Outset

Qualitative research has always been rationed by a single constraint. A skilled moderator can run perhaps six or eight in-depth interviews in a day, and each one produces an hour of recording somebody then has to analyse.

Everything about the practice follows from that, which is why so much customer understanding gets outsourced to listening tools such as Brandwatch instead. Sample sizes stay small, so findings are directional. Studies take weeks, so research happens before a decision rather than during it. And because it is expensive, it gets commissioned for large questions and skipped for the hundred smaller ones where it would also have helped.

Outset's proposition is that an AI moderator removes the constraint. That is a substantial claim about a practice built by people who take moderation seriously, and it deserves examining on both sides: what genuinely improves when you can run hundreds of conversations, and what a machine asking the questions cannot do.

AI Moderation Is The Real Product

• An AI moderator runs the interview in video, voice or text, probing follow-up questions dynamically rather than reading a fixed script.
• Scale is the point — hundreds of conversations where a human moderator could run a handful, which changes the kind of question research can answer.
• Synthesis is automatic, with tagging, reports and highlight reels produced without someone watching forty hours of video.
• Digital Twins simulate participants grounded in real human data — a genuinely different thing from interviewing people, and the part to treat most carefully.
• Compliance is real: SOC 2, GDPR and HIPAA stated, with screening for non-genuine participants, plus a participant pool spanning 85+ countries.

Research Terms Worth Defining

Five terms that decide whether a study's findings will hold up.

Moderated interview

A conversation guided by a moderator who can follow up in the moment, as opposed to a fixed survey. The follow-up is where the value sits.

Dynamic probing

Asking unscripted follow-up questions based on what the participant just said. The capability that separates AI moderation from an automated survey.

Digital twin

A simulated participant generated from data about real people. Useful for piloting a guide; not observation of anyone, and never to be reported as what customers said.

Say-do gap

The persistent divergence between what people report they would do and what they actually do. The reason stated preference data — simulated or not — needs behavioural validation.

Skip logic

Rules that route a participant past irrelevant questions based on earlier answers, keeping a long guide tolerable and the data clean.

Scale Changes What Research Can Ask

The most interesting consequence of AI moderation is not cheaper interviews. It is that a different class of question becomes answerable.

With eight interviews you can explore a theme and form a hypothesis. You cannot compare segments, because you do not have enough people in any one of them. With three hundred you can — and comparison is where most product decisions actually live.

Speed compounds this. Research that returns in days rather than weeks can inform a decision while it is still being made, instead of validating one already taken. Every researcher knows the experience of delivering findings a fortnight after the team shipped.

Multilingual capability matters for the same structural reason. International qualitative research has historically meant either recruiting moderators in each market or accepting that everyone is interviewed in English, which quietly filters your sample to the more educated and more Anglophone. An AI moderator conducting interviews in the participant's own language removes a real distortion.

The honest counterpoint is that scale does not convert qualitative research into quantitative research. Three hundred AI-moderated conversations is still not a representative sample unless it was recruited to be one, and a large number of interviews can produce misplaced confidence in exactly the way a small number does not.

The constraint on qualitative research was never the questions. It was that one skilled person could only ask them eight times a day.

What An AI Moderator Cannot Do

Skilled moderation is not the delivery of questions, which is why a transcription tool such as Otter.ai captures a session without replacing the moderator. It is noticing what is not being said.

A good moderator registers a hesitation before an answer, a shift in posture, a topic the participant steers away from twice, an answer that contradicts something said fifteen minutes earlier. They then decide, in the moment, whether to pursue it — and the decision draws on judgement about what this particular person is comfortable with.

Dynamic probing on stated content is a genuine capability and Outset has it. Probing on the unsaid is a different order of problem, and it is where the most valuable qualitative findings have always come from.

There is also a candour question that cuts both ways. People sometimes disclose more to a machine, particularly on embarrassing or sensitive topics, because there is no social judgement in the room. They also invest less effort, give shorter answers and disengage more readily, because there is no person whose time they feel obliged to respect.

The practical consequence is not that AI moderation is inferior; it is that it is different, and it suits some studies far better than others. Concept testing and structured comparison work well. An exploratory study into something people find difficult to talk about is where a human moderator still earns their fee.

Digital Twins Deserve A Careful Look

Alongside real interviews, Outset offers Digital Twins — simulated participants described as grounded in real human data.

This deserves the most scrutiny of anything in the product, and the reason is not that simulation is illegitimate. It is that simulated responses arrive in the same interface, in the same format, and read exactly like interview data from actual people.

Used deliberately, there are reasonable applications — one being to check what published work already established, which a literature tool such as Elicit does better still. Piloting a discussion guide before spending real recruitment budget. Stress-testing whether a question is ambiguous. Getting a rough directional read on a decision too small to justify a study at all.

Used carelessly, the failure mode is severe. A model trained on how people describe their preferences reproduces stated preferences, and the gap between what people say and what they do is the single most important fact in consumer research. A simulation cannot surprise you with the thing nobody thought to ask, and being surprised is what research is for.

The rule worth adopting is a firm one: never present simulated output as what customers said. If a finding reaches a decision-maker, it must be labelled as simulation, and any decision that matters should rest on data from actual people.

The site's own framing — grounded in real humans — is doing careful work. Grounded in is not the same as collected from, and buyers should hold that distinction more firmly than the marketing does.

• Use simulation to pilot a guide, never to produce a finding you will act on.
• Label simulated data unmistakably in every artefact that leaves the research team.
• Ask what data the twins are grounded in, how recent it is, and how it was collected.
• Validate any simulated finding against a small real sample before it influences a decision.
• Never use simulation for anything involving a protected characteristic or a regulated claim.

Method Coverage And Where It Actually Fits

Outset supports concept testing, usability and UX evaluation, creative testing, brand research, market strategy, persona development and diary studies, across video, voice and text with intelligent skip logic.

That is a broad list and the methods are not equally suited to AI moderation. It is worth being specific rather than accepting the list wholesale.

Concept and creative testing fit best. The participant reacts to a stimulus, the questions are largely known in advance, and the value comes from volume and comparison — exactly what this removes the constraint on.

Usability testing works with an important caveat: what people say about an interface and what they do with it diverge sharply, and observation of behaviour remains the higher-value data. AI moderation captures the account rather than the action.

Diary studies are an interesting fit, because the burden there is longitudinal contact rather than depth in any single session — an area where automation genuinely helps and where human moderators are underused anyway.

The weakest fit is deep exploratory work: understanding a difficult experience, researching a sensitive category, or any study whose value depends on the moderator noticing something nobody anticipated.

Designing A Study That Tests The Tool

The way to evaluate a research platform is to run a study whose answer you already know, so you can grade the output rather than admire it.

The bar is step four. If the AI-moderated study misses a finding your existing method surfaced, that gap tells you exactly which studies you can trust it with.

1. Choose a study you have already run with human moderation in the last year, where you know what the findings were.
2. Rebuild the discussion guide in Outset and run it with a comparable sample — same segment, same size where budget allows.
3. Read the raw transcripts before reading the automatic synthesis. Judge whether the probing followed up the things you would have followed up.
4. Compare the findings against your known result. Note specifically what the AI-moderated study missed, because that gap defines its safe scope for you.
5. Separately, run the same guide against Digital Twins and compare all three. The difference between simulated and real is the most useful number you will get from this exercise.
6. Only then decide which of your study types move to this method and which stay with a human moderator.

Costing It Against Your Current Research Spend

Access is subscription-based, with no published amounts and no self-serve entry visible — the route in is a conversation with the vendor.

In research platforms this normally means the figure scales with study volume, seats, and whether participants are recruited from the vendor's pool or supplied by you. Bringing your own participants via shareable links is supported, and that changes the cost structure substantially.

The comparison to make is not against other software. It is against what you currently pay for the same research — moderator time, recruitment, incentives and analysis. That comparison is where AI moderation makes its case, and it is a calculation only you can run.

Two things to establish before committing: whether participant incentives are included or billed separately, since they are frequently the largest line in a research budget, and whether Digital Twin usage is metered differently from real interviews. The second matters because the cheaper option is the one you must use most carefully.

We do not publish amounts, as terms change and are negotiated per customer.

Outset: The Platform Details