Every research platform now has an AI moderator, or is shipping one this quarter. There are more than fifteen tools in the category as of 2026, and the pitch is usually some version of the same promise: qualitative depth at survey speed, with nobody running the call.
Most of that will keep failing for about another ten years. Models keep getting better and it will not help much here, because what breaks is on the participant's side of the call.
There is a real job here and the tools do it. If you need to hear two hundred people react to three packaging concepts, an AI moderator is a better instrument than the open text field at the bottom of a survey. It asks follow-up questions, it runs all two hundred sessions at once, and results come back in a day instead of a month.
That job is breadth, and it used to be done badly by surveys, so nobody should miss the old way.
In a real interview there is a moment, usually ten or fifteen minutes in, where the person decides you are worth being honest with. Before it you get polite answers that don't really mean anything, and after it you get what actually happened, including the parts that make them look bad.
That decision is about the human being on the other end, so you can write a perfect guide and still collect polite answers for forty minutes if the person didn't open themself to you. Nobody opens themself to a bot. A senior employee experience researcher we interviewed in May said it as a participant rather than as an expert: he would sign up for an interview, find out it was AI moderated, and drop off the call.
There is also nothing to lose by being lazy. With a person in the room, a thin answer has a price: someone asks for an example, and the participant hears that what they said was not enough. Take the person away and that price goes to zero, so people answer just enough to get to the next screen, and they make things up more comfortably, because making things up to software does not feel like lying to anyone.
This gets worse over time, because the more normal it becomes to talk to machines, the less effort anyone spends on a single conversation with one. Models will keep improving through all ten of those years, and the people being interviewed will still want a person on the other end.
A participant says yes and means maybe. The same researcher told us that this is exactly what a transcript loses: the text is flat, the participant says yes, and how they said it is gone.
Models are getting better at hearing that, but what matters more is what the moderator does with it. A human hears the hesitation, drops the remaining eight questions, and spends the rest of the call on the thing that just cracked open, which is a decision to change the study while it is running and to carry the blame if it was the wrong call.
A platform is built to finish the guide and throwing it away is not a move it is allowed to make.
Recruiting is where the money goes. Published 2026 benchmarks put a sixty minute session with a B2B professional at $150 to $500 in incentives, the fee a participant is paid for their hour, plus $50 to $150 in recruiting costs to find them. Senior decision makers run $400 to $750 and up, while a researcher's hour, fully loaded, runs $80 to $150.
So the participant costs more per hour than the researcher, often several times more, and putting a bot in front of a $600 participant to save $120 of researcher time is optimizing the cheap side of the ledger.
The arithmetic flips on cheap consumer panels, where a participant costs $50 and you are running two hundred of them, which is exactly where AI moderation already belonged.
This is a large part of why Zernote helps the researcher run the call instead of running it for them. Our first value says it plainly:
"Augment the researcher, never replace them. Research is a craft. Our job is to carry the cognitive load."
Zernote works on both ends of that. During the call it listens along and suggests follow-ups, and after the call it turns the recording into something you can question directly in Claude, but the person asking the questions is still a person.
write to me at mironbeneval@zernote.com.