How an AI interviewer works

What an AI interviewer is

An AI interviewer runs a qualitative interview with one person at a time. It works from a research brief, not from a script: the brief says what you need to understand, and the interviewer decides in the moment which question gets closer to it. That is the main difference from an online form or a chatbot survey, which ask the same questions in the same order and, at best, add a canned follow-up.

Narratia is built on this idea. Its interviewer is meant for any field where knowledge, opinions or memories live only in people’s heads: research, product work, knowledge transfer, journalism and genealogy.

What happens in a single turn

Before each reply the interviewer does three things:

1. Reads what the person said. The whole answer, including asides, hesitations and corrections.
2. Checks it against what it already knows. The brief, the permitted supporting material and everything said earlier in the conversation.
3. Picks the next step. Ask a follow-up, reflect back what it understood, ask for a real case, or move to a topic the brief still needs.

Because the next question comes from the last answer, two people talking about the same topic get different interviews. The interviewer refers back to what someone already said instead of asking again, and when it gets something wrong and is corrected, it says so and updates the picture.

The interviewing techniques it follows

Good human interviewers share a few habits, and the interviewer is designed around them. Narratia’s standard lists 12 marks of a professional interview, drawn from classic qualitative research. Among them:

• Short questions, long answers. One question per turn, so the person does most of the talking.
• A real case instead of generalities. “The last time this happened, what did you do?” gets closer to behavior than “Do you usually…?”
• Reflecting back. The interviewer says what it understood and lets the person confirm or correct it.
• No leading, judging or repeated questions. It is designed not to put words in the person’s mouth, and these behaviors are evaluated in testing.
• Reading the room. It tells a joke from sarcasm, fatigue or a test, and replies in the same tone, without a stock “I understand how you feel”.

Before, during and after the interview

Before

Preparation uses the study brief and permitted supporting material, such as documents you attach. Earlier conversations are used only with separate access and permission. From this the interviewer knows which questions are still open and checks them during the conversation.

During

The person joins by link in the browser and talks by voice or types. Each language has its own interviewer: Spanish speakers talk to Camila, English speakers to Sam, Russian speakers to Anya. The person knows who they are talking to: the participant notice explains that an AI runs the conversation, what is recorded, who receives the answers and how long they are kept. Asked directly whether it is human, the interviewer says it is an AI.

After

You choose one of two kinds of result: transcripts and summaries for your own analysis, or the same plus findings measured against the research question you set. Findings rest on verbatim quotes with speaker and time, so any line can be checked against the transcript. Preliminary material is marked as unconfirmed, and if the evidence does not answer the question, the result is reported as partial instead of being stretched to fit.

How the interviewer is tested

An interviewer that improvises needs more testing than a form. Narratia keeps a catalog of 149 hard situations: rudeness, nonsense, attempts to derail the conversation, requests for other people’s data, silence and “uh-huh” to everything. The interviewer talks with AI models playing real people, a separate model takes part in the review, and each new version is compared with the previous one on the same respondents. Testing catches some failures; real interviews can still reveal mistakes. That is why every finding stays tied to a quote you can verify.

Where an AI interviewer fits

• Surveys and social research: real interviews instead of open-ended survey questions.
• User research and customer discovery: why customers chose you, why they left, what gets in their way. See customer discovery interviews with an AI interviewer.
• Knowledge transfer and offboarding: what a departing or key employee knows, written up with quotes for review before a successor uses it.
• Journalism: accounts from sources and witnesses, with quotes tagged by speaker and time.
• Genealogy: voice interviews with relatives; facts for the family tree arrive as suggestions to check, not as finished entries.

Not sure an interview is the right tool at all? Compare the options in AI interviewer vs survey vs human interviewer.

What to look for in an AI interviewer

An AI interviewer is software that runs an interview the way a skilled human would: it prepares, asks questions, listens to the answers and decides what to ask next. What separates a good one from a scripted chatbot is adaptivity — the next question should come from what the person just said, not from a fixed list.

When comparing tools, it helps to check a few things:

• Adaptive questioning. Does it follow the thread of the conversation, or does it march through a prewritten script regardless of the answers?
• Neutral technique. A good interviewer avoids leading, judging and repeated questions, because each of these pushes the person toward a particular answer or closes a topic early.
• Checking understanding. It should reflect back what it heard and confirm it with the person, rather than silently assuming.
• Preparation. It should study the brief and any permitted material before the conversation starts, so its questions are grounded in the topic.

These behaviours matter in any field where knowledge lives only in people's heads — research, product work, knowledge transfer, journalism, genealogy — because the value of the interview depends on the person speaking freely and in their own words.

Narratia is built around exactly this. It asks the next question from what the person just said, reflects back what it understood, checks it with the person and asks for real cases instead of generalities. It is designed to avoid leading, judging and repeated questions, and these behaviours are evaluated in testing. During the conversation it can show a running summary and takes corrections on the spot, so misunderstandings are fixed while the person is still there. Examples on the site come from demo interviews with test personas. Break me if you can