QuestionPunkQuestionPunk
Start free

How AI interviews work

Understanding AI-driven conversational surveys.

Explains the mechanics of AI interviews: adaptive follow-ups, prompt engineering, context preservation across turns, and how the AI moderates across languages. Also covers privacy and moderation considerations.

aiinterviewsconversational10-30 minutesIntermediateResearchersProduct managers

Steps

  1. Add an AI interview question
    In the survey editor, click + Add and choose AI interview. In the two-step setup, enter what you want to learn, then choose Text chat or Voice under How should it be conducted?.
  2. Configure the AI interviewer
    Click the Interviewer to pick a personality or create your own (name, description, and example questions that guide the style). To read the prompt behind it, open AI model > Advanced prompt > Configure.
  3. Set the question mode
    Under Length, choose how the AI decides when to stop asking follow-ups: Fixed Count (set an exact number of follow-up questions), AI Decides (the AI stops when it has enough depth, guided by min/max limits and your stopping criteria), or Time-Based (interview runs for a set duration). Voice interviews are time-based only.
  4. Choose a model
    Select the AI model for the interview. New text AI interviews default to OpenAI GPT-5.4 Mini (fast, economical). On Pro you can choose from 163 models across 16 providers, including Claude Haiku 4.5, Sonnet 4.6 through Sonnet 5.5 (balanced), Opus 4.6 through Opus 5.5 and Fable 5.1 (deepest follow-ups), OpenAI (GPT-6, GPT-5.6/5.5/5.4, o3/o4 series), Google (Gemini 3.8/3.7/3.6/3.5), xAI (Grok 4.7/4.6/4.5), Meta (Llama 4), Mistral, DeepSeek, and Moonshot (Kimi K3). The Free plan offers Claude Haiku 4.5, GPT-5.4 Mini, and Gemma 4 31B. Existing surveys keep their saved model. See Choosing a model for trade-offs.
  5. Pick the response quality
    Under AI model, set Response quality for how text follow-ups are generated. Fast (default) generates a single question in one step, the most responsive option. Slow generates five candidate follow-ups, grades each with personality-specific criteria, then picks the highest-ranked one: higher quality for important interviews, at the cost of latency.
  6. Choose what it can refer to (optional)
    Click What it can refer to to let the AI use the respondent's earlier answers, or insights from this survey or other surveys, during the interview. The AI can use these references to ask more contextual follow-ups. You can also add per-respondent context that feeds into the interview dynamically.
  7. Monitor and review
    After launch, review AI interview transcripts in the Results tab. Each conversation shows the full exchange between the AI and the respondent.

AI interviews dynamically adapt follow-ups based on participant answers. The system probes for clarity and depth to produce higher-quality qualitative data than traditional open-text fields.

The AI preserves context across conversation turns, so follow-up questions reference what the respondent has already said. This creates a natural, interview-like experience.

Three question modes control the length of a text interview: Fixed Count (exact number of follow-ups), AI Decides (AI stops when it has enough depth, with configurable min/max and stopping guidance), and Time-Based (runs for a set duration). Voice interviews always run for a set duration.

Response quality controls how text follow-ups are produced. Fast (default) generates a single question in one step and is best when responsiveness matters. Slow generates five candidate questions, grades each with personality-specific criteria, and picks the highest-ranked one. Use it for high-stakes interviews where question quality matters most.

Text AI interviews work in all 130+ survey languages, and voice interviews in 25. The interviewer replies in the language the respondent selected, while keeping the interview goals you defined.

What it can refer to lets the AI use answers from earlier questions in the survey, so follow-ups can build on what the respondent already shared in previous sections. Per-respondent context can also be injected dynamically.

You can read the full prompt the interviewer uses, edit its context, add your own sections, and review every exchange in the response data.