An AI Interviewer That Asks About What Users Just Did

UserTold runs GPT-Live-1 voice interviews inside your product. A participant tries a real task, the interviewer asks about what they just did in a planned voice debrief, and the Evidence stays linked to the recording and page context it came from.

Run your first interview · See a debrief · Pricing

Powered by GPT-Live-1

GPT-Live-1 handles the natural full-duplex conversation, including interruptions. UserTold's delegated backend agent follows the Study, reasons over grounded page and product context, and decides when to use approved tools, so the spoken interview remains responsive without losing the research plan.

See a debrief

Task: Find the current plan and where to change it.

Observed: The participant opens account settings, goes back, and opens account settings again.

Interviewer: “When you returned to account settings, what were you looking for?”

Participant: “I thought the plan would be under my account. I didn't know it belonged to the workspace.”

Evidence for review: One path plus one explanation is a reason to look at where plan controls live. It is not proof that every user hits this, or that moving the controls is the fix.

For the broader product flow, read the research walkthrough.

How the live debrief works

1. Set the research goal. Choose what you need to learn and give the participant a neutral task.
2. Observe without interruption. With consent and the available permissions, UserTold captures speech, interactions, navigation, and page context while the participant uses the product. Screen recording is available on supported desktop browsers.
3. Prepare a running evidence summary. When Pre-debrief analysis is on for the planned debrief, UserTold periodically summarizes recent speech, product behavior, and page changes. It keeps three things apart: what the participant said, what the product recorded, and what the summary infers.
4. Ask grounded follow-up questions. The debrief opens from the most recent summary and the study's goals, and the interviewer picks each question from that context and the answers so far. With no summary, it falls back to general questions instead of guessing at what happened.
5. Review evidence after the interview. Separate post-interview processing extracts source-linked Evidence. Review the recording and grouped Findings before deciding whether to send anything to a tracker.

Pre-debrief analysis prepares the debrief; it does not interrupt an observation segment to rescue the participant. The Study Runtime reference explains the handoff and recovery behavior.

What can I learn?

Your question	Start here
Why do new users stop during setup?	Investigate onboarding and activation.
What need is behind a feature request?	Identify user needs.
Why is choosing a plan difficult?	Research your pricing page.
Why do users keep their old workaround?	Learn why users switch.

UserTold interviews the users your team can already reach. It does not recruit participants or guarantee responses. Use the method-selection guide to decide whether an in-product interview fits your question.

Set up an interview

Create a Project and add its script to your site once. The new Project comes with a starter interview; build a Study when you need a specific task, observation, and debrief. Check its goals, flow, invitation, and visibility, then run the whole thing yourself before you invite anyone.

For an observation study, write the task instruction, add an Observe segment, and follow it with a Talk debrief. Turn on Pre-debrief analysis for that Talk when its questions should draw on what just happened. Research Script Templates has tasks and neutral questions to adapt; Quickstart covers installation and verification.

Use the dashboard, or connect an agent through MCP or the CLI. Linear and GitHub are optional.

Questions before you start

Does it ask everyone the same questions?

The Study sets the goals and the flow. Inside a Talk segment the interviewer follows up on what the participant says and on the evidence it has been given. Scripted introductions and task instructions stay scripted, and the debrief is a planned segment rather than a reaction to every pause.

What happens without screen sharing?

Screen capture depends on the browser and on the participant's permission. Voice and product events still give you context, but they are not a stand-in for a recording you do not have. Check what was actually captured before you treat anything as recorded evidence.

Can users decline recording?

Participants see a recording disclosure and choose whether to start, and they can stop the interview at any point. Your invitation should say what the research is for and that taking part is voluntary. See Participant Consent and Privacy.

What does the team receive?

The Interview record, the transcript and whatever behavior context was captured, and Evidence once processing finishes. Related Evidence can support draft Findings. A person or a project-aware agent reviews the source alongside the synthesis; marking something reviewed does not open an issue anywhere. See how to analyze interviews.

What does it cost?

UserTold charges for recorded interview time. See current pricing for Managed AI and bring-your-own-key options before running a Study.

Run one focused study

Pick the question you need answered, set up your first interview, and review the source before you invite anyone else. Start with one real workflow and follow the explanation through to a decision.