How to Write a Discussion Guide for an AI Moderator (2026 Template)

Qualitati Research Team · 2026-05-19 · 10 min read

Last updated: May 19, 2026

Short answer

An AI moderator discussion guide is not a human script with the human removed. It is a structured object: a research goal, 4–7 sections tied to research questions, an opening prompt per section, 2–4 probe prompts with stop conditions, and explicit instructions for when to move on, when to push back, and when to escalate to a human. Write it that way and an AI moderator will run rigorous interviews at scale.

Why discussion guides need to change for AI moderators

Traditional discussion guides are written for humans who can read the room, improvise, and remember the research goal. AI moderators read instructions literally. They will probe forever if you do not tell them to stop, skip critical topics if you bury them in prose, and ask leading questions if your wording is loose.

A guide built for an AI moderator is closer to a structured config than a Word document. The good news: once you adopt that structure, you get more consistent interviews across hundreds of participants, easier downstream thematic analysis, and a guide you can iterate on like code.

Key takeaways

Anchor every section to a research question. If a section does not map to a research question, cut it.
Write opening prompts as questions, not topics. “Tell me about the last time you…” beats “Explore onboarding.”
Give the AI explicit stop conditions: max probes per section, signs of saturation, signs of participant fatigue.
Add a small set of escalation triggers (distress, off-topic loops, safety issues) that hand off to a human or end the interview gracefully.
Pilot with 5–10 participants before scaling. AI moderators surface bad guide design faster than humans do.

The 7-part AI moderator discussion guide template

Use these seven blocks as the skeleton of every guide. Each block is short on purpose; AI moderators perform better with terse, unambiguous instructions.

1. Research goal (one sentence)

State the single decision this study informs. Example: “Understand why first-week users abandon the import flow so we can prioritize fixes for Q3.” The AI uses this to judge when a tangent is worth pursuing.

2. Participant context

Tell the moderator who it is talking to and what is already known: role, recent behavior, segment, language. Keep it factual. This grounds personalization without scripting it.

3. Sections mapped to research questions

Pick 4–7 sections. Each section has: a research question (internal), an opening prompt (spoken), 2–4 probe prompts (conditional), and a transition line.

4. Probe library with stop conditions

For each section, list probes the moderator may use and the conditions under which it should stop probing and move on. Without stop conditions, AI moderators over-probe and burn participant goodwill.

5. Behavior rules

Define what the moderator should never do: lead, suggest answers, evaluate, give product advice, reveal hypotheses. Also define recovery behavior for silence, one-word answers, and off-topic drift.

6. Escalation and exit rules

Specify when to end early (distress signals, time cap, off-scope), how to thank the participant, and how to confirm consent for follow-up. For sensitive topics, route to a human reviewer.

7. Closing block

One open question (“Anything important we did not cover?”), one demographic confirmation if needed, and a clear sign-off. Always close cleanly — truncated endings hurt completion and data quality.

Worked example: onboarding study

Below is a compressed example for an onboarding research goal. In production each section would also carry a research-question ID for downstream coding.

Section	Opening prompt	Probes	Stop condition
First impression	Walk me through what you did the first time you opened the app.	What were you trying to accomplish? What surprised you? Where did you hesitate?	3 probes max; move on if user repeats themselves.
Import flow	Tell me about the last time you tried to import data.	What were you expecting? What did you do when it failed? What did you try next?	Stop when a concrete failure or workaround is named and described.
Help-seeking	When you got stuck, what did you do?	Where did you look first? Did you ask anyone? What would have helped?	2 probes max; do not suggest help channels.
Decision to continue	What made you decide whether to keep using it?	What almost made you stop? What kept you going?	Stop after a clear pivotal moment is described.

AI Interview Readiness Scorecard

Before you launch, score your guide on these eight dimensions, 0–2 each. A guide that scores below 12/16 is not ready.

Goal clarity: one sentence, one decision.
Research-question mapping: every section ties to a research question.
Prompt phrasing: opens with experience questions, not opinion questions.
Probe discipline: probes are listed; stop conditions are explicit.
Neutrality: no leading language, no product framing.
Escalation rules: distress, off-topic, and exit rules are written.
Length realism: estimated time fits the participant cap (typically 15–25 minutes for AI-moderated text or voice).
Pilot plan: 5–10 pilots scheduled before scaling.

Common failure modes (and how to fix them)

Over-probing: the AI keeps asking “Can you say more?” Fix: cap probes per section and add a saturation signal (“move on once the participant has named a specific instance with context”).
Leading questions: the AI inherits your hypothesis. Fix: remove product names and value-laden adjectives from prompts.
Topic drift: participants pull the conversation elsewhere. Fix: explicit transition lines plus a “return-to-section” rule.
Thin answers: one-word replies dominate. Fix: anchor openings in past behavior (“the last time you…”) rather than abstract opinions.
Missing data for analysis: sections do not map cleanly to research questions, so coding later is painful. Fix: enforce the research-question ID on every section.

Voice vs text guides

Voice interviews tolerate shorter prompts and benefit from more frequent acknowledgments; text interviews tolerate longer prompts but need stricter probe caps because participants drop off faster. For voice studies, also note any acoustic signals you plan to analyze later — see our overview of voice analytics in user interviews.

Where Qualitati fits

Qualitati is an AI user research platform designed around exactly this kind of structured guide. When you create a study, you define sections, research questions, probes, and stop conditions in one place; the AI moderator runs interviews in text or voice across 10 languages, and ThemeLens AI maps codes back to the research questions you wrote in the guide. Start free with 30 credits — no credit card required — and pilot your first guide with five participants before scaling.

Limitations and trade-offs

AI moderators are not a fit for every study. Highly sensitive topics (trauma, clinical, regulated industries) still benefit from human moderators or a hybrid model such as Qualitati’s Active Listener mode, where a human leads and the AI surfaces real-time prompts. AI moderators also depend on guide quality more than humans do: a vague human-written guide can be rescued by a skilled moderator on the fly; a vague AI guide produces vague data at scale. Build pilots into every study and treat guides as living artifacts you revise weekly.

FAQ

How long should an AI moderator discussion guide be?

Typically 4–7 sections, 1–2 pages of plain text when exported. Longer guides usually signal that research questions have not been prioritized.

Can I reuse my human discussion guide?

Use it as a starting point, but rewrite each section into the opening-prompt + probes + stop-condition format. Plain prose guides under-perform with AI moderators.

How many probes per section is too many?

Two to four. Beyond that, completion rates drop and participants start repeating themselves, which adds noise to analysis.

Should I tell the AI my hypothesis?

Share the research goal and decision context, not the hypothesis. Hypotheses tend to leak into prompts and bias responses.

Do I still need to pilot if the AI is consistent?

Yes. AI consistency means a bad guide produces bad data consistently. Pilot with 5–10 participants and revise.

What about multilingual studies?

Write the master guide in one language, then have the platform translate prompts and probes. Review translated prompts with a native speaker before launch — idioms in probes are the most common failure point.

Conclusion

An AI moderator is only as good as the guide it runs. Treat the discussion guide as a structured object — goal, sections, prompts, probes, stop conditions, escalation rules — and you get interviews that are both rigorous and scalable. Try the template in Qualitati with your next study and iterate from real pilot data instead of guesswork.