How AI is changing primary and secondary research — stage by stage, the good and the bad

How widely AI is already used in research

Adoption is no longer experimental. Qualtrics’ 2026 market research trends report, based on a survey of about 3,000 researchers in 14 countries, found that 53% use AI regularly and around 90% have tried it. Among teams using synthetic data, 45% called it their most reliable data source — a finding that deserves caution given the independent evidence below.

The useful question is not whether to use AI, but which stages of a study it improves, which it weakens, and what controls keep findings trustworthy.

Secondary research: scoping and source discovery

AI can map a market in minutes — segments, players, regulation and likely data sources — giving a fast first view before deeper work.

• Better: much faster orientation in unfamiliar markets and quicker identification of where evidence probably exists.
• Worse: AI can invent reports, sources and numbers, and it often misses paywalled, offline or regional-language sources such as industry body reports, government notifications and local press.
• Guardrail: treat AI output as a list of leads to check, not as findings.

Secondary research: data gathering and synthesis

AI-assisted search and extraction can pull company filings, websites, job postings, reviews, pricing pages and news into structured tables, and summarise long reports quickly.

• Better: volume that a team could not read carefully by hand, and side-by-side comparison of documents.
• Worse: figures quoted out of context (wrong year, currency or geography), summaries that smooth over caveats and disagreements between sources, and polished errors that are hard to spot.
• Worse: some scraping can breach website terms, so collection methods need checking.
• Guardrail: every extracted figure carries its source link, date and definition, and a reviewer opens the source.

Secondary research: market sizing and modelling

AI speeds up building top-down and bottom-up models, running scenarios and checking arithmetic.

• Better: faster iterations and easier sensitivity testing.
• Worse: a tidy market size can be built on invented or mismatched inputs — polished output does not mean sound evidence.
• Guardrail: make every assumption explicit and triangulate top-down against bottom-up estimates and expert or buyer input.

Primary quantitative research: questionnaires, sampling and fieldwork

AI helps draft questionnaires, flag leading or double-barrelled questions and translate surveys quickly. Panels use AI to score fraud and target niche B2B audiences.

The biggest new threat sits in data collection. In a study published in PNAS, Dartmouth researcher Sean Westwood built an AI agent that answered online surveys like a careful human and evaded standard bot-detection checks 99.8% of the time. His warning was that researchers can no longer assume survey responses come from real people.

• Better: faster questionnaire drafting and translation; automated detection of speeders, straight-liners and copied open-ends.
• Worse: traditional attention checks no longer reliably catch AI respondents, so data can look clean and still be fake.
• Worse: AI translations can lose nuance, particularly for regional languages and mixed-language responses.
• Guardrail: verified B2B samples, identity checks for key respondents and human review of open-ended answers.

Primary quantitative research: coding and analysis

Coding thousands of open-ended answers used to take days. AI can produce a first-pass codeframe in minutes, in many languages, and explain crosstabs and statistical tests in plain language.

• Better: faster turnaround and more open-ended questions become practical to analyse.
• Worse: themes can be over-simplified, and sarcasm, cultural context and rare but important answers can be lost.
• Worse: easy analysis encourages pattern-fishing, which produces chance “significant” results.
• Guardrail: a researcher reviews the codeframe and plans the key analyses before looking at the data.

Synthetic respondents and digital twins

Synthetic respondents are AI models that answer questions as if they were a type of person. They are useful for testing whether a questionnaire flows, generating hypotheses and rough early reads before fieldwork.

As a replacement for real people, the evidence is weak. Pew Research Center tested AI “digital twins” built from its own panel and found they misestimated public opinion, with especially large errors for some subgroups. A study that found synthetic surveys broadly comparable to human samples on average answers was co-authored by the co-founder of the synthetic-data company that funded it, and positioned synthetic data as a complement for exploratory work.

• Better: cheap, instant pre-testing and hypothesis generation.
• Worse: answers can match averages while missing the spread of real opinion and the views of smaller groups — often the groups a business decision depends on.
• Guardrail: never present synthetic outputs as findings, and say clearly when they have been used.

Primary qualitative research: interviews and focus groups

AI drafts discussion guides, screens candidates, transcribes and translates interviews, finds themes across many transcripts and can now moderate interviews itself.

• Better: AI-moderated interviews can run many conversations at once across time zones and languages, and some participants may share more openly with an AI.
• Worse: an AI moderator can follow a thread but struggles to read a room — rapport, sensitive topics, senior B2B executives and what is left unsaid. Most claims about AI moderation quality come from the vendors themselves.
• Worse: transcription errors with accents, jargon and mixed languages can change a finding; AI synthesis favours what is said most over what matters most, and can paraphrase or invent quotes.
• Worse: fake experts and professional respondents can use AI to pass screeners and sound informed.
• Guardrail: humans moderate senior and sensitive interviews, every quote is traced to the transcript, and expert experience is verified.

Reporting and advice

AI produces faster drafts, charts and audience-specific summaries. The risk is that a report looks equally professional whether the evidence behind it is strong or weak. Clients can also run their own AI desk research, which raises the bar for what a research partner must add: verified evidence, judgement and accountability for the recommendation.

Business implications

• Use AI for speed in desk research, transcription, coding, analysis and drafting — and expect faster turnaround, not lower standards.
• Ask research partners how respondents are verified and how bots and fake experts are screened out.
• Treat synthetic respondents as a pre-test tool, never as evidence for a decision.
• Require every figure to have a source and every quote to have a transcript reference.
• Keep human moderators for senior, sensitive and high-stakes interviews.