Adaptive Interviews

The AI conductor doesn't just ask questions — it observes participant behavior in real time and adapts how it engages. During usability tests, it watches silently, detects when someone is stuck, steps in with just enough help, then steps back. No human moderator required.

Three Engagement Modes

Every segment of a study defines a base mode that controls how the AI interacts with the participant. The AI can temporarily escalate to a higher mode when it detects the participant needs help, then return to the base mode automatically.

Observe

The AI watches silently. The participant uses your product naturally — navigating, clicking, typing — while the AI records everything: page navigation, clicks, spoken thoughts (via microphone), and UI state changes.

Best for: Usability testing, task completion studies, watching real workflows.

Speak

The AI provides one-way verbal guidance. It can read a task aloud, clarify instructions, or offer a nudge. The participant listens and responds. If the participant starts speaking while the AI is talking, it immediately stops (barge-in) — the participant always has priority.

Best for: Task instructions, directed questions, gentle coaching when participants seem lost.

Talk

Full two-way voice conversation. The AI asks questions, listens to answers, and follows up in real time. This is the closest to a human research interview — the AI probes emotional moments, asks "why" questions, and explores unexpected insights.

Best for: Discovery interviews, JTBD conversations, debriefing after an observation segment.

Mode Escalation

Modes have a hierarchy: observe < speak < talk. The AI can escalate one level at a time:

Observe → Speak: Offer verbal guidance when the participant seems stuck
Speak → Talk: Start a conversation when the participant needs more support

Each segment defines a mode ceiling — the maximum mode the AI can reach. If a segment is set to observe, the AI can escalate to speak but never to talk. If set to speak, it can reach talk. The AI never exceeds the ceiling.

Stuck Detection

The AI runs heuristic checks — no API calls, no delay — to detect when a participant needs help:

Signal	What It Detects
Silence	No speech for 30+ seconds during an observe segment
Confusion	Keywords like "I don't understand", "this is confusing", "unclear"
Help request	"Can you help?", "How do I...?", "I'm lost"
Navigation loop	Same page visited 3+ times in 60 seconds
Quick bounce	Landed on a page for less than 5 seconds then left
Stale session	No navigation or speech for an extended period
High engagement	Multiple substantive responses suggesting readiness for deeper conversation

These signals fire instantly — the participant gets help within seconds, not minutes.

Bounded Interventions

When stuck detection fires, the AI starts a bounded intervention:

Escalate — Mode steps up one level (e.g., observe → speak)
Help — AI delivers a short, contextual message (1-2 sentences)
Return — After the message plays or the participant responds, mode automatically returns to the base level

Interventions have a 60-second timeout. If the AI's help message finishes and the participant is ready, it returns to the base mode immediately. The AI never stays escalated indefinitely.

Common intervention messages:

Silence: "Take your time — let me know if you need any help."
Navigation loop: "It looks like you might be looking for something. Can I help point you in the right direction?"
Confusion: "I can see this might be confusing. Let me help."
Help request: "Of course! What are you trying to do?"

For complex situations (multiple signals at once), the AI generates a custom contextual message based on the current task, recent transcript, and what the participant was trying to do.

Real-Time Observation Pinning

As the AI monitors the session, it can pin moments — flagging high-signal observations in real time:

Struggling moment: Participant encountered friction
Workaround: Participant invented a hack to get around a problem
Desired outcome: Participant expressed what they actually want
Surprising insight: Something unexpected that's worth capturing

These pinned moments are recorded with timestamps and context, then passed directly to the post-session signal extraction pipeline as strong priors — confirmed observations that help the AI produce more accurate analysis.

How It Connects to Signal Extraction

After a session ends, the conductor's real-time observations feed directly into signal extraction:

Pinned moments become strong priors in the extraction prompt, marked as "identified in real-time with full conversational context"
Completed goals are confirmed by direct observation, not inferred from transcripts
Mode escalations (interventions) indicate moments of struggle — these map directly to struggling moment signals
Navigation history and page snapshots provide behavioral context beyond what the transcript alone captures

This two-pass approach — real-time observation during the interview, then deep analysis after — produces richer, more accurate signals than transcript analysis alone.

Designing for Adaptive Behavior

When creating study scripts, you can control how the AI adapts:

Mode per segment

Set each segment's base mode to match your research goal:

{
  "segments": [
    { "id": "intro", "mode": "talk", "title": "Introduction" },
    { "id": "task", "mode": "observe", "title": "Complete checkout" },
    { "id": "debrief", "mode": "talk", "title": "Discuss experience" }
  ]
}

Advance conditions

Tell the AI when to move to the next segment:

url:/success — advance when the participant reaches a specific page
all_goals_completed — advance when all goals are marked done
Free-text hints — "advance when the participant has expressed their frustration with the form"

Custom step-up triggers

Define per-segment conditions for when the AI should escalate:

question — step up when the participant asks a question
confusion — step up on confusion signals
engagement — step up when participant shows high engagement

Max duration

Set a timeout per segment. If the participant hasn't advanced naturally, the AI moves them along — preventing sessions from getting stuck indefinitely.

Tips

Start with observe for usability tests. Watch first, then debrief with talk. The combination produces the richest signals.
Trust the stuck detector. The heuristics are tuned to avoid false positives — it won't interrupt participants who are thinking quietly.
Keep interventions natural. The AI's default messages are conversational, not robotic. Participants often don't realize they're being guided.
Review pinned moments. They're the AI's real-time highlights — check these first when reviewing session results.
Mix modes in a single study. A study that starts with talk (context), moves to observe (task), and ends with talk (debrief) leverages all three modes effectively.