bmc-ai-scaled-customer-interviewing

Uses AI-moderated interviews to run customer discovery conversations at a scale no human team could match — hundreds of short interviews in parallel — then treats the aggregate as a filter: mines it f

SKILL.md

GitHub

BMC AI-Scaled Customer Interviewing

Purpose

Direct customer interviews are the strongest evidence a hypothesis can get, but a human team can only run so many before time runs out. This skill uses AI-moderated interviews to remove that ceiling — running many short conversations in parallel — and treats the resulting mass of transcripts as a filter, not a final answer: mine it for the most interesting outliers, contradictions, or recurring patterns, then call those specific people back personally. Use this as a first, wide pass before or alongside the personal follow-up interviews bmc-proxy-expert-validation and direct customer research already call for.

Anchored in research

Grounded in a well-documented, mainstream 2026 practice, independent of any single named source: roughly 40% of B2B SaaS product teams now report AI-moderated interviews monthly, up from under 10% in 2024; early-stage founders in 2026 run a median of 47 completed interviews per discovery round, up from 8–12 in 2022, enabled specifically by AI moderation removing the scheduling and synchronous-time cost; a typical AI-moderated conversation runs under $5 in compute. Multiple named platforms exist for this specifically (Perspective AI and others), and the underlying mechanism — a software moderator running many sessions in parallel, asking unscripted follow-ups, and synthesizing themes across every transcript — means research throughput is no longer capped by researcher headcount.

Method

  1. Design the interview for AI moderation, not just transcription. The value here isn't automating note-taking on human-run calls — it's a software moderator that asks unscripted follow-up questions live, based on what the respondent actually says, the same way a skilled human interviewer would probe an interesting answer. Confirm the chosen tool actually does this, not just records and summarizes.
  2. Recruit at volume, using low-friction channels — a newsletter invite, an embedded flow on a signup or thank-you page, an outbound message offering a short (5-10 minute) conversation. The low per-conversation cost only pays off if recruitment volume is genuinely high; a handful of AI-moderated interviews doesn't unlock the advantage this method offers.
  3. Run interviews in parallel, at a volume a human team couldn't match — tens to hundreds of conversations depending on the audience's size and reachability, not a handful.
  4. Treat the aggregate as a filter, not a conclusion. Don't try to read every transcript individually — instead, synthesize across all of them for: the most frequently recurring theme, the sharpest contradiction between what different respondents say, and the most surprising individual outlier that doesn't fit the pattern.
  5. Personally follow up on what the filter surfaces, not on a random sample. Call back the specific respondents behind the most interesting outliers or the clearest pattern-breakers — these personal, human-led follow-ups are where the deepest insight actually comes from; the AI-moderated pass exists to find WHO is worth that follow-up, not to replace it.
  6. Check for AI-moderation-specific bias before trusting the aggregate. Respondents may behave differently with a software moderator than a human one — more candid on sensitive topics, less candid on ones where they'd normally read social cues from an interviewer. Note this explicitly as a limitation of the aggregate data, not just of the individual transcripts.
  7. Compare this method's role to the pack's other validation sources, so it isn't used as a like-for-like substitute: bmc-proxy-expert-validation gets pattern-rich signal from adjacent professionals, not target customers directly; bmc-experiment-method-selection decides whether to build something or use a proxy test at all. This skill's distinct contribution is volume — direct customer conversations at a scale personal outreach alone can't reach.

What this skill does NOT do

  • Doesn't replace personal follow-up interviews — it's explicitly designed to identify WHO to follow up with personally, not to substitute for that follow-up.
  • Doesn't guarantee representative sampling just because the volume is high — recruitment channel bias (who sees the invite, who's willing to talk to a bot) still applies and should be checked, not assumed away by volume alone.
  • Doesn't work well for topics that genuinely need human trust or rapport to surface honest answers (deeply sensitive, high-stakes, or relationship-dependent topics) — use judgment about which questions suit AI moderation and which don't.

Refinement notes

  • Which AI-interview tool or approach has actually produced the cleanest signal in your own practice?
  • What's a real case where the aggregate filter surfaced an outlier worth a personal follow-up that a small human-only interview batch would have missed entirely?
  • How do you personally handle the AI-moderation bias risk (Step 6) — have you seen a clear case of respondents behaving differently with a bot vs. a human interviewer?

Continue from here

  • Use alongside: bmc-proxy-expert-validation/SKILL.md — a different, complementary customer-understanding source (adjacent professionals, not target customers directly).
  • Use alongside: bmc-experiment-method-selection/SKILL.md — decides whether to build or use a proxy test; this skill is one form of cheap, fast direct-customer testing within that decision.
  • This pack's shared guardrails: ../../CLAUDE.md

References

  • ../../references/bmc-source-material-notes.md — source material background
  • ../../../../ai-strategy-and-governance/references/ai-native-reshuffle-heuristics-research.md — selection and grounding notes for this skill and its siblings
  • ../../CLAUDE.md — this pack's shared guardrails