AI Customer Interviews: How They Work, When They Beat a Human Call, and How to Run One

An AI customer interview is a research call where an AI persona asks your questions, follows up on what the customer says, and returns a transcript and analysis, with no researcher on the line. You write the goals and the script, send a link, and the call runs on the customer's schedule. It works best when you need many conversations on a known topic. It works worse when you do not yet know what to ask. Below: how a run works, when to choose AI over a human, a sample script, and consent.

Here is a common way research stalls. You plan to talk to 30 customers. Every call needs a calendar slot, a note-taker and a person with a free afternoon, so you get through six. The other 24 stay a spreadsheet of names.

What an AI customer interview is

It is the same research method you already know, with the interviewer swapped. Nielsen Norman Group defines a user interview as a method where the interviewer asks participants questions about a topic, listens to their responses, and follows up with further questions. Nothing in that definition requires a human. The interviewer has to ask, listen and follow up.

An AI customer interview is that loop run by an AI persona on a live voice or video call. The customer sees a named persona. The persona opens the conversation, works through your questions, asks follow-ups, and ends the call. When it finishes you get a recording (if recording is on), a transcript, and a scored analysis.

Two terms get mixed up here, so pin them down:

• Generative research explores. You do not have a hypothesis yet. You want to learn what the customer's world looks like. Customer discovery and Jobs To Be Done interviews live here.
• Evaluative research checks. You have something specific, a feature, a pricing page, an onboarding flow, and you want to know whether it works for the person using it.

AI interviewers are strongest on the second kind and on the structured half of the first. Section three explains why.

How one runs, end to end

Here is the whole flow on Waterr. Four steps, and only the first needs real thinking from you.

1. Write the goals and the script. In Waterr a conversation starts as a scenario: the persona, a meeting script that says how to open, what to cover and how to handle situations, and a set of goals. Goals are the criteria the session is scored against, and the participant does not see them during the call.
2. Share a link, or embed it. Every scenario has a share link in the form waterr.ai/{username}/{scenario-slug}. Flip "Allow anyone with link" on and anyone you send it to can start. Or restrict it and send invite-only links. The link guide walks through the options.
3. The customer joins and consents. They see a welcome page with the persona's name and avatar, grant mic permission (camera optional), accept the consent prompt if you turned it on, and the conversation starts. Nobody scheduled anything.
4. You read the output. After the session you get the transcript and a post-meeting evaluation: a score per goal with written feedback, an overall score, and transcript highlights tied to those goals. You can export results as CSV, receive them by webhook, or pull them through the API.

That is the full loop. If you want the API-level picture, what an AI meeting API is covers the primitives, and the customer interview use case is the page to open if you would rather run one than read about it.

One practical note on goals. The docs frame them as evaluation criteria for a participant, so for research you write them as coverage checks: "Customer described the last specific time they hit this problem." "Customer named what they tried before." The score and feedback then tell you which interviews got you what you needed and which drifted.

When AI beats a human interviewer, and when it does not

This is the section the vendor pages skip. Be honest with yourself here, because picking the wrong method wastes more time than not interviewing at all.

AI is the better choice when:

• You need volume. Thirty conversations in a week is hard to schedule by hand. It is a link in 30 inboxes.
• Your customers are in different timezones. The call happens when the customer is free, not when you are.
• You need consistency. Every participant gets the same questions in the same order with the same follow-up rules. That matters when you are comparing answers across segments.
• Interviewer bias is a worry. A human who is invested in the product drifts toward the answers they hope to hear. NN/g names this problem directly: if the interviewer asks many leading questions, the validity of the data will be compromised. A script that cannot get excited is a guard against that.
• The topic is known. You are validating a pattern, testing a message, or collecting feedback after an event, not hunting for the right question.

There is also a softer effect. Customers are ~10x more expressive when answering an AI than when recording themselves cold. That comparison is against self-recorded video. A skilled human interviewer is a different baseline, so do not read it as a claim that AI beats a good researcher in a room.

A human should run it when:

• The topic is sensitive. Grief, health, money trouble, layoffs. Rapport and judgment about when to stop are not things to outsource.
• You have no hypothesis. Early discovery depends on a researcher noticing the odd sentence and abandoning the guide to chase it. A script, even one with good follow-up rules, is a script.
• The relationship is the point. A strategic account, a churn-risk executive, a design partner. The call is also a relationship moment.
• You need to read a room. If body language and the mood in the room matter to the question, be in the room.

A workable rule: run your first five to ten interviews yourself to find the questions. Hand the question set to an AI to run the next fifty. Keep humans on the accounts where the call itself matters.

NN/g's broader advice applies either way: combine methods, because each has strengths and weaknesses. Interviews tell you what people say about their problems. They are poor at predicting what people will do with a feature that does not exist yet.

Writing a script that probes

The script decides everything. A well-prompted persona with a lazy script produces polite, useless transcripts. Four rules:

Ask about the past, not the future. "Tell me about the last time you..." gets a real story. "Would you use a feature that..." gets a guess. NN/g makes the same point: users cannot reliably predict their behaviour with something they have not interacted with.

Avoid leading questions. "Don't you find reporting slow?" plants the answer. "How do you get your weekly numbers today?" does not.

Write follow-up rules, not just questions. In the meeting script, tell the persona what to do when an answer is vague: ask for one specific example, ask what happened next, ask what they did instead. Two follow-ups per question is a sensible ceiling.

Budget the time. Decide how long the call is and how that splits across questions. Six questions with two follow-ups each fits in about fifteen minutes of talking. Ten questions turns into a survey read aloud.

A sample six-question script for a discovery-style call about how people run weekly reporting:

1. Walk me through the last time you put together your weekly numbers. What did you open first?
2. What part of that took longer than you expected?
3. What did you do the last time something in the numbers looked wrong?
4. What have you tried to make this easier? What happened?
5. If you could hand off one part of this tomorrow, which part, and why that one?
6. Is there anything I should have asked and did not?

Notice what is missing: any mention of your product. Questions one to five are all about their past behaviour. Question six is the escape hatch that often produces the best material.

Consent and recording

Tell people they are talking to an AI, and tell them if the call is recorded. Do it before the call starts, not in a footer.

Waterr has an optional consent prompt shown before the session begins. Per the docs, participants see a notice that they are about to interact with an AI persona, information about recording if it is enabled, and your own message if you add one. They accept or decline. If they decline, they do not join.

Two details worth knowing. First, the prompt is a gate before the session. It is not a stored, auditable consent record, so if your compliance team needs an audit trail, capture consent in your own backend first. Second, the docs list GDPR, CCPA and BIPA as regimes where a consent prompt helps, and note that healthcare, education and financial services often require explicit consent for recorded interactions. That is not legal advice, and neither is this post. Your counsel decides what your market requires.

For how to word the ask, see the guide to getting consent. Short, plain language, and say what happens to the recording.

Reading the output

Thirty transcripts are a pile unless you have a method. This is the practical order:

1. Scan the scores first, then ignore them. Per-goal scores tell you which interviews covered the ground. Interviews where the customer never described a concrete incident are weak data. Mark them and move on.
2. Read every transcript once, fast. You are looking for repeated phrases, not conclusions. Mark the sentence, not the paragraph.
3. Code in two passes. First pass: tag each highlight with a short label ("manual copy-paste", "waiting on finance"). Second pass: merge labels into themes. This is thematic analysis, and you do it, not the tool. The scored analysis points you at the moments; the judgment is yours.
4. Count by segment. A theme that shows up in 12 of 30 interviews but only among one customer size is a finding about that segment, not about all customers.
5. Stop when it stops surprising you. That is saturation. For qualitative usability tests NN/g's well-known figure is five participants, and that is for one tightly scoped test. Interviews across several segments need more. Add interviews until new ones stop adding new themes inside each segment.

Export to CSV if you want scores in a spreadsheet next to your tags, or use the webhook to push each finished session into wherever your team already keeps research notes.

Embedding interviews in your product

The best feedback arrives in the moment. Someone just finished onboarding. Someone just cancelled. Someone just hit the third error in an hour. A link in an email a day later catches the memory after it has been smoothed over.

The embed guide covers the options. Waterr scenarios drop into a page with an iframe, in three formats: inline, a floating button, or a small corner avatar. The page must be served over HTTPS because the iframe needs camera and microphone access. A public scenario works with a plain snippet. A private scenario, or one where you want to pass the signed-in user's identity, uses a short-lived embed token that you mint server-side with POST /v1/embed-tokens, so your API key never reaches the browser.

If hiring is your case instead, the sibling write-up is the AI interviewer API. If you are training reps rather than researching customers, see the AI sales roleplay API. The primitives are the same. The goals and the script are what change.

Frequently asked questions

Can AI conduct customer interviews automatically? Yes. An AI persona can run the full call: open it, ask your questions, follow up on vague answers, and close. You write the goals and script and send a link. The customer joins when they are free, and you get a transcript and scored analysis afterwards without a researcher attending.

Are AI customer interviews as good as human ones? For structured questions on a known topic, they are consistent and scale well. For open discovery, sensitive subjects, or high-value relationships, a skilled human does better. Use a human to find the right questions, then use AI to ask them many times.

How many AI interviews do I need? Run until new interviews stop adding new themes within each segment. A tightly scoped usability test can saturate at around five participants, per NN/g. Interviews across several customer types usually need more. Treat any fixed number as a starting point, not a rule.

Do I need consent to record an AI interview? Tell participants they are talking to an AI and whether the call is recorded. Waterr offers a pre-join consent prompt that covers both and blocks anyone who declines. It is a gate, not an audit record. What the law requires depends on where your participants are, so ask counsel.

Can I run AI interviews inside my own app? Yes. Embed a scenario with an iframe as an inline panel, floating button or corner avatar. Public scenarios need only the snippet. Private scenarios, or ones that identify the signed-in user, use a short-lived server-minted embed token. Your site must be on HTTPS.

The 24 conversations that did not happen are the real cost of customer research. Take the first five yourself, write down the questions that earned their keep, and then see how many of the rest you can get back by sending a link instead of finding an hour.