Choose by question type, not by budget. AI moderation suits briefs where the questions are known and the value is in n — pricing checks, feature ranking, channel confirmation. Human moderation earns its cost where the useful answer only surfaces because someone noticed what the expert was steering around.
What does an AI moderator do well?
It runs a discussion guide without drift. Across forty conversations a human moderator will phrase question six slightly differently forty times, and some of those phrasings will lead the answer. An AI moderator does not, which makes the resulting set genuinely comparable.
It also adapts. A competent AI interviewer will follow up on an incomplete answer, ask for a number when given a range, and return to a skipped question — the mechanical parts of moderation, done consistently.
What it does not do is decide that the guide was wrong. An AI moderator will run a badly framed question forty times with perfect consistency, which produces a clean dataset built on a flawed premise. A human running the same guide would usually notice by the fourth conversation that the question was landing oddly and say so. Consistency is only an advantage once the guide is right.
| AI moderated | Human moderated | |
|---|---|---|
| Strongest at | Consistency across sessions, and n at speed | Noticing what is being avoided |
| Best question type | Known questions — pricing, ranking, channel confirmation | Discovery, sensitive topics, contested markets |
| Real cost driver | The expert's fee, which does not change | The moderator's hour, plus the expert's fee |
What does a human moderator catch?
Evasion. When an expert answers a slightly different question from the one asked, or gives a fluent answer in a register that has changed from the previous ten minutes, that is information — and it is the signal a text-adaptive system is least equipped to read.
The findings that justify commissioning primary research at all tend to come from exactly there. An expert who will not say something directly will usually indicate that there is something not being said, and a human hears it — one of the reasons for putting a moderator on the call at all.
Which questions suit which method?
Known questions, large samples, comparability: AI. Pricing sensitivity, feature ranking, channel confirmation, anything where you already know what you need to establish and the value is in establishing it across enough people to be confident — the same territory an expert opinion survey covers.
Unknown questions, sensitive topics, contested markets: human. Discovery work, where the point is to find out which question matters, is not substitutable — you cannot instruct a guide to look for something nobody has identified yet.
Can you run both on one brief?
Yes, and it is usually the right answer. Run AI moderation wide to find where the variance sits, then put a human moderator on the handful of conversations where the variance turned out to be interesting. That is also how a forty-interview program stays inside three weeks.
The sequencing matters. Running the human interviews first tends to anchor the guide on what those few people said, which is precisely the bias the larger sample was supposed to correct.
What does each actually cost?
Per session, AI moderation is cheaper — the moderator's hour is not in the price. Per finding, the picture is less clear, because the expert's fee dominates a B2B interview cost and does not change with who asks the questions.
That is why the economics published by consumer-research tooling do not transfer. When a panel respondent costs fifteen dollars, removing the moderator is most of the cost. When the respondent is a former operator billing several hundred an hour, it is a fraction.