AI-moderated interviews help you collect rich qualitative feedback at scale without scheduling sessions, coordinating calendars, or manually moderating every conversation.
This guide walks you through everything you need to launch your first study, from project setup to analysis, plus the best practices you should know before going live.
We've broken it into the following sections:
Step 1: Create and configure your project
Your project is the home for the study. Every completed interview uploads and stores here automatically.
Set Up a Project
Create or select an existing project in Marvin. All completed participant interviews will automatically upload and store directly here.
Configure Settings
Set up your project-level AI settings and toggle on automatic PII (Personally Identifiable Information) plotting.
Organize with Tags and Labels
Establish file tags (e.g., for personas or regions) and note labels (e.g., for themes or features) ahead of time. Consistent tagging and a clear labeling taxonomy support robust segmentation, better AI annotation, and cross-project insights.
Learn more:
Step 2: Design a structured discussion guide
A well-structured discussion guide is the foundation of a successful study, and the clearer it is, the smoother the experience for participants.
Keep it Single-Topic
Focus on one specific topic at a time. Avoid "double-barreled" questions that mix multiple topics into a single prompt.
Use the forward slash
Put a forward slash in front of each question you want the AI to ask verbatim. This tells it to ask the question aloud and link the response to it, which also aids analysis. Text without a slash becomes background context for the AI.
Add Warm-Up Scripts
Include introductory text or technical instructions at the beginning of the guide to welcome the participant.
Test a prototype (optional)
In the discussion guide settings, click "Add Prototype" and paste a link (Figma, Lovable, or a custom URL), making sure permissions are set to "anyone with the link can view." During the session, the AI moderator shares its own screen to display the prototype, so participants don't need to log into their own accounts.
Learn more:
Step 3: Set up and customize your interview
Go to your project, click Add New Files, and select AI Moderated Interview. This one screen covers general info, moderator behavior, probing, and compliance.
General info and participant requirements
White-label the interview with your logo, name the requester, and enter an interview topic (a synopsis the participant and the AI both see). Choose whether video, screen sharing, name, and email are mandatory, optional, or off, or keep the session anonymous.
Response format
Turn-based (highly recommended): The AI asks a question and pauses completely; the participant thinks, records, then clicks a button (or key) to move on. It gives people time to think and is highly resilient to background noise.
Real-time conversation: A natural back-and-forth where the AI detects when the participant stops speaking. Requires a quiet environment for best results.
Voice, subtitles, and context
Pick an AI moderator voice and enable subtitles. Add detailed research goals and background (audience, terminology, what you want to learn). The more context you provide, the smarter the AI's follow-up questions. It also switches languages automatically if a participant responds in another language.
Branching logic
Beyond a single linear script, you can build multi-layer branching so the interview adapts to each participant's answers. Use it to route respondents down different question paths based on what they say, skip sections that don't apply to them, or probe specific segments more deeply. This keeps each conversation relevant and respondent-driven.
Probing level per question
Keep it brief: The AI asks the question and moves on.
Probe a little: 2 to 3 organic follow-up questions based on the response.
Probe more: 3 to 5 follow-up questions to dig deeper.
Advanced and compliance (optional)
Add legal or compliance opt-in language, auto-add participants to your research panel, set an end date or response limit, and provide a redirect URL to a thank-you page, survey, or incentive site.
Learn more:
Step 4: Pilot, prepare, and launch
Pilot internally
Run through the interview yourself to check pacing, tone, follow-up behavior, and prototype visibility. Refine the guide and prompt until the flow feels right.
Prepare participants
Ask them to use a quiet space and headphones, and tell them in your invitation that the study is AI-moderated so no one expects a human to join. You can also have the AI introduce itself as an AI.
Generate and distribute the link:
Click Create for your unique study link. Share it by email, post it in online spaces (Reddit, LinkedIn), or embed it in-app. Participants complete it whenever they like, no scheduling required.
Step 5: Manage and analyze results
Manage active studies
Go to Manage AI Moderated Interviews to view, manage, or turn off active interviews.
Automatic saving
Completed sessions upload back into your project automatically, fully transcribed.
Analyze with AI
Use Marvin's AI to query across all your data, locate specific comments, and run thematic analysis. It relies on the timestamped conversation rather than manual notes, which reduces bias.
Learn more:
Step 6: Essential best practices
Keep sessions under 30 minutes: The AI can get “wonky” after about 30 minutes. For complex, multi-hour, or group studies, use a human researcher.
Default to turn-based mode: It gives participants time to think, avoids pressuring them, and handles background noise far better than real-time.
Write a structured, single-topic guide: Avoid double-barreled questions, and use the forward slash for questions you want read verbatim.
Provide deep context and clear goals: The AI is only as good as its prompt. Rich research goals and context produce smarter, more natural follow-ups.
Test and pilot internally first: Refine pacing, tone, and instructions before sending the link to real participants.
Set expectations early: Tell participants up front they'll be interviewed by an AI and that no human will join.
Mind usability-test limits: Design prototype tasks to be straightforward and fully functional so participants don't get stuck.
Lean on AI for sensitive, global, and multilingual research: Participants often feel less judged by an AI, and it adapts to time zones and languages automatically.