AI in Market Research: What It Can and Cannot Replace in Singapore's Qualitative Research
Some months ago (on a slow Friday afternoon) I ran a transcript from one of our groups through an AI summary tool, mostly out of curiosity. It was a pricing discussion about a household product, conducted in English with generous helpings of Mandarin and Singlish, and the tool came back in about a minute with a tidy paragraph. Participants, it said, "responded positively to the lower price point". I had been in that room. When one woman in her sixties said "wah, so cheap ah", the table went quiet, two people exchanged a look, and someone else asked what was wrong with it. Cheap, in that moment, meant suspicious. The summary had read every word correctly and understood none of the meaning. I remember thinking that the tool was very good, and that this was exactly the problem.
Here's the tension. AI is now genuinely useful in research, and it has a place in our own workflow (for transcription, for first-pass coding, for checking whether a discussion guide repeats itself). Pretending otherwise would be dishonest, and a little precious. But the sales pitch has run well ahead of the evidence, and some buyers in Singapore are now being told that synthetic respondents or AI moderators can replace the room altogether. This post is my attempt to sort the claims into what AI can take over, what it can assist, and what it cannot touch, at least not with the tools and evidence we have in 2026. I'll try to be specific about both halves, because the useful answer lives in the detail.
How fast Singapore is adopting AI at work
The adoption numbers are real, though they depend heavily on how you ask (which, for a researcher, is the most interesting part). IMDA's Singapore Digital Economy Report 2025 found that AI adoption among SMEs tripled from 4.2 per cent to 14.5 per cent in 2024, while adoption among larger firms rose from 44 per cent to 62.5 per cent, and that 73.8 per cent of workers reported using AI tools at work. Then, in a September 2026 speech at the SME Centre Conference, Senior Minister of State Low Yen Ling cited an SCCCI survey in which 77 per cent of SMEs said they had adopted AI, with 45 per cent starting on off-the-shelf tools.
I read those two figures several times. Fourteen and a half per cent and 77 per cent cannot both describe the same behaviour. They almost certainly measure different things (organised deployment in one case, somebody in the company using a chatbot in the other), and a year apart. That gap is a small lesson in itself, and it is the lesson this whole post keeps returning to. The definition in the question shapes the answer, and a machine that summarises answers without interrogating the question will pass the confusion straight through. Policy is pushing in one direction regardless. Enterprise Singapore's Budget 2026 measures include a new Champions of AI programme and an expanded Productivity Solutions Grant covering more AI-enabled solutions. So research buyers will meet AI in their suppliers' workflows whether they ask for it or not. The useful question is where.
What AI already does well in qualitative research
Start with the unglamorous work, because that is where the gains are clearest. Transcription used to take days and now takes hours (for a two-hour group, at least), and for clean English audio the quality is good. Translation of stimulus and first drafts of discussion guides come out faster (a human still has to fix the idioms). First-pass coding, where a tool tags every mention of price or trust across a stack of transcripts, gives an analyst a map before she starts reading. And desk research, the background reading before a brief, has become quicker to assemble, provided somebody checks every source it produces, since these tools can state invented facts with complete confidence.
None of this replaces judgement. It replaces clerical time, and that is worth having. In our experience the hours saved at the clerical end get spent where they should have been spent all along, rereading the moments in a session that didn't fit. Actually, let me be more precise. AI doesn't save research time overall so much as move it, from typing to thinking, and a supplier who tells you it has halved the cost of insight has probably halved the thinking too. Call this range of tasks the Replaceability Spectrum, since the question is less whether AI belongs in research and more how far along the process it can go before the output starts to degrade.
The Replaceability Spectrum
Transcription sits near the left end, with a caveat I will come to. Coding sits a little further along, useful as a draft and risky as a final answer. Moderation sits past the middle, and interpretation sits at the far right. The further right a task sits, the more it depends on knowing who is in the room, what they are not saying, and why.
Synthetic respondents and the problem of the average
The boldest claim in the market is that you can skip participants altogether and ask a large language model to play them. It is an appealing idea, cheap and instant (and I understand the pull when budgets are tight), and the early academic evidence suggests caution. A 2024 study in Political Analysis by Bisbee and colleagues generated synthetic survey answers with ChatGPT and compared them with real American survey data. The synthetic averages looked reasonable. But the responses showed less variation than real people, relationships between variables often differed from the real data, small changes in prompt wording shifted the results, and the same prompt produced noticeably different answers when rerun a few months later.
A newer preprint benchmarking LLM-simulated survey responses (still under peer review, so treat it with some care) went further. Across several models, the simulated answers did not beat a simple demographic lookup, the models behaved as if demographics predicted attitudes far more strongly than they actually do, and in segment-targeting tasks they pointed teams to the wrong segment in 50 to 72 per cent of the American cases tested. That second finding is the one that worries me most. A synthetic panel does not only miss detail. It invents tidy differences between groups, then hands them to a marketing team as a targeting decision.
The research industry's own bodies have landed in roughly the same place. The MRS Delphi report on synthetic respondents sees a role in hypothesis generation and early exploration, and warns that language models can struggle to represent marginalised groups, falling back on stereotypical answers that do not reflect real diversity. I think that framing is about right. A synthetic panel can help you write better questions. It cannot tell you the answers, because it has only ever read what people wrote down, and the most commercially useful thing about people is how often they do something other than what they wrote down.
Why Singapore is a harder test for AI
Now bring all of that to Singapore, where three things make the problem sharper (at least three that I can see). The first is language. Our sessions move between English, Mandarin, Malay, Hokkien, and Singlish within a single answer, and the switch itself carries meaning (people often drop into dialect exactly when they get honest). Researchers at A*STAR, building MERaLiON, a speech and language model for Singapore, note that Singlish "exhibits many unique words and usage patterns that deviate from standard English", and that most existing audio models are built for high-resource languages and struggle with regional adaptation. That is the caveat on transcription. For a clean English interview, the machine is excellent. For a kopitiam-style group where three languages share one sentence, it needs a bilingual human checking the output line by line.
The second is indirectness. Singaporeans are polite in groups, and politeness is data. A participant who says "can try" about a concept is often closing the door gently, and the room knows it from the half-second pause before the answer. Transcripts don't record pauses well, and summaries record them not at all. We wrote about this at length in projective techniques that work in Singaporean focus groups, which exist precisely because direct questions produce socially safe answers here.
The third is representation. Singapore is a small market with meaningful differences between ethnic, religious, and generational groups, and those are the groups a synthetic model is most likely to flatten into stereotype, if the MRS panel's warning holds. Getting multicultural audience research in Singapore right is already difficult with real, carefully recruited participants. A model trained mostly on Western internet text is starting from a long way behind.
So what is actually lost when the transcript is all you have?
The three layers of a qualitative answer
My first attempt to answer that was a simple split between what people say and what they mean. It was too blunt. Or rather, it skipped a middle layer that AI can partly reach, and that middle layer is where the most interesting tooling work is happening. I now think of every answer in a session as carrying three layers of meaning, and I have been calling them the Three Layers of Meaning, which is not a clever name but it is at least an honest one.
The Three Layers of Meaning
Words
What was literally said. Transcription and coding tools capture this layer well, especially in clean English audio.
AI handles this layerSignals
Tone, pauses, laughter, a switch into Hokkien, the glance across the table. Partly detectable, rarely interpretable.
AI can flag, a human readsContext
Who is in the room, what face is at stake, what went unsaid, and why "so cheap" meant something was wrong.
Only the moderator holds thisThe words layer is where AI earns its keep. The signals layer is contested (some tools can now flag laughter, long pauses, or a change in speaking pace, which is genuinely helpful when you are reviewing twenty hours of audio). The context layer is where interpretation happens, and it depends on things that were never recorded. The moderator knew that the woman who said "so cheap" had spent the previous ten minutes describing a counterfeit purchase. The machine knew only the sentence. Our guide on analysing focus group data without losing the insight is really a long argument for protecting that third layer.
A decision table for research buyers
If you commission research, the practical question is which parts of a project you should expect AI to do, and which parts you should insist a person does. Here is how I would split it today, task by task.
| Research task | What AI does well | Where it falls short | Who should own it |
|---|---|---|---|
| Desk research and brief preparation | Gathers background quickly | States unsourced facts confidently | Human checks every source |
| Discussion guide drafting | Spots gaps and repetition | Misses local phrasing and sensitivities | Moderator writes, AI reviews |
| Transcription | Fast and accurate for clean English | Code-switching, dialect, crosstalk | AI drafts, bilingual human corrects |
| Moderation | Consistent scripted follow-ups | Cannot read the room or push gently | Human moderator |
| Coding and synthesis | A first map of themes | Treats frequency as importance | Analyst, with AI as a first pass |
| Synthetic respondents | Hypotheses to test later | Compressed variance, invented segment gaps | Never a substitute for fieldwork |
The coding row deserves one more sentence. A tool that counts how often a theme appears will tell you the most common complaint, and the most common complaint is rarely the one that explains behaviour. In one group (a household category, some years back), the remark that reframed an entire project was made once, quietly, by the participant who had spoken least. Frequency-based synthesis would have buried it on page nine.
A check we run on every AI-assisted analysis: take the five boldest lines in the AI summary and find the exact clip in the recording for each one. Then ask whether a person who was in the room would have written that line the same way. Where the answer is no, the gap usually sits in the signals or context layer, and that is where the report needs a human rewrite.
What AI moderators can and cannot do
AI-moderated interviews are the newest offer on the market (some pitch decks call them scalable qual), usually a chat or voice agent that asks scripted questions and follows up on keywords. For some jobs they are fine. A large number of short, low-stakes conversations about a website flow, where you mostly want reactions in participants' own words, can be run this way. I suspect they will get better quickly.
But moderation, at its core, is deciding what not to ask. A good moderator notices that a participant went quiet when her husband's spending came up, and decides to come back to it later, one to one, in a different way. She notices that a younger participant is echoing an older one and gives him a private way to disagree (a written card, a later question). In in-depth interviews on sensitive topics such as debt, illness, or family care, the whole value lies in a person deciding, moment by moment, how far to go. An agent following a script cannot make that call, and I'm not sure we would want it to. The same concern runs through our comparison of online and in-person qualitative research in Singapore, where the medium itself changes what people are willing to say.
Governance questions to ask any AI-enabled supplier
If your research supplier uses AI (and most now do, whether they say so or not), you are entitled to ask how. ESOMAR's 20 Questions to Help Buyers of AI-Based Services is a good starting list, with sections asking how the supplier provides human oversight of its AI system and what its data governance protocols are. Two of those questions matter most in Singapore, I think. Where do the recordings and transcripts go, and who checks the output before it reaches you?
On the first, the PDPC's advisory guidelines on personal data in AI systems, issued in March 2024, were written for AI recommendation and decision systems. The good-practice principles still map neatly onto research data. They encourage data minimisation, pseudonymising or de-identifying personal data as a basic control, and anonymising datasets as far as possible. Focus group recordings contain faces, voices, and very personal stories, so a supplier uploading raw audio to a public AI tool should be able to explain exactly why that is acceptable. On the second question, the Model AI Governance Framework for Generative AI from IMDA and the AI Verify Foundation lists accountability and content provenance among its nine dimensions, and flags hallucination as a known risk that calls for human verification. For research, provenance is simple to state. Every claim in a report should trace back to a real participant, in a real session, saying a real thing.
Put together, the governance questions and the evidence point to a workflow rather than a verdict. Here is roughly how we run AI-assisted projects now, and I expect the order will hold even as the tools change.
The Assisted Analysis Workflow
Human brief
The decision the research must inform is set by people, with AI used only for background reading.
Human fieldwork
Real participants, a moderator in the room, recordings kept in controlled storage.
AI first pass
De-identified transcripts, draft codes, flagged pauses and code-switches.
Human meaning
The moderator checks every headline against the clip and writes what it means.
What this means for research budgets in Singapore
You might be expecting me to say that AI should make research dramatically cheaper. For some parts of a project it does lower cost, mainly transcription and first-pass coding. But the expensive parts of good qualitative work were never the typing. They are recruiting the right participants, running sessions well, and interpreting them honestly, and none of those has become much cheaper. Our honest pricing guide to market research in Singapore sets out where the money actually goes. If a quote has fallen sharply because of AI, I would ask which of those three steps was removed.
There is also a risk to the brief itself. When fieldwork seems cheap and instant, it is tempting to skip the hard thinking about what decision the research should inform, and a brief that gets results still starts with that question. Some of the most interesting recent projects I have seen, from how students and parents use AI in Singapore's schools to why digital banking customers hold back their trust, were about AI itself. They needed real people precisely because the technology was changing how those people behaved, and no model trained on last year's text could have known that. This is also why our focus groups and wider market research practice in Singapore still put a moderator in the room.
The machine hears the words, the room holds the meaning
AI will take over a lot of the clerical work in qualitative research, and it should. It will get better at flagging signals. It may well become a decent interviewer for simple, low-stakes questions. What it cannot do, on the evidence available in 2026, is stand in for the people whose behaviour you are trying to understand, or for the person who sat with them and knew what "so cheap" meant that afternoon. I could be wrong about how long that holds. The tools are improving faster than anyone predicted, and I would rather revise this post in two years than pretend certainty I don't have. But for now, in a market where one answer can carry three languages and a polite refusal, the most valuable thing in a research project is still the room, and the person paying close attention in it.
What research buyers ask about AI in qualitative research
Can AI replace focus groups and in-depth interviews?
Are synthetic respondents reliable for market research?
What parts of qualitative research can AI do well?
How should research buyers in Singapore evaluate AI-enabled suppliers?
Does AI make market research cheaper in Singapore?
Deciding where AI belongs in your next research project, and where real participants still matter
AI can speed up transcription and first-pass analysis, but it cannot stand in for the people whose behaviour you need to understand or the moderator who reads what they leave unsaid. We design Singapore research projects that use AI where it is reliable, keep humans in the room where it counts, and trace every finding back to a real participant.
Request a quote →
