Can AI Replace People in Market Research? What AI Can and Cannot Do in 2026
Short answer: AI can replace part of market research, especially secondary research, transcribing and categorizing interviews, and analyzing large volumes of text, which genuinely cuts time by tens of percent. But in 2026 AI still cannot replace four things: observing what customers do but do not say, asking follow-up questions when an informant says something unexpected, taking responsibility for the conclusions that decisions are based on, and understanding Thai context that is not in the training data.
The organizations that benefit most are those that use AI to speed up the mechanical parts of the work and then put the time saved into the work that needs real people, not those that cut the human work and hope AI will do all of it.
What AI does well in market research in 2026
Searching and synthesizing secondary data. Work that used to take researchers several days, gathering industry reports, competitor information and government statistics, can now be done in a few hours. The caveat is that you must check the source of every number, because AI can still produce figures that look credible but do not exist.
Transcribing and categorizing interviews. Thai-language transcription that used to take 4 to 6 hours per hour of interview now takes a few minutes, and AI does a good job of initial theme grouping (thematic coding). This is where qualitative research saves the most time.
Analyzing large volumes of text. For example, tens of thousands of customer reviews, social media posts, or open-ended survey answers. Work that used to rely on reading a sample can now cover everything and pick up patterns.
Checking data quality. Detecting survey respondents who answer randomly, answer abnormally fast, or contradict themselves is more accurate and faster than checking by hand.
Helping design research instruments. Drafting initial interview guides or questionnaires, spotting leading questions, and suggesting angles the team might overlook. AI works well as an assistant, but someone who understands the business question still needs to make the final call.
What AI still cannot do
1. Observe what customers do but do not say
A great deal of high-value research comes from seeing the gap between what people say and what they do. Customers say they choose on quality, but in a real store they always pick the product at eye level first. Customers say the sign-up process is easy, but during testing they pause for three seconds on the same screen every time. This kind of data is not in any text AI can read, because it was never said out loud.
2. Ask follow-up questions when something unexpected comes up
The value of an in-depth interview lies in catching small signals and digging further. When an informant hesitates before answering, contradicts what they said five minutes earlier, or tells a story unrelated to the question that turns out to matter more, a good interviewer drops the script and follows that thread. AI that interviews from a question guide still does this poorly. On sensitive topics, some groups of informants may even answer a machine more readily than a person, but AI still cannot tell which answers are worth probing further and which are simply answers that sound good.
3. Take responsibility for conclusions
When research leads to an investment decision worth tens of millions, someone has to be able to confirm what the conclusion is based on, why it can be trusted, and where its weaknesses lie. AI can give answers but does not take responsibility for the outcome. That responsibility is part of what clients are paying for.
4. Understand Thai context that is not in the training data
Much Thai consumer behavior is not recorded as English-language text on the internet, from the dynamics of decision-making in Thai families to the differences between markets in Bangkok and the provinces. When AI answers questions like these, it often responds with stereotypes that sound good but are wrong in the details that matter for decisions.
A caution about synthetic respondents
In 2026 there are tools that offer to use AI to simulate survey respondents instead of collecting real data (synthetic respondents). The selling point is that they are much faster and cheaper.
Where they genuinely work is early-stage testing, such as checking whether survey questions are understood the same way by everyone, or screening a large number of ideas at an early stage before real testing.
Where they still cannot substitute is any work where the numbers will be used for investment decisions or reported to the board. AI generates answers from patterns in data it has seen before, which means it is good at predicting the average of the past but misses what is changing, and what is changing is the one thing that gives research its value.
The rule we use: if a wrong answer would cost money, collect data from real people.
How to use AI in the right places in a research project
The approach we use in NXT projects:
- —Use AI in preparation, exploring secondary data, drafting hypotheses and drafting instruments, so the work starts from a higher baseline.
- —Always collect real data from real people for every conclusion that will be used in a decision.
- —Use AI to speed up processing, transcription, theme coding and text analysis, so the team has time for interpretation.
- —Always have people interpret and decide, especially in explaining why the numbers came out the way they did.
- —Check every number AI provides before it goes into a report the client will use.
The time saved therefore comes from processing work, such as transcription and theme coding, not from cutting steps that need people. The time it takes to schedule interviews and collect data in the field stays the same, so projects with a lot of qualitative work can be shortened more than projects that are mainly surveys.
Frequently asked questions
- Can I use AI to do market research myself, without hiring a research firm?
- Yes, for exploratory and low-risk work, such as understanding the overall industry before a meeting or screening ideas at an early stage. But for decisions with high financial stakes, there should be real data collection and someone accountable for the method and the conclusions. The most cost-effective approach is to use AI to do the homework first, then hire professionals only for the parts that need to be truly accurate.
- How much can AI reduce the cost of market research?
- In the cost structure of a research project, what AI can reduce is analysis and processing time, which is usually about 20 to 35 percent of the project. What it cannot reduce is the cost of reaching the sample, informant incentives, and researchers' time collecting data, which is the larger share. Expecting AI to cut research costs in half is therefore unrealistic for work that requires real data collection.
- Is customer data that I give to AI for analysis safe?
- It depends on the tool and the agreement. The points to check are whether the data is used to train models, where it is stored, and whether it complies with the PDPA. Safe practice is to always remove personally identifiable information first, use enterprise-grade tools with clear agreements, and state this in the consent form given to informants from the start.
- When can synthetic respondents replace real data collection?
- They can be used in work whose results do not feed directly into decisions, such as testing whether a questionnaire is understood before it is used for real, or screening a large number of ideas at an early stage. But they should not replace real data in work where numbers must be reported to the board or investors, or used to support investment.
- Will AI replace all market researchers in the next 3 to 5 years?
- What is likely to happen is a shift in the mix of work. Processing and synthesis will be done almost entirely by AI, while the work left to people is designing the right questions, collecting data in real contexts, and turning findings into decisions, which is already the highest-value part. Researchers who only do processing will be hit hard, while researchers who do interpretation and advisory work will become more valuable.
This article reflects the state of technology and practice as of 2026, which is changing quickly. The time and cost-share figures in this article are estimates based on the team's experience, not results measured in any particular project; actual results vary with the nature of the work. This article was prepared by NXT Consulting Group and translated from the Thai original.
More in this series
Share this article
