A qualitative text interview starts with one open question, then uses AI to ask focused follow-up questions based on what the respondent has already said. It is useful when a static open end would capture the headline answer but not the reasons, experiences, or trade-offs behind it.
The interview remains bounded: you choose the maximum number of follow-ups, and the AI can finish earlier when the research objective has been covered.
Choose the right text instrument
Use a regular text_question when every respondent should receive one fixed prompt. It is the better choice for short recall, feedback, names, labels, or an answer that only needs validation or clarification.
Use a qualitative_text_question when the next question should depend on the respondent's answer. Good uses include exploring a decision, reconstructing an experience, understanding uncertainty, or probing a trade-off.
Do not use a qualitative text interview merely to turn one long question into several turns. Each possible follow-up adds respondent effort and live AI processing.
Design the interview
Write a useful opening question
The opening question is respondent-facing. Make it neutral, specific enough to start a useful conversation, and broad enough to allow an unanticipated answer.
For example, "What makes choosing a current account difficult?" creates room to discuss fees, trust, comparison, eligibility, or previous experiences. "Why are bank fees confusing?" assumes both the topic and the respondent's opinion.
Treat instructions as the research brief
instructions guides the AI interviewer but is never shown to the respondent. State the objective, useful areas to explore, and important boundaries. Do not write a script or prescribe a fixed sequence: the AI receives the answered transcript and chooses one focused next question at a time.
Good instructions are concise and directional:
Understand the moments that create uncertainty and the trade-offs people make.
Explore concrete examples, but do not ask for account numbers or other financial identifiers.
Avoid instructions that demand a conclusion, repeat the opening question, combine several questions in each turn, or encourage collection of data the study does not need.
Set a sensible follow-up bound
max_follow_ups is the maximum number of AI-generated questions after the opening turn. The default is 3, and 0 produces an ordinary single-turn open end without requesting an AI follow-up.
Start with one to three follow-ups for most survey studies. Increase the bound only when the additional depth justifies the extra respondent time and AI processing. The AI may stop before reaching the bound when the opening question and instructions have been sufficiently explored.
Use a stable identifier
Set id so every turn shares a stable reporting ID. This keeps the interview connected across survey revisions and makes its reporting identity explicit.
Program the interview
s.qualitative_text_question(
"What makes choosing a current account difficult?",
instructions=(
"Understand the moments that create uncertainty and the trade-offs "
"people make."
),
max_follow_ups=3,
id="current-account-difficulties",
)
The generated Survey API reference is authoritative for the current signature and parameter types.
What respondents experience
The respondent first sees the opening question and enters a text answer. In a live interview, MX8 Labs sends the opening question, researcher instructions, and the complete answered transcript to OpenAI. The AI either returns one focused follow-up or marks the interview complete.
After each follow-up answer, the same process can run again until the AI finishes or max_follow_ups is reached. Only one follow-up is presented per turn. Reaching the bound ends the interview without another AI request, and the survey continues to its next element; there is no configurable closing script.
If a respondent leaves and resumes, MX8 Labs reconstructs the interview from its persisted, ordered turns. Previously answered turns and already-generated follow-up decisions are reused so answers are not lost or duplicated.
Test before fielding
Simulated surveys follow a deterministic interview path and make no external AI calls. They are suitable for checking routing, persistence, bounds, and downstream reporting, but their placeholder questions and answers do not predict the quality or wording of live probing.
Live interviews use OpenAI to decide whether to ask another question and what to ask. Review several live test interviews before launch to check that the opening question and instructions produce focused, non-leading probes. Each live follow-up decision is an external AI call, so the configured bound also caps the number of possible calls for that interview.
Analyze the results
Reporting treats the interview as one open-ended response per respondent, not as a separate response for every turn. MX8 Labs orders the turns and assembles the complete question-and-answer transcript before open-end coding, base calculation, and Example Verbatims processing.
This means:
- The analytical base counts each respondent once for the interview.
- The coding model receives the complete interview context.
- Example Verbatims uses the assembled transcript rather than selecting an isolated answer.
- Raw response outputs retain the individual turn rows, exact prompt, and turn order for audit and detailed analysis.
See Open end coding in MX8 Labs, Example Verbatims reports, and Exporting Respondent Transcripts for the downstream workflows.
Handle disclosure and sensitive text
Tell respondents that an AI interviewer will process their answers and may ask follow-up questions. Use consent and disclosure language appropriate to the study, audience, and jurisdiction.
Free text can contain personal or sensitive information even when it was not requested. Ask only for information the study needs, tell respondents what not to disclose where appropriate, and put those boundaries in the researcher instructions as well. Apply the same access, retention, and export controls used for other raw respondent data.
Current limitations
Qualitative text interviews are bounded, asynchronous text interactions. They do not provide live moderator takeover, voice or video interviewing, a configurable closing script, or a fixed multi-stage discussion guide. The AI chooses focused probes from the opening question, instructions, and answered transcript; researchers should review the resulting transcripts and retain responsibility for the final interpretation.