Voice of the customer research, from surveys to AI interviews

John Provine · Co-Founder & CEO
November 10, 2026 · 9 mins

Voice of the customer research turns what customers say, in their own words, into decisions. Here are the methods, a program to run, examples, and what AI adds.
The phrase is older than the software that now carries its name. In 1993, Abbie Griffin and John Hauser described the job as three tasks: identify customer needs, structure them, and set priorities among them. Their raw material was the customer interview, and their paper asks how many customers a team needs to interview and whether focus groups beat one-on-one conversations.
Interviews gave way to surveys for a practical reason. As we wrote in 2024, it isn't practical to interview 1,000 people by hand, so they get radio buttons and checkboxes instead. A survey can count how many customers hold a view you already knew to ask about. It can't tell you what you didn't think to ask.
An AI interviewer changes that arithmetic. It can hold hundreds of conversations at once, ask a follow-up when an answer is thin, and do it in the customer's language. That puts the interview back at the center of a voice of the customer program, at survey sample sizes. This guide covers the methods, how to build a program, examples from two teams, and where AI helps and where it doesn't. We build a platform for this work, so I'll say where a claim is about our own product.
What is voice of the customer?
Voice of the customer (VoC) is the practice of collecting what customers say about their needs and experiences, in their own words, and using it to decide what to build, fix, or say. The term comes from quality management. Griffin and Hauser wrote about it as one part of quality function deployment, a method for turning customer needs into product requirements.
A satisfaction score such as NPS or CSAT is one input. The score tells you how many customers are unhappy. Voice of the customer research tells you what they were trying to do when it went wrong, in words you can quote to the team that has to fix it.
Voice of the customer research methods
Six methods cover most programs. Each answers a different question, and most teams need two or three of them.
| Method | What it tells you | Where it falls short |
|---|---|---|
| Surveys (NPS, CSAT, custom) | How many customers hold a view you already knew to ask about | Fixed questions can't follow a surprising answer, and typed answers run short |
| One-on-one interviews | The reasons behind a behavior, with follow-up questions | A researcher runs one session at a time, so samples stay small |
| Focus groups | How a group reacts to a concept, and the words people use for it | A few voices can steer the room, and sessions are slow to schedule |
| Support tickets, reviews, and call transcripts | What breaks, at volume, with no fieldwork | You hear only from people who complained, and you can't ask a follow-up |
| In-product feedback | Reactions at the moment of use | Short answers, and only from current users |
| AI-moderated interviews | Interview depth at survey sample sizes, in many languages | Needs a human review, and a skilled moderator is still better on sensitive topics |
The last row is the newest, so it has the most to prove. Two researchers at the London School of Economics, Friedrich Geiecke and Xavier Jaravel, ran AI-led interviews and had the transcripts graded by a team of sociology PhD students from Harvard and LSE. The graders rated the AI-led interviews "approximately comparable to an average human expert (under the same conditions)," and respondents wrote 142% more words than they did in open text fields. The paper is now in the Review of Economic Studies.
That is one study, and "comparable to an average expert" leaves room for your best moderator to do better. It is enough to treat the AI interview as a research method, and to check its output the way you'd check a junior researcher's.
How to build a voice of the customer program
A program is research on a schedule, with a named person who owns what happens to the findings. These are the seven steps I'd keep.
- Start from a decision. Write down the decision the research will inform and who makes it. "What do customers think of onboarding?" has no owner. "Should we cut the second onboarding screen?" does.
- Pick the method by the question. Counting calls for a survey. Reasons call for an interview. One study can do both: a rating question, then an interviewer who asks why.
- Recruit past your happiest customers. A feedback email reaches the people who stayed. Add customers who left and people who never bought, which usually means a research panel.
- Ask what happened before you ask what they want. "Tell me about the last time you returned an order" gets a story. "Would you use a faster return flow?" gets a polite yes.
- Keep the quote attached to the theme. Count how many people raised each theme, and keep the sentence or the clip that shows it. A count without a quote gets argued with. A quote without a count gets dismissed as one person.
- Send it to the owner, and write down what changed. A finding that reaches the person from step 1 within the week still has a decision to change.
- Run it again. One study is a snapshot. Needs move when the product, the price, or the market moves.
Voice of the customer examples
Two teams that have published their work with us show what the output looks like. One is a mature insights team. The other is a startup that had no users yet.
Ancestry, a question from leadership in the field within hours. Ancestry has more than 3 million subscribers and a Consumer Insights and UX Research team. Leadership wanted to know how inflation was changing subscriber behavior. "We launched a Voicepanel study within hours to understand how financial pressures were influencing subscriber decisions," said Kendra Shapiro, Head of Consumer Insights & UXR. The study found that many subscribers were considering canceling, and the findings went to leadership to guide retention strategy. Read the case study.
Hint, a list of user truths before there was an app. Hint, the AI for the home, started research with no product and no users. It recruited homeowners from our consumer panels, put rough ideas in front of them, and rolled the findings up monthly into a top 10 list of user truths that every new hire and every project starts from. One entry reads "homeowners want a coach, not a search engine." Homeowners rejected automatic appointment booking, so scheduling shipped as opt-in. The app launched less than five months after the first study. Read the case study.
In both, the output was a decision someone made with a customer's words attached.
Voice of the customer analysis
A hundred interview transcripts take longer to read than a hundred survey rows. Until recently that meant a researcher coding each answer by hand.
AI does the first pass now. It groups answers into themes, counts how many people raised each one, and pulls the quotes that show it. On Voicepanel each theme links to the quotes, audio, and video behind it, and you can cross-tabulate by segment or ask questions of the data in chat.
Check the first pass before you share it. A summary favors the common answer, and the finding that changes a roadmap is sometimes the one three people said. Read the quotes under each theme, and look for the answers that fit no theme at all.
What AI changes in voice of the customer research
AI changes three things about the fieldwork. Samples get bigger, because an AI interviewer talks with hundreds of people at the same time. Markets open up, because the interviewer and the analysis work across languages, and ours covers 37. And the cycle shortens. Kendra Shapiro of Ancestry said her team was "able to hear from hundreds of consumers in minutes."
It leaves the hard parts where they were. Someone still has to pick a question worth asking and own the decision at the end. A human moderator is still the better choice for a sensitive subject or a visit to a customer's home. And an AI that interviews customers is a different product from an AI that plays one. Synthetic respondents answer in place of your customers, and I've written about where they hold up and where they don't.
For what AI does across a whole study, see AI customer research, explained for research teams.
Voice of the customer tools
Tools sort by the method they run. These descriptions come from each vendor's own site, the same sources behind our comparison of 20 research platforms, and one of the tools is ours.
- Survey platforms. Qualtrics has the deepest survey toolset in our comparison, and its AI adds text follow-ups to a survey. SurveyMonkey is the easiest to start with, on a free plan with public prices. Neither runs voice or video interviews.
- In-product surveys. Sprig puts short surveys inside a web or mobile product, which catches a reaction at the moment of use.
- Research repositories. Marvin pulls calls, support tickets, and surveys into one place and answers questions with citations. It analyzes feedback you already have and doesn't recruit participants.
- AI interview platforms. Voicepanel runs AI interviews in voice, video, and chat, with survey questions and usability tasks in the same study, and recruits from 10+ panel partners, or from your own customers on Enterprise. Pay per project starts at $5,000, and the prices are public.
Voice of the customer FAQ
What is the difference between voice of the customer and customer satisfaction?
A satisfaction score such as CSAT or NPS measures how customers rate an experience. Voice of the customer research collects what they say about it in their own words, so you learn the reason behind the score. A program can use both: the score to track a trend, the interviews to explain it.
What is a voice of the customer program?
Customer research on a schedule, with a named owner for the findings. It sets which decisions the research informs, which customers you hear from, which methods you use, and who acts on the results.
What questions should you ask in a voice of the customer interview?
Start with what happened: "Tell me about the last time you..." Then ask why, and what the customer did next. Ask about past behavior before opinions, and leave questions about your own ideas for the end.
How often should you run voice of the customer research?
Often enough to match how fast you ship. A team that releases every two weeks learns little from an annual study. Hint's team rolls its findings up every month.
Can AI run voice of the customer interviews?
Yes, with real customers answering. An AI interviewer asks your questions in voice, video, or chat and follows up on what each person says. In a London School of Economics study, sociology PhD students rated AI-led interviews approximately comparable to an average human expert. A person still writes the questions and checks the analysis.
If you want to see what a voice of the customer study looks like on our side, the study reference lists every block you can combine. Talk to us to run one.