On this page
Can an AI agent do market research?
Partly, under a researcher. An AI agent for market research can moderate customer interviews at scale, draft short surveys, synthesize responses and write cited desk research. A researcher still frames the question, approves the interview guide, checks sources and quotes, and decides what the findings mean. Treat it as a fieldwork and first-draft team that never sleeps, not as the person who owns the answer.
Two companies describe AI interviews in production, both through their vendor's customer stories. Listen Labs lists 150+ global, multilingual interviews and a 24-hour turnaround among Microsoft's results. A Microsoft research manager is quoted saying traditional research "takes 4 to 6 weeks to get to insights." Outset says Away, a luggage brand with a one-person UX research team, ran 75 interviews overnight, and a study that used to take a month took less than two weeks.
What the agent does and what stays with you
Agent
0of 8
Agent with review
6of 8
- Turn the business question into a research plan
- Write a short survey from a plain-language goal
- Moderate one-to-one customer interviews
- Synthesize interview responses into themes
- Desk research on market size, trends and players with citations
- Produce report drafts from a tracking dataset
Human
2of 8
- Recruit and screen the right participants
- Decide what the findings mean and recommend action
8 tasks: 0 agent, 6 agent with review, 2 human
| Task | Who does it | Why |
|---|---|---|
| Turn the business question into a research plan | Agent with review | OpenAI documents deep research proposing a research plan you review and edit before it runs. |
| Write a short survey from a plain-language goal | Agent with review | Morning Consult says its Smart Survey designs a survey of up to 10 questions from a plain-language description; whether to check the wording before fielding is up to you; we recommend it. |
| Moderate one-to-one customer interviews | Agent with review | Outset and Listen Labs customer stories describe AI moderators running 75 and 150+ interviews; we recommend a researcher approves the guide, since neither story shows the AI writing it. |
| Synthesize interview responses into themes | Agent with review | Outset says its synthesis tools let Away's researcher skip hours of footage review and go straight to the biggest takeaways, which she then interpreted. |
| Desk research on market size, trends and players with citations | Agent with review | OpenAI says deep research outputs include citations or source links so you can verify them, which means a person must. |
| Produce report drafts from a tracking dataset | Agent with review | Morning Consult says its AI Research Agents generate editable PowerPoint reports from its daily survey data. |
| Recruit and screen the right participants | Human | No source read for this page documents an agent choosing the sample on its own; recruitment criteria remain the researcher's call. |
| Decide what the findings mean and recommend action | Human | In Away's story the researcher draws the conclusion; ESOMAR asks suppliers whether final decisions are made by humans. |
The split is between gathering and judging. The sources here show agents running interviews, drafting surveys, summarizing transcripts and writing cited reports. None shows an agent choosing the sample or making the recommendation.
The two interview stories draw that line themselves. Away's researcher said she left "the interview process and the first analysis portion" to Outset, and then found the result herself: shoppers were using AI tools mid-journey, not only for discovery as she had assumed. At Microsoft, according to Listen Labs, the researcher reaches "hundreds of users at one third of the cost." That is her quote; the story does not say what the baseline was.
Desk research is the other half of the job. OpenAI documents its deep research tool as proposing a research plan you can edit, searching the web or sites you choose, and returning a report with "citations or source links so you can verify the information." The tool hands you the means to check; it does not check for you.
A data vendor offers agents on top of its own panel. Morning Consult says it runs 30,000 consumer interviews a day and that its AI Research Agents turn that data into editable PowerPoint reports on brand reputation, customer targeting and company performance. Its Smart Survey designs a survey of up to 10 questions from a plain-language goal, with results "in as little as 24 hours." These are the vendor's own descriptions.
A worked example: one interview study
This is an illustration, not a customer story. Assume, for this example, that you want to know why trial users of your app do not convert to paid.
You write the decision ("change onboarding or change pricing") and three questions. The agent drafts a 15-minute discussion guide; you cut one leading question and approve it. The interview platform invites 60 trial users from your CRM export, and the AI moderator interviews 42 of them over two days, probing when answers are vague.
The agent's synthesis says 18 of 42 participants mention price, and quotes three of them. You open the transcripts. Two quotes match. The third merges two answers from the same participant into one sentence, so you replace it with the original. You also find that 11 of the 18 mentioned price only after the moderator asked about it, which weakens the theme. You write the recommendation: test onboarding first. The agent collected and sorted; you weighed the evidence.
What you need to provide
Start with the decision, not the topic. "Should we enter the German market this year" gives the agent something to answer; "research Germany" does not.
Give it the audience and access: screening criteria, a customer list or panel account, and read access to past studies and analytics you allow it to use. The agent touches personal data, often voice and video of participants, plus your customer records. Keep access to the study in scope, and check the provider's retention and model-training terms against your privacy notice.
Name the sources it may cite. List the paid reports you license, the sites it may search, and anything off limits. OpenAI's tool lets you restrict research to specific sites, which is one way to enforce this.
Decide how participants will know they are talking to an AI. ESOMAR's checklist asks suppliers how they make clear when AI is used in any part of the service. Consent and disclosure rules differ by country and by panel; ask counsel which apply to your studies.
Job brief template
Use this as the brief for a vendor, an operator or your own team. Replace the bracketed values with your own.
Copy-ready brief
- Goal
- Each research question gets a findings report within [10] business days, with every claim traced to a participant response or a cited source.
- Deliverable
- A findings report saved to [shared folder]: a one-page summary, themes with participant counts, supporting quotes linked to transcripts, a cited desk-research section with a source list, and a list of open questions.
- Inputs
- The business decision the research must inform and the [3–5] questions behind it
- Target audience definition and screening criteria, plus access to your customer list or a panel provider
- Approved discussion guide or survey draft, or permission for the agent to draft one for your review
- Read access to prior research, CRM exports or analytics you allow it to use
- A list of sources the agent may and may not cite, such as paywalled reports you license versus unvetted websites
- Acceptance criteria
- Every quote in the report matches a transcript word for word on a sample of [10] quotes.
- Every figure or market claim links to a source a reviewer can open; on a sample of [10], none is fabricated or misattributed.
- Sources older than [24] months are flagged as dated, and paywalled sources the reviewer cannot open are listed separately.
- Each theme states how many participants raised it, and the counts match the transcripts on spot check.
- Questions the evidence does not answer are listed as open, not filled in.
- Systems
- AI interview or survey platform (for example Outset or Listen Labs)
- deep research tool with web and file access (for example ChatGPT deep research)
- shared drive or research repository
- CRM or panel provider for recruitment
- Constraints
- The agent touches personal data from participants (voice, video, text answers) and your customer lists; limit access to the study in scope.
- Participants are told when an AI conducts the interview; check what consent and disclosure your jurisdiction and panel require.
- Synthetic or simulated respondents are labelled as such and never mixed with real responses.
- Check the provider's data retention and model-training terms against your privacy notice.
- Volume
- Example: 1–4 studies a month, each with 20–150 interviews or one desk-research report
- Cadence
- Example: per project, with a monthly tracking report
- Human checkpoints
- Study design: the researcher approves the questions, audience and discussion guide before fieldwork.
- Fieldwork: the researcher reads a sample of early interviews and adjusts the guide.
- Findings: the researcher checks quotes, counts and citations against transcripts and sources.
- Recommendations: the researcher or decision owner writes them.
The quote and citation checks do most of the work. If every quote matches a transcript and every figure opens to a real source, review is quick. If they do not, you are paying for the research twice.
Risks and human checkpoints
The report cites a source that does not exist, or attributes a figure to the wrong publisher.
How you notice it
Open a sample of citations every report; OpenAI's documentation says sources are provided so you can verify the information.
Desk research relies on stale data or skips paywalled reports the agent cannot read.
How you notice it
Require a date on every source and a list of sources it could not access; ESOMAR notes AI may have limited knowledge of events after a particular date.
Synthesis misquotes participants or overstates how many raised a theme.
How you notice it
Spot-check quotes and counts against transcripts before the report is shared.
The sample is biased, or synthetic data is presented as real customer views.
How you notice it
Ask the supplier ESOMAR's questions on bias, representativeness and how synthetic data is flagged; reject reports that do not separate the two.
A confident headline finding drives a decision before anyone checks it.
How you notice it
No finding leaves the research team without the researcher's sign-off on the evidence behind it.
Human checkpoints
- Study design: the researcher approves the questions, audience and discussion guide before fieldwork.
- Fieldwork: the researcher reads a sample of early interviews and adjusts the guide.
- Findings: the researcher checks quotes, counts and citations against transcripts and sources.
- Recommendations: the researcher or decision owner writes them.
The main risk is not an obviously wrong report. It is a plausible one: a market figure from a source that does not say it, a theme counted generously, a quote tidied into something the participant never said. ESOMAR's buyer checklist names the underlying problems: generative AI "can sometimes produce biased, incorrect or otherwise misleading answers" and may have limited knowledge of events after a particular date. It asks suppliers what measures they take to identify hallucinations, and, for synthetic data, how they validate it against real-world results and flag it apart from data from real people.
ESOMAR published the checklist in March 2024. It stays here because it is the industry body's buyer checklist for AI-based research services and no newer edition was found.
How to run a trial
Run the agent on a question you have already answered, so you know what good looks like. The length is your call; our suggestion is one full study plus one desk-research report.
- Pick a past study with known findings, or run a small new one in parallel with your usual method.
- Ask your shortlisted suppliers ESOMAR's 20 questions in writing, especially those on output validation, limitations and human oversight.
- Approve the guide, then read the first five transcripts before fieldwork continues.
- Check 10 quotes against transcripts and 10 citations against their sources. Count every mismatch.
- Compare the agent's themes with your known findings. Note what it missed and what it overstated.
- Time your review. Hours spent checking are the cost you trade against fieldwork and drafting time.
Pick a past study
With known findings, or run a small new one in parallel with your usual method
Ask your shortlisted suppliers ESOMAR's 20 questions
In writing, especially on output validation, limitations and human oversight
Approve the guide, then read the first five transcripts
Before fieldwork continues
Check 10 quotes and 10 citations
Count every mismatch
Compare the agent's themes with your known findings
Note what it missed and what it overstated
Time your review
Hours spent checking are the cost you trade against fieldwork and drafting time
If the agent meets the brief's criteria, let it run fieldwork and first drafts for one study type, and keep design and recommendations with your researcher.
For the brief, see how to write a job description for an AI agent, and for structuring the test, how to run a paid trial. Related pages: the market research analyst role, competitive intelligence, deep research reports, literature review and the research overview.
Questions buyers ask
Are AI interviews as good as a human moderator? The sources here are customer stories, not comparisons. Away's researcher said the "qualitative nuance is similar to a human moderator." No independent head-to-head study was found for this page; run your own parallel test.
Can I skip real respondents and use synthetic ones? ESOMAR asks suppliers what validation compares synthetic outputs with primary research or real-world results. Without that evidence, treat synthetic answers as hypotheses to test with real people.
What does it cost? The vendors cited here ask buyers to book a demo or contact sales, so this page shows no price range.
Sources
- How Microsoft uses Listen Labs to better understand their customers
Listen LabsVendorAccessed Nov 14, 2026
- How Away's Lone Researcher Used Outset to Conduct 75 Interviews Overnight
OutsetVendorAccessed Nov 14, 2026
- Morning Consult Intelligence: Always-On Market Platform
Morning ConsultVendorAccessed Nov 14, 2026
- Deep research in ChatGPT
OpenAI Help CenterPrimaryAccessed Nov 14, 2026
- 20 Questions to Help Buyers of AI-Based Services for Market Research and Insights (March 2024)
ESOMARPrimaryAccessed Nov 14, 2026
What we could not verify
- Both production accounts (Microsoft with Listen Labs, Away with Outset) are vendor-published customer stories. No independent study comparing AI-moderated interviews with human-moderated ones on the same question was found for this page.
- Microsoft's "one third of the cost" is a quote from a Microsoft researcher in Listen Labs' story; the baseline it compares against is not stated.
- Agent pricing is not shown. Listen Labs, Outset and Morning Consult direct buyers to a demo or sales contact on the pages read, so the page has no cost comparison.
- No source was found measuring how often AI-written research summaries misquote participants or misattribute a source in market research specifically.
- Rules on recording, consent and AI disclosure for research participants vary by jurisdiction and were not researched beyond ESOMAR's buyer checklist.
Related
- AI agents for research: jobs they do and how to hire one
- Can AI replace market research analysts? Task by task
- AI agent for competitive intelligence: what it does
- AI deep research agent: what it can do for reports
- AI agent for literature review: what it does and needs
- AI agent job description: how to write one, with template
- How to test an AI agent before buying: run a paid trial