How to write a discussion guide for AI moderators
Learn how to structure questions, pacing, and branching logic so an AI moderator can run interviews that produce real insights.
How to write a discussion guide for AI moderators
A discussion guide for an AI moderator is a structured script that defines the questions an AI will ask, the order it will ask them, and the logic that determines which follow-up probes fire based on what a participant says. Writing one well is the difference between an AI interview that produces rich, usable data and one that feels like a survey with a chatbot skin.
This guide covers what to include, how to format it, and the specific differences between writing for an AI versus a human moderator.
Why the guide matters more with AI moderation
When a human moderator goes off-script, they draw on years of interviewing experience to decide what to probe and how. An AI moderator does not have that fallback. It follows the structure you give it. If you write vague questions, it asks vague questions. If you forget to include a follow-up probe for an important answer type, that thread simply does not get explored.
This places more responsibility on the researcher at the guide-writing stage. The quality of your output is largely determined before the first interview starts.
The anatomy of an AI moderator discussion guide
A well-structured guide has five components:
1. Session framing
The opening section tells the AI how to introduce itself and set expectations for the participant. This should include:
- A brief explanation of what the AI will do (“I’ll ask you some questions about your experience with…”)
- A statement about confidentiality and data use
- An instruction for the participant to speak freely and at length
- A confirmation that there are no right or wrong answers
Keep this section short. Participants are more willing to engage once they understand the format, but a long preamble loses them before the real questions start.
2. Warm-up questions
One or two low-stakes questions that get the participant talking before the substantive topics begin. For example: “Tell me a bit about your role and how long you’ve been in it.” These questions do two things: they calibrate the AI to the participant’s communication style, and they lower the participant’s guard before you ask about harder topics.
3. Core questions
This is the substance of your guide. For a 20 to 30 minute interview, aim for 6 to 10 core questions. Each core question should:
- Be open-ended (“Walk me through the last time you…”)
- Target one specific topic or behavior
- Include 2 to 4 pre-written follow-up probes
The follow-up probes are what separate a good AI guide from a mediocre one. For every core question, think through the most likely answer types and write a probe for each. If you ask “What made that experience frustrating?”, a participant might describe a workflow problem, a communication breakdown, or a tool failure. Write a different probe for each path.
4. Branching logic
Branching logic tells the AI how to route the conversation based on what a participant says. Not every platform supports complex branching, but most support at least basic conditional follow-ups: if the participant mentions X, ask Y; if they say they have never done X, skip to Z.
Document your branching logic clearly, even if you are using a platform that handles it through a visual interface. Writing it out in plain language first helps you catch gaps before you configure it in the tool.
5. Closing and debrief
End the guide with a standard closing sequence: an opportunity for the participant to raise anything they wanted to mention, a brief thank-you, and any instructions about next steps or compensation. The AI should not simply cut off the conversation when the last question is done.
Question types that work well for AI moderation
| Question type | Example | Why it works |
|---|---|---|
| Retrospective experience | ”Describe the last time you had to…” | Grounds the participant in a specific memory |
| Process walkthrough | ”Take me through how you typically…” | Reveals steps and decision points |
| Comparison | ”How does this compare to how you used to…” | Surfaces change over time |
| Prioritization | ”If you could only fix one thing, what would it be?” | Forces clarity and reduces hedging |
| Hypothetical (with caution) | “If that wasn’t available, what would you do?” | Useful for fallback behavior research |
Avoid questions that can be answered yes or no, questions that contain the answer (“Did you find it confusing?”), and abstract questions that don’t anchor to a real experience.
Formatting for AI consumption
Most AI moderation platforms read your guide as structured text or through a purpose-built configuration interface. Regardless of the format the tool expects, write your guide with the following conventions:
Label everything. Use clear labels like Q1, Q1a, Q1b to distinguish core questions from follow-ups. This makes it easier to review, revise, and troubleshoot.
Write tone instructions explicitly. If you want the AI to respond with empathy when a participant describes something difficult, write that instruction in the guide (“If the participant expresses frustration or negative emotion, acknowledge it briefly before asking the follow-up”). Human moderators pick this up intuitively; AI moderators need it spelled out.
Define fallback questions. For each core question, include a fallback that the AI can use if the participant gives a very short or unclear answer. Something like “Can you tell me a bit more about that?” is generic but often enough to keep the conversation moving.
Set a word limit expectation. Some platforms let you configure how the AI interprets silence or very short answers. If yours does, use it.
What to leave out of an AI moderator guide
Moderator notes and reminders to self. These are useful in a human guide, but they confuse AI parsing. Keep the guide clean and separated from your internal research notes.
Complex multi-part questions. Participants struggle with these in human interviews. In AI interviews, they are worse: the AI may only capture a response to the first part. Break compound questions into separate items.
Highly sensitive or crisis-adjacent topics without a human handoff plan. If your topic has any chance of surfacing distress, build in an explicit instruction for the AI to pause and provide a resource, or plan for human review before the interview ends.
Piloting your guide
Before you launch a study, run two or three pilot sessions. Review the transcript with these questions:
- Did the AI ask the right follow-ups, or did it stick to the script when it shouldn’t have?
- Were any questions confusing or interpreted in a way you didn’t intend?
- Did the session cover all the topics you needed?
- Were participants engaged throughout, or did engagement drop at a specific point?
Revise based on what you find. A guide that has been piloted and revised will consistently outperform one that goes straight to a full sample.
Connecting AI guide design to your broader research workflow
Writing a strong AI moderator guide is one part of a larger skill set. Understanding what AI-moderated interviews are and how they work gives you the foundation for making guide-writing decisions. Knowing when AI versus human moderation is the right choice helps you design studies that fit the method to the question.
For the question bank itself, 50 user interview questions with examples is a practical reference for building your core question set, and 40 examples of interview questions for qualitative research covers the broader range of question types applicable to AI-moderated contexts.
Platforms like CleverX combine AI moderation with a verified panel of 8M+ B2B and B2C participants across 150+ countries. This means you can design your guide once and run it at scale without the scheduling overhead of coordinating individual human-moderated sessions. The best AI-moderated interview platforms in 2026 covers your platform options in detail.
Frequently asked questions
What is a discussion guide for an AI moderator?
A discussion guide for an AI moderator is a structured script that tells the AI what questions to ask, in what order, and how to respond to specific answer types. Unlike a human moderator guide, it must be more explicit about branching logic and follow-up triggers, because the AI cannot improvise the way a skilled human researcher can.
How long should a discussion guide for an AI interview be?
Aim for enough content to fill 20 to 30 minutes of conversation. In practice, that means 6 to 10 core questions with defined follow-up probes for each. Guides that are too long overwhelm the AI with branching decisions; guides that are too short produce shallow data.
What types of questions work best for AI moderation?
Open-ended questions work best. Questions that ask participants to describe a process, recall a specific experience, or compare two options give the AI material to probe further. Avoid hypotheticals and leading questions, which tend to produce vague or socially desirable answers regardless of who is moderating.
Can an AI moderator handle sensitive topics?
AI moderators can handle moderately sensitive topics if the guide includes explicit instructions on how to respond when a participant signals distress or discomfort. For highly sensitive research, such as mental health or trauma-adjacent topics, a human moderator or a hybrid approach is more appropriate.
How is writing for an AI moderator different from writing for a human moderator?
Human moderators can read body language, pick up on hesitation, and deviate from the guide when something unexpected comes up. An AI moderator follows the script, so you need to anticipate likely answer paths and pre-write the follow-up probes for each one. Branching logic, explicit tone instructions, and fallback questions are required in an AI guide; they are optional in a human one.
How do I test my AI moderator discussion guide before launching a study?
Run at least two pilot interviews with colleagues or friendly participants. Review the transcript to check whether the AI asked appropriate follow-ups, whether any questions confused participants, and whether the guide hit the topics you needed. Revise before launching to your full sample.