AI-moderated customer & user research interviews

Qualitative research gives the "why" behind numbers, but a human moderator can run only a few interviews a day and each study takes weeks. An AI moderator follows a research guide, asks open questions, probes answers adaptively, works in dozens of languages and can run hundreds of interviews in parallel, with a synthesis ready within hours. It trades some interviewer intuition for breadth, speed and cost, and is best used next to, not instead of, human researchers.

Problem & value

• Today: product, UX and insight teams recruit participants, schedule slots, moderate each session, transcribe and code the material. Studies are therefore small (often 8-20 interviews), slow and costly, and multilingual research needs local moderators. Many questions are simply answered with a survey that lacks depth.
• AI value: no scheduling (participants join on demand, any time zone), consistent guide adherence, follow-up probing that surveys cannot do, native-language interviews, and automatic transcription plus theme synthesis. This allows qual at quant-like sample sizes.
• Evidence: vendors (Outset, Listen Labs, Conveo, Strella) report studies with hundreds of participants and turnaround in hours or days; most of these claims are vendor-authored and independent validation is thin [1][2]. Professional bodies stress that buyers should ask about explainability, bias and human oversight before adopting AI research services [3].

Variants

• Product discovery & UX: concept tests, usability sessions with screen share, churn and win/loss interviews, onboarding feedback.
• Market research agencies: brand, ad and packaging studies across countries with panel recruitment.
• Customer-success / CX: post-purchase or post-support interviews triggered from a CRM event; structured follow-up to NPS comments (see LANG-013).
• Public sector & NGOs: consultations and citizen-experience studies (human review of sensitive topics).
• Employee research: engagement or change-management interviews; beware workplace power imbalance and the Art. 5 emotion-recognition ban (see Risks).
• Not this use case: candidate screening (CONV-019, high-risk) and scripted outbound surveys (CONV-008).

How it works

1. research brief + interview guide (goals, topics, probes, stimuli)
2. recruitment / invitation (panel, CRM list, in-product intercept) + consent & AI disclosure
3. live interview: ASR
4. moderator LLM (guide state, probing, no-leading check)
5. TTS / text (voice, video with screen share, or chat; language auto-detected)
6. quality guards (off-topic, fraud/bot detection, safety escalation, time budget)
7. transcripts, per-interview summary, tagged quotes
8. cross-interview synthesis (themes, segments, counts, evidence links)
9. human researcher review
10. report / insight repository

• Moderator agent: an LLM holds the guide as a state machine (must-cover topics, optional probes), decides when an answer is rich enough, and rephrases to avoid leading or double-barrelled questions. Spoken interviews need low-latency streaming ASR/TTS or a speech-to-speech model so that turn-taking feels natural.
• Synthesis: the same pipeline as qualitative coding (LANG-019): codebook, quotes linked to transcript timestamps so every claim can be verified.
• Human-in-the-loop: researchers write and pilot the guide, review a sample of transcripts, validate themes against raw quotes, and own conclusions. Sensitive topics trigger handover or a stop-and-signpost response.

Data preparation

Data needed: research objectives and the interview guide, a participant list or panel with consent records (CRM, newsletter tool, customer database), stimuli such as concept images, prototypes or product videos, and prior research for context (reports in the DMS, old transcripts, survey exports).

Typical gaps: vague research questions, customer lists with duplicates and no consent flags, outdated contact data, previous findings scattered over slide decks and personal drives, and stimuli in unusable formats.

Preparation steps: 1. Sharpen objectives and turn them into a structured guide with probes (PREP-016 for gold-standard examples). 2. Clean and deduplicate the contact list and record consent and purpose (PREP-025, PREP-011). 3. Pull prior research into one searchable collection with metadata (PREP-009, PREP-010). 4. Define a code scheme for themes before analysis (PREP-019). 5. Pilot the guide with 3-5 internal test interviews and compare AI summaries with a human reading (PREP-018). 6. Plan pseudonymisation of transcripts that leave the team (PREP-022).

Tools: CRM export, a spreadsheet, the research platform's import, Whisper for own transcripts. Rough share of effort: 10-20 percent, low. See GUIDE-003.

Implementation paths

Which path for whom

• SMB: Path A. Mature AI-moderation research platforms exist, many with EU hosting, and a study is a standard, project-based process without deep system integration.
• Enterprise: Path A. Research is a commodity function and panels, translation and analysis come bundled; move to Path B only when interviews must be integrated into an internal insight repository or involve highly confidential concepts.
• Developer: Path B. Build a custom moderator on speech and LLM APIs when you need your own guide logic, panel handling or analysis pipeline.

Path A - Buy (recommended)

• Purpose-built platforms: Outset, Listen Labs (recruiting network, enterprise service), Strella, Conveo (Belgium-based, EU), Glaut (Italy, EU), Feedbk and Koji; classic research vendors add AI moderation to existing platforms (e.g. Qualtrics, UserTesting, Dovetail for the analysis side) [1][2].
• Check: EU data residency and DPA, voice/biometric handling, languages, guide control, quote-to-source traceability, fraud detection, export of raw data, recruiting included or not, model providers used (sub-processors).
• Choose A when: you want results in days without engineering; research is recurring.

Path B - Build on cloud APIs

• Dialogue: frontier LLM with tool calling for guide state (Claude, GPT, Gemini, Mistral from the EU); realtime voice via OpenAI Realtime API or Gemini Live API, or a cascaded stack with ASR (Deepgram, AssemblyAI, Azure Speech, Speechmatics - UK) and TTS (ElevenLabs, Azure) [4][5].
• Transport: LiveKit (open-source, EU-hostable) or Daily/Twilio for web audio/video rooms.
• Synthesis: long-context LLM with structured outputs plus an embedding index for quote retrieval.
• Choose B when: you need custom guides, own panel, CRM triggers or a white-label experience.

Path C - Self-host (open source)

• ASR/diarisation: Whisper large-v3 / faster-whisper, NVIDIA Parakeet or Canary, pyannote.
• LLM: open-weight models such as Qwen3, Llama or Mistral families (~30-70B for good probing quality), served with vLLM; TTS: Kokoro, Piper or similar open models; orchestration: LiveKit Agents or Pipecat (both open source) [5].
• Hardware: HW3 for concurrent voice sessions; text-only chat interviews run on HW2. Ops effort is real: latency tuning, voice quality, and evaluation of probing quality in each language.

Path D - Custom training

Rarely justified. Possible: fine-tuning a small model on expert-moderator transcripts to improve probing style or domain vocabulary, or a classifier for low-quality or bot responses. Requires some hundreds of annotated, consented transcripts.

State of the art (2026)

• Voice-first moderation with real-time turn-taking, interruption handling and 40+ languages is now standard in leading products; video interviews add screen sharing and stimulus walkthroughs.
• Reported effects: participants often disclose openly to a non-judgemental AI, and adaptive follow-ups yield richer answers than static surveys, but evidence is mostly vendor-produced; independent comparisons with human moderators are still few [1][3].
• Synthesis agents build themes with linked evidence and let analysts "chat with the study".
• Market consolidation: well-funded start-ups and acquisitions by research-platform vendors are reshaping the field; verify vendor status before contracting.
• Limits and failure modes: shallow probing on emotional, non-verbal or culturally subtle topics; leading or repetitive follow-ups; ASR errors with accents and noisy rooms; LLM summaries that overstate consensus; synthetic or fraudulent participants (people or bots farming incentives); self-selection bias in who accepts an AI interview; inconsistent quality across languages. Mitigate with piloting, transcript audits and evidence-linked reporting.

Shared resources & combinations

• Same stack (asr + llm + tts + streaming): CONV-002 voicebots, CONV-008 outbound surveys, CONV-018 digital humans; guardrails and evaluation harnesses are shared with CONV-001.
• Downstream: LANG-019 (qualitative coding of human-run interviews), LANG-016 (topic modelling), LANG-013 (feedback and NPS analysis). One speech stack on HW3 can serve all.
• Boundary: screening or assessing job candidates is CONV-019 and falls under high-risk rules.

Effort, cost & KPIs

• Path A: days to set up; pricing is typically per study, per completed interview or subscription; recruiting incentives are often the larger cost.
• Path B: 1-3 months for a reliable voice pipeline, evaluation set and consent flow.
• Path C: 3-6 months plus GPU operations.
• KPIs: completion rate, median interview length and depth (words per answer, probes per topic), guide coverage, share of leading questions (audited), participant satisfaction, fraud rate, time from brief to insight, agreement of AI themes with human-coded sample.

People & change

Involve management (which decisions will the study inform), the department lead who owns the question (marketing, product, UX) and the human researcher or analyst who validates the synthesis. IT sets up accounts and integration, the DPO and legal clear participant consent, recording, processor contracts and non-EU transfers before the first invitation. Customer-facing teams should know that their customers are being approached.

Works council: usually no, because participants are customers or panellists and not staff. It becomes relevant if employees are interviewed (e.g. internal user research) or if employee-level analytics arise (§87(1) no. 6 BetrVG); then also inform it under §90(1) no. 3. See GUIDE-004.

Data protection: interviews are personal data and voice recordings; inform participants, disclose that the moderator is AI (Art. 50 AI Act), ask for explicit consent, set retention limits, and run a DPIA if special categories could come up. Support AI literacy of researchers (Art. 4): limits of AI probing, hallucinated quotes, bias in themes. Pilot with one low-risk study and run the same questions with a human moderator for comparison. Typical concerns: "no empathy and shallow probes" (offer a human interview option, check transcripts), "participant trust" (be open about AI). Humans verify every quote and decide on conclusions. See GUIDE-009, GUIDE-005 and GUIDE-008.

Risks, regulation & compliance

• EU AI Act: typically limited risk: participants must be informed they are talking to an AI system (Art. 50(1)) and synthetic voices or avatars may need labelling. Art. 5 warning: inferring emotions from voice or face in the workplace or education is prohibited; do not use emotion-recognition features for employee research [6]. Using interviews for hiring decisions moves into Annex III high-risk (CONV-019).
• GDPR: voice, video and free-text answers are personal data and may reveal special-category data (Art. 9) without being asked. Define lawful basis (consent is typical), purpose limitation (research only, no re-use for marketing), minimisation, retention and deletion, DPA with all sub-processors, third-country transfer safeguards; run a DPIA for large or sensitive studies. Voice and face recordings must not be used for biometric identification [7].
• Research ethics: ESOMAR/ICC guidance and national codes (e.g. ADM, BVM in Germany) require transparency, voluntary participation, honest incentives and human accountability [3]. Minors and vulnerable groups need extra safeguards.
• Safety: distress or self-harm disclosures need defined escalation and signposting.
• Mitigations: clear AI disclosure and consent, opt-out for a human interviewer, no emotion inference, pseudonymised exports, evidence-linked synthesis, human review of findings.

Getting started

1. Pick a low-risk study (e.g. concept feedback, onboarding) and write a clear objective and guide.
2. Draft the consent text and AI disclosure; involve the data-protection officer early.
3. Trial 2-3 platforms (Path A) with the same guide, 20-30 participants each, and compare to a small human-moderated baseline.
4. Audit transcripts for leading questions, missed probes and fraud; refine the guide.
5. Define how themes are validated and how findings feed the insight repository.

Sources

1. Outset - AI-moderated research platform - https://outset.ai
2. Conveo - AI-moderated market research services overview - https://conveo.ai/insights/top-ai-moderated-market-research-services
3. ESOMAR - 20 Questions to Help Buyers of AI-Based Services - https://esomar.org/publications/esomar-20-questions-to-help-buyers-of-ai-based-services
4. OpenAI - Realtime API guide - https://platform.openai.com/docs/guides/realtime
5. LiveKit Agents (open source) - https://github.com/livekit/agents
6. Regulation (EU) 2024/1689 (AI Act), Art. 5 and 50 - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. Regulation (EU) 2016/679 (GDPR) - https://eur-lex.europa.eu/eli/reg/2016/679/oj