region beyond “enough” is exactly the region you could never afford to visit. So you never saw the curve go flat. You saw it climbing, you ran out of money, and you told yourself it had probably levelled off somewhere just past where you stopped. Sometimes it had. Sometimes you were still on the steep part and calling it a summit.
Saturation you can’t afford to overshoot isn’t a finding. It’s a guess with good posture.
Cheap interviews turn the curve into something you can watch
Drop the cost of an interview far enough — no scheduling, no moderator in the room, no write-up done by hand — and the region past “enough” stops being off-limits. That’s the shift AI User Interviews is really for. Not “run a hundred instead of five” as a vanity number, but run until the new-theme count actually hits zero and stays there, and then keep running a little longer to prove it stayed. For the first time the flat part of the curve is somewhere you can go stand.
And once you can go there, saturation stops being a hunch you defend in a readout and becomes a shape you can point at. You don’t argue that you’ve probably heard enough; you show the line climbing, bending, and flattening, with the last three interviews adding nothing — and the flatness is the evidence. The question “did we talk to enough people?” gets an answer with a picture attached, which is a different kind of answer than a number someone chose because it was Thursday.
What changes when “enough” is observable
The obvious win is that you stop under-sampling — no more five-and-pray, no more shipping a decision built on a curve that was still rising when the money ran out. But the subtler win is the other direction: you stop over-sampling out of anxiety. Teams that can’t see saturation tend to either quit too early and hope, or keep booking interviews long past the point of learning anything, because “one more, to be safe” feels responsible when you can’t tell whether it’s necessary. A visible flat line ends both. It tells you when you’ve heard enough and when you’re wasting your time, which are the same fact read from two directions.
There’s a quieter consequence too. When enough is a shape rather than a target, you stop picking the number in advance. You don’t commit to “we’ll do twenty” and then discover at twelve that you were done, or at twenty that you weren’t. You run, you watch the curve, and the data tells you where to stop — which is how it was always supposed to work, and never could, because watching the curve meant paying to render the part of it you already suspected was flat.
None of this makes the questions any better; that’s still your job, and a bad question saturates just as neatly as a good one, on nonsense. What changes is that the decision to stop moves out of the budget and back into the research, where it belonged the whole time. You stop when the shape tells you to. And now you can see the shape.