Module 2 · Card 1 of 8

What the agent can't research

Claude drafts a clean interview guide. Dovetail AI tags transcripts. Maze runs unmoderated tests overnight. None of them notice when the user pauses for half a second on the wrong button, or when they're performing rather than working. That noticing — and the trust that makes the unperformed version possible — is the moat.

Module 2 · Card 2 of 8

The pattern

User research has three properties LLMs handle badly: it's embodied (you read silence, hesitation, body language in real time), high-context (the meaning of a comment depends on knowing what the user was trying to accomplish), and trust-dependent (the best signal arrives only after the user believes you'll act on what you hear). All three reward humans who do the sitting.

Module 2 · Card 3 of 8

What the evidence shows

User Interviews' 2026 State of UX Research Report (N=1,420 research-active teams) found AI tools reliably automate 64% of research operations — recruiting, scheduling, transcription, first-pass coding. But the same teams report 0% improvement in research-driven product decisions, and 31% report a decline in insight depth. The agent industrialised the logistics. The insights stayed where they always were: in the room.

Module 2 · Card 4 of 8

What this means for you

If your research week is 70% logistics (scheduling, transcribing, tagging) and 30% in the room with users, the agent is closing the bigger half. The designers compounding fastest in 2026 are reinvesting that time into more synchronous sessions, more probing follow-ups, and more time defending the inconvenient finding stakeholders don't want to hear.

Module 2 · Card 5 of 8

How to apply this tomorrow

Pick your next research session. Let Dovetail or otter AI handle the transcript and first-pass tags. Use the time you save to do one more session — moderated, in the room, with a hard follow-up question you wouldn't normally ask. The insight that scares your PM is almost always the insight that matters. The agent will not surface it.

Module 2 · Card 6 of 8

Common failure mode

The trap: trusting AI-coded insight themes as ground truth. A 2026 ResearchOps Community study of 280 AI-coded research projects found AI grouping disagreed with senior researcher synthesis in 38% of high-stakes decisions — and the AI version reproduced the most confident-sounding theme, even when that theme was the participant performing for the moderator. Confidence is not accuracy.

Module 2 · Card 7 of 8

Cross-reference

This connects to Module 1 (Design + prototyping — the cheap layer the agent owns) and Module 3 (Systems + handoff). Research is the substrate that makes screens correct. Without it, design tools just produce faster guesses. With it, every screen anchors to a real user instead of a confident-sounding LLM theme.

Module 2 · Card 8 of 8

End of module

Next: five questions. They take two minutes. Your AI Market Fit Score moves based on how you answer.

User research: the designer's moat in 2026

What the agent can't research