Can you trust AI with your customer research?

AI tools can turn 40 customer interviews into a polished summary overnight. That saves time. It is also how marketing teams end up signing off a website redesign or a public campaign on findings nobody checked. Here is where AI earns its place in your research, where it puts you at risk and how we split the work at JBi Digital.

• Give AI the admin: transcribing calls, tagging survey responses and grouping large volumes of feedback into themes.
• Keep people on the judgement calls: running interviews, noticing what customers do not say and deciding which findings matter.
• Treat every AI summary as a first draft until someone traces each finding back to a real quote or recorded behaviour.

AI summaries sound equally confident whether the research behind them is thorough or thin, so weak findings can reach sign-off without anyone noticing.

Picture the week. You have a redesign brief, a campaign about to launch and a head of marketing who wants evidence before they approve the budget. An AI tool offers to process 40 customer interviews overnight. By morning you have clean slides ready to present.

The summary reads the same whether it came from 40 in-depth conversations or 10 rushed ones. Nothing in it tells you, or your director, which. If the campaign then underperforms, "the research said so" will not protect you if the tool had already dropped the detail that mattered.

The stakes rise in the public sector and regulated industries, where much of our work sits. There, a research finding can shape decisions for years:

• Decisions can face an audit, a freedom of information (FOI) request or public challenge.
• Accessibility standards such as the Web Content Accessibility Guidelines (WCAG) expect you to show how real users were considered.
• A problem missed in research gets built into a website or service and stays there until the next redesign.

Nielsen Norman Group, which has tested a range of AI research tools, found that some give experienced researchers modest speed gains but none come close to replacing human expertise in synthesis. It also warns that the appearance of competence hides real limitations (Nielsen Norman Group via Rosenfeld Media).

AI is good at the repetitive work that eats up research time. Maze reports that adoption of AI in user research has grown by 32% since 2024, and analysing research data is the most common use (Maze). Used well, it helps in three places:

• Sorting large volumes of feedback. Give it hundreds of survey responses, support tickets or reviews and it will group recurring themes in hours. It shows you what people say most often. It cannot tell you why.
• Transcribing and tagging interviews. It removes hours of note-taking, as long as someone who was in the session checks the tagged quotes against what was actually said.
• Early checks on wording. It can flag whether a menu label, landing page heading or button text is likely to confuse people. Treat the result as something to test with real users, not as a finding.

Once your question moves from what customers do to why they do it, AI starts to smooth over the detail you need.

• It hides contradictions. People often say one thing and do another. On a public service website, for example, users may ask for more detailed guidance, while the analytics show them leaving pages with long blocks of text. A researcher spots that clash and asks about it. An AI summary tends to settle on the most common answer, so the contradiction, often the most useful finding, disappears.
• It misses what people do not say. A transcript records words. It does not record the long pause before a sensitive question or the internal politics that stop a stakeholder speaking freely. AI tools cannot pick up body language, tone or the setting a session takes place in (Rosenfeld Media).
• It sounds certain when the evidence is thin. An AI summary rarely says "only three people mentioned this" or "participant four disagreed". The tone stays confident however little sits behind it.
• Synthetic customers are not customers. Some tools offer AI-generated personas you can "interview". When Nielsen Norman Group compared these against three of its own studies with real participants, it found the answers were too shallow to be useful (Nielsen Norman Group). Synthetic users tend to confirm what you already think. Real users challenge it.

We use AI to save time on the admin, and our researchers stay responsible for every finding we hand over. JBi Digital has worked with enterprise and public sector organisations since 2008, and our research combines stakeholder workshops, user interviews, usability testing, behavioural analytics and journey mapping (JBi Digital). AI supports that work. It does not replace any part of it.

Our work for the British Council shows what that looks like in practice. For its TakeIELTS website, our discovery included 42 user interviews across nine countries. That research mapped a five-stage test-taker journey (researching, booking, preparing, managing and returning) and underpinned the plan to bring more than 120 country sites together on a single Drupal 10 platform (British Council case study). A tool can transcribe and tag 42 interviews. Deciding what test-takers in nine countries actually need from a website is still a researcher's job.

This is how we split the work:

Our rule is simple: if a finding cannot be traced back to a real conversation or recorded behaviour, it does not go in the report.

We also control where research data goes. Interview recordings and survey responses often contain personal data, so we only process them in approved tools, in line with our ISO 27001 certification and Cyber Essentials Plus accreditation. Pasting customer transcripts into a free AI tool is a common way for data to leave an organisation without anyone noticing (JBi Digital AI glossary).

1. Ask what each finding is based on, including how many people and which method.
2. Find at least one real quote or recorded behaviour behind every key finding.
3. Look for what is missing, such as minority views, unusual drop-offs and answers that did not fit.
4. Ask someone who ran or attended the sessions to review the summary.
5. Label anything AI-generated or unchecked as a hypothesis, not a finding.

Not by cutting conversations with real customers. Using AI to justify fewer interviews only gets you to the wrong answer faster.

Use it to cut the admin instead. As a rule of thumb, look at which 80% of the admin, such as tagging, transcribing and formatting, can safely go to AI. Put the time you save into the 20% that explains why customers decide what they decide. That is the part your campaign or redesign depends on.

If your next campaign, redesign or content plan rests on customer research, the evidence has to hold up when your head of marketing, director or board asks where it came from. That is hard to guarantee when the findings came out of an AI tool nobody had time to check.

Our research and strategy team can help you:

• Check AI-generated research summaries against the interviews, survey data and analytics behind them.
• Set up a research process where AI handles transcription and tagging, and researchers lead the interviews and the sign-off.
• Turn checked findings into a clear brief for your campaign, website redesign or landing pages.

New to terms like AI hallucination and human-in-the-loop? Our AI glossary for marketers explains the AI terms that come up most often in marketing, so you can question what a research tool claims to do before you rely on it.

Planning a campaign or redesign built on customer research? Talk to us before your findings go to sign-off.

Send us a message