The interview agent runs the conversations. You configure the cohort.

Stage 01 is how the corpus fills. Our AI conversation agent runs structured fifteen to forty-five minute interviews with the people whose knowledge you want to capture: senior leadership, brand operators, store managers, franchisees, advisors, customers. Each conversation runs over voice through the browser, or over a regular cellphone call when the browser is not available. The unit cost is three cents per enriched conversation, end to end.

This page is the product view of that stage: how you launch a cohort, who gets interviewed, what the operator console shows you, and what lands in your corpus when the agent finishes a call. The interview protocol the agent follows (anchor-gated probing, neutrality enforcement, the wrap protocol) is documented under Methodology. The quality controls that audit the agent on every call are documented under Trust & Security.

From the operator console, you define a cohort: the stakeholder tier (a typical four-tier shape covers Tier 1 direction, Tier 2 brand operators, Tier 3 front line, and Tier 4 customers), the geographic scope, the role filter, and the cadence. The platform handles recruitment outreach, calendar coordination, and the consent flow. Workers and customers see the consent language before any conversation begins. The recruitment funnel is visible to you in the console at every step.

For a Pulse-tier deployment, the target is around 250 conversations in one market. For a Macro tier, around 1,000 across a region. For a large beauty-retail network, 50,000 across 35 markets on a weekly refresh. The platform handles all three scales on the same architecture. The Cloud Harness ships ready for the largest case at the smallest deployment.

The agent opens every interview on a concrete recent event in the respondent’s domain: “walk me through the last time a customer asked you for a product you did not stock.” The agent will not advance to the next theme until the anchor is locked, because abstract questions produce abstract answers and abstract answers do not enrich. Neutrality is checked on every utterance before delivery. The wrap follows a fixed closing sequence (crystallization, residual, farewell) so the last two minutes carry the highest-fidelity content of the conversation. The full protocol is on the Methodology page.

The mechanical comparison with adjacent capture methods is clear. A human qualitative researcher costs $400 to $1,200 per interview fully loaded, and every interviewer drifts. A survey runs at single digits per response but fixes a Likert ceiling on every answer. The agent runs at three cents per enriched conversation and executes the same protocol on conversation one and conversation ten thousand. The middle of the cost-fidelity curve did not exist until the AI interview agent arrived.

Every conversation the agent runs is audited by an independent audit agent against six protocol KPIs: context delivery, profile coverage, brand-perception coverage, no re-asking, neutrality, and graceful close. A conversation with poor execution has its intelligence discounted in the composite quality score. The full audit description is under the Interviewer Quality Audit.

The original audio enters through a zero-retention contract with ElevenLabs. Inside your tenant, an in-memory worker strips names, person identifiers, and any phrase on the redaction list. You edit the redaction list in the console: add the names of competitors, internal project codes, or anything else that should never appear in a transcript. The cleaned transcript is then rendered back through ElevenLabs as a synthetic voice. The original audio is never written to disk by any system we operate, and ElevenLabs discards it under the contract.

What lands in your corpus is the cleaned transcript and the synthetic-voice audio. Re-identification by voice requires a recording to exist. There is none. The full description of the chain is under Synthetic Voice Substitution.

A typical retail cohort runs the agent against four tiers of the same brand. Tier 1 (direction) is senior leadership, who give the strategic narrative the company tells itself. Tier 2 (brand operators) is marketing, communications, merchandising, and design, the people who execute the brand. Tier 3 (front line) is store managers and franchisees, who watch the brand land in the market. Tier 4 (customers) is the external shoppers who decide whether the brand actually works.

The same brand-perception question hits all four tiers with calibrated wording. The corpus then lets you compare the four answers as four facts with different segment vectors. “The brand is a value-fashion destination” might be true at Tier 4 today, contested at Tier 2 today, and was true at Tier 1 in 2023. That structural disagreement across tiers is the kind of finding the layer was built to surface, and it is invisible to surveys that only ever ask one tier.

The operator console shows you the cohort fill rate, the recruitment funnel at each stage, the conversations currently in flight, the per-conversation quality scores from the audit agent, and the running cost. A conversation that fails the pre-enrichment gate (under 60 seconds, fewer than six exchanges, low signal) is flagged with the gate reason and excluded from aggregates. You see it on the console. The intelligence record is preserved in the audit trail, never deleted.

You can listen to any synthetic-voice recording from the console (the original recording does not exist) and read the cleaned transcript side by side with the ten typed dimensions the enrichment pipeline extracted. Every quote in every dimension links back to its chunk identifier. Every enrichment call links back to its row in the audit trail. When you spot something that looks wrong, the path from the dashboard cell to the LLM call that produced it is two clicks.