AI in Market Research: When to Use It and When to Push Back

AI is everywhere in market research right now. New tools promise to simulate respondents, conduct interviews, and generate insights without a human in the room. Some of it is genuinely useful. Some of it should make you nervous.

Our point of view is this: AI makes good research better. It does not make researchers optional. Below, we break down how we think about AI in research holistically as well as where we stand on the AI tools and methods getting the most buzz right now.

The Way We Think About AI In Market Research

Not all AI applications are created equal. The ones we’re most excited about are the ones that make our team sharper: faster at processing information, better at spotting patterns, more equipped to bring fresh angles to a research challenge. The ones that give us pause are the ones designed to cut humans out of the process entirely.

Here’s what AI can’t do, regardless of how sophisticated it gets. It can’t read a pause. It can’t sense when someone is telling you what they think you want to hear. It can’t follow an unexpected thread. It can’t bring human judgment to a finding that doesn’t fit neatly into a pattern. And it can’t take responsibility for a recommendation that turns out to be wrong.

Research is fundamentally about understanding people. People are messy, contradictory, and surprising in ways no model has fully figured out yet. The best research happens at the intersection of good methodology and human judgment. AI can strengthen both. It just can’t replace either.

Where we’re all in: AI for tactical research operations work

There’s a category of AI tools we use regularly and recommend without hesitation. What they have in common: a human is always setting the direction, interpreting the output, and making the calls. AI is doing the heavy lifting in the background.

A few places we’ve found it genuinely valuable:

Analyzing open-end responses. When you have 500 or 1,000 open-end responses, human coding is slow, expensive, and inconsistent. Two coders rarely categorize things exactly the same way, and fatigue sets in fast. AI processes the full dataset consistently and quickly, surfacing themes a human might miss simply because they were on response 400 of 600. A researcher still interprets what those themes mean. But the starting point is better.
Survey design. Even experienced researchers have blind spots. AI is a useful thought partner for drafting question sets, flagging leading or double-barreled language, and pressure-testing flow before a survey goes live. It won’t design your study for you, but it will catch things you might have read past for the tenth time.
Report writing and narrative generation. Staring at a dataset and trying to find the story is one of the most time-consuming parts of research. AI can draft chart callouts, summarize key findings, and build a first-pass narrative that gives your team something to react to rather than starting from a blank page. The final product still requires a human who understands the client’s business. But you get there faster.

Our take: Use AI to do the time-consuming work faster and more consistently. Keep a human in the driver’s seat for everything that requires judgment, interpretation, and context. That’s when research gets good.

Where we’re in, with caveats: AI-moderated interviews

Traditional qualitative research, the kind with a human moderator leading a one-on-one interview or a focus group, is one of the most powerful tools in research. It’s also time-intensive and expensive to do at scale. AI-moderated interviews have emerged as an answer to that problem.

Instead of a human moderator, an AI conducts the conversation (example in this earlier blog post). It asks questions, responds to answers, and probes for more detail. A growing number of platforms have made this widely accessible, and it’s getting a lot of attention as a faster, lower-cost alternative to traditional qual

Here’s where it genuinely improves on the traditional open-end survey question:

It probes. A standard open-end asks a question and accepts whatever answer it gets. AI moderation follows up, asks for clarification, and pushes for more detail in ways a static survey never could.
It scales. You can hear from hundreds of people in an open-ended format without the cost or time of recruiting and scheduling human-moderated sessions.
It’s consistent. Every respondent gets the same questions in the same way, which makes responses easier to compare and analyze.

But it has real limitations that matter:

It can’t go off script. A human moderator notices when something unexpected comes up and chases it. AI follows its programming. If the most interesting insight is hiding in an unplanned direction, AI won’t find it.
It can’t read the room. Tone, hesitation, body language, the way someone’s energy shifts when a topic makes them uncomfortable. AI captures and interprets none of that.
It can’t connect threads across conversations. A human moderator who completes ten interviews in a day starts to notice patterns, tensions, and outliers in real time. AI treats each conversation as its own isolated event.

Our take: AI-moderated interviews are a meaningful upgrade to the open-end survey question. They are not a replacement for a human moderator. Use them when you need scale and depth on a defined topic. Bring in a human moderator when you need real discovery.

Where we have mixed perspectives: AI respondent screening and data quality

Before any research study begins, someone has to make sure the right people are in it. And once the data comes in, someone has to make sure it’s worth using. These are two of the most time-consuming and unglamorous parts of research: screening respondents before they enter a study, and reviewing data quality after responses come in.

Traditionally, both rely on static rules and manual review. They work, but they have limits. Screeners can’t adapt based on what someone says. Data quality checks can’t catch every bad actor at scale. And both are vulnerable to respondents who game the system to collect an incentive without doing the work. Bad respondents mean bad data, and AI is making both problems meaningfully easier to solve.

AI is changing this in a few distinct ways:

Conversational screeners. Instead of a static form, an AI conducts a short qualifying conversation, asks follow-up questions based on what someone says, and makes a real-time judgment about whether they belong in the study. It’s much harder to game than a checkbox screener.
Behavioral fraud detection. AI monitors how respondents move through a survey, flagging suspicious patterns like copy-pasting answers, unusually fast completion times, or response patterns that suggest a bot or a professional survey taker just trying to collect an incentive.
Open-end quality scoring. AI evaluates written responses in real time, identifying gibberish, AI-generated text, or answers that are clearly off-topic or disengaged.
Consistency tracking. AI can flag respondents who contradict themselves across questions or across studies, catching people who misrepresent who they are to qualify.

Our take: not all of these are equal, and the right level of human involvement depends on how much judgment the decision requires.

When the criteria are objective and the patterns are clear, AI can largely take the wheel. When the decision carries real consequences for who gets included or excluded from a study, a human needs to stay in the loop.

Behavioral fraud detection and open-end quality scoring. AI takes the wheel. The volume of decisions makes human review impractical, and the signals AI reads are concrete and measurable.
Conversational screeners. AI as augment. Humans set the qualifying criteria, AI enforces them with more nuance and adaptability than any static form could.
Consistency tracking. Humans stay involved. Decisions about who gets excluded from a study carry real consequences and deserve a human review layer.

Used well, AI screeners don’t just protect data quality. They free researchers to focus on what actually requires human judgment.

Where we’re skeptical: synthetic respondents

Synthetic respondents are AI-generated personas that simulate how real people would respond to survey questions or research prompts. The AI builds these personas by drawing on its training data, which includes everything it has ingested: surveys, social media, consumer research, behavioral data, published studies, and more.

It looks for patterns in how certain types of people think, talk, and respond, and uses those patterns to generate simulated answers. Instead of recruiting actual humans, you feed the AI demographic and psychographic parameters and it produces responses based on what it has learned about people like that. No recruiting, no fieldwork, no waiting. The appeal is obvious.

But here’s the critical thing to understand: synthetic respondents are only as good as what the AI has actually been exposed to. If the AI has seen a lot of data about a population and a topic, the simulated responses may reasonably reflect reality. If it hasn’t, the AI is essentially making educated guesses and presenting them as research findings. That distinction should drive every decision about whether synthetics are appropriate for a given project.

Here’s when synthetic respondents can be defensible:

Well-known populations on well-researched topics. General consumers, broad demographic groups, widely studied behaviors and attitudes. These are populations and subjects the AI has ingested a lot of data on. The simulated responses have a reasonable foundation.
Early-stage directional work. If you need a rough signal to pressure-test a concept or narrow down hypotheses before investing in real fieldwork, synthetics can play a supporting role.

Here’s when we’d push back hard:

Niche or specialized populations. B2B buyers, technical professionals, niche hobbyists, underrepresented communities. AI doesn’t have enough training data on how these groups actually think and behave. The responses it generates reflect gaps in its knowledge or existing bias in limited data sets, not the reality of your audience.
Any topic with real stakes. Product launches, pricing decisions, messaging strategy. If a client is making a significant business decision based on the research, synthetic respondents are not a sound foundation.
Emerging or fast-changing topics. AI training data has a cutoff. If the topic is new or rapidly evolving, the simulated responses reflect a world that may no longer exist.

Our take: the right question to ask before using synthetic respondents is not “can we do this?” It’s “does the AI actually know enough about these people and this topic to simulate them accurately?” If the answer is yes, synthetics can be a useful and efficient tool. If there’s any doubt, talk to real people.

Where we’ll always land

Our POV today is not our POV forever. AI in research is moving fast, and we’re paying attention. Tools that feel experimental right now will mature. Limitations we’re flagging today may get solved. We’re not skeptics for the sake of it.

What won’t change is the standard we hold any tool to. We’ll adopt AI when it makes us faster, sharper, and more efficient without compromising what research is actually for: accurately representing the people we’re studying.

That means we’ll always ask:

Does this tool risk flattening the voices of niche, underrepresented, or hard-to-reach populations?
Does it introduce bias by over-relying on data that skews toward majority groups or well-documented behaviors?
Does it get us to better insights, or just faster ones?

Speed and efficiency are worth pursuing. But not at the cost of the respondent. The moment a tool starts making research more about what’s convenient to measure than what’s true for the people behind the data, we’ll push back. That’s not a limitation of how we think about AI. It’s the whole point of why research matters.