The conversation agent is the only instrument that holds cost and fidelity at the same time.

Every conversation that enters the platform is a fifteen-to-forty-five-minute voice interview with our AI conversation agent. The agent opens on a concrete event in the worker’s recent experience, gates each probe against an anchor, never advances past a theme until the anchor is locked, and closes on a structured wrap. The output is a transcript, a typed field set, and twenty to forty enrichment-ready chunks.

The cost is three cents per enriched conversation. The protocol on conversation number one is the protocol on conversation number ten thousand. That combination is what makes the ten-dimension layer above the conversation possible. A human qualitative researcher produces fidelity at four hundred to twelve hundred dollars per interview. A five-point survey produces scale at the cost of throwing away ninety-five percent of what the worker knows. The middle did not exist until the AI-guided structured interview arrived.

An abstract question produces an abstract answer. The advisor who is asked “what do you think of our assortment” will tell you the assortment is fine. The advisor who is asked “walk me through the last time a customer asked for a product you did not stock” will tell you about the foundation shade a Tier-3 customer wanted last Tuesday and the substitute they bought reluctantly down the street. The first answer enriches into nothing. The second answer enriches into a pain point, a substitution edge in the brand graph, a market signal, and a customer-segment behavior fact.

The interview agent opens every theme on a concrete event and refuses to advance until the anchor is locked. The anchor is the interview’s evidence base. Every claim downstream traces back to a moment the worker actually witnessed. Anchor failure is a hard failure in the audit. It blocks the score for the conversation and surfaces in the operator dashboard the day the interview lands.

A leading question kills the brand-perception map. Ask “don’t you think the new fragrance launch is well-positioned” and every response biases toward yes. The neutrality protocol is the set of rules that keeps the agent from doing this. No value-loaded adjectives in a probe. No implied desired answer. No social-desirability primes (“most advisors find…”, “our best stores tell us…”). Before the agent delivers any next utterance, the utterance runs against a neutrality classifier. A flagged utterance is rewritten before it reaches the worker.

Neutrality also de-frames evaluation. The conversation is not a performance review. The interview script names this explicitly: the worker hears, in the opening, that nobody on their management chain will see what they said, that the audio is replaced with a synthetic voice before storage, and that the system is interested in what actually happens on the floor, including the things that do not match the official story. Workers who hear that opening on a real call disclose at depth on the kinds of topics that surveys and manager-led conversations systematically miss.

The last two minutes of an interview produce the highest-fidelity content. The worker has warmed up, the instrument is trusted, and the residual thought (the thing they were going to say but did not get to) surfaces on the way out. A faded ending throws that content away. The wrap protocol is three structured beats: a crystallization (the agent repeats back the one or two claims the worker put forward most clearly), a residual (the agent asks if there is anything the worker wanted to say and did not), and a farewell. The audit grades the close on those three beats. A graceful close is worth four points on the execution score.

This is the part of the methodology that surprises most operators. It is also the part with the strongest empirical record. The literature on self-disclosure to conversational AI, surveyed across twenty-six controlled studies, finds a consistent pattern: when the topic carries any social weight (a friction with a colleague, a workaround that diverges from policy, a customer complaint the worker did not log), people disclose more to a structured AI interviewer than to a human one, and far more than to a written survey.

The mechanisms behind the effect are named in the research. The AI agent carries no evaluative capability the worker has to manage. There is no career consequence in the room. The anonymity (the worker knows the audio is replaced with a synthetic voice before storage, knows the transcript never reaches their manager by name) reduces impression management. The conversational format builds rapport that a survey cannot. The voice modality matters: workers self-disclose more when speaking than when typing, and more when listening to a voice than when reading text on a screen. Each of these effects is documented across multiple independent studies in the published review.

This is why we run voice. A typed chatbot would lose the disclosure depth. A manager-led interview would trigger social-desirability bias. A survey would fix the format ceiling at five points and throw the rest away. The structured voice agent is the only instrument that gets the worker to talk at depth, on topics that carry social weight, at protocol-grade fidelity, at three cents per conversation. The choice of interviewer is the choice of how much the worker tells you.

One nuance from the same literature. Embodiment (a face on the agent) tends to reduce disclosure on sensitive topics. The face reintroduces the very evaluation cue the anonymity removed. The agent has no face. It is a voice, a name, and a protocol. The choice is deliberate.

Voluntary participation skews to engaged workers. A corpus built only from the people who said yes to an email blast is a corpus of advocates. The methodology requires a sampling design that goes in before capture begins: stratified by tier (direction, brand operator, front line, customer), with quotas per tier, and with outreach scripts calibrated for the worker who would not normally raise their hand.

In an early commercial deployment, the sample covered four tiers across direction, brand operators, store managers and franchisees, and external customers. For a large beauty-retail deployment running across many markets, the sampling design adds geography, store format, and customer segment as quota axes. The sampling design ships with the engagement. It is in place before any voice call is scheduled.

The agent will not advance past an unlocked anchor. The execution score penalises the conversation if it tries. The agent will not re-ask a question the worker has already answered (this is the highest-weighted KPI in the audit: seven points). It will not deliver a probe that failed the neutrality classifier. It will not invent a quote. Every key verbatim is verified as a literal substring of a real worker turn before persistence, and any quote that fails the substring check is deleted rather than rewritten. It will not flatten the worker’s Quebec-French register into Parisian French or into generic English. The source-language verbatim stays in its original register as the provenance record.

A mechanism in the pipeline enforces each refusal. The neutrality classifier blocks the probe. The hallucination filter deletes the invented quote. The cross-contamination check clears any brand-perception field that mentions a brand the worker never named. The agent is one part of the instrument. The audit and the filters are the other part. Both ship together.

A human interviewer drifts. Every qualitative researcher primes a little differently on a Friday afternoon than on a Monday morning, leads a little differently in the twenty-third interview of the week than in the first. That drift is a real ceiling on human-run capture at scale. The AI agent does not drift. The system prompt that configures the agent is verified by SHA-256 hash before every call. If the deployed hash does not match the registered version, the call fails closed. The protocol cannot quietly change between deployments.

In an early commercial deployment, the same protocol ran across the full sample, across four tiers, with measurable variance only on the dimensions we expect to vary (richness depends on what the worker actually had to say). On a large beauty-retail deployment running across many markets, the same property holds. The instrument is the same instrument in Quebec, in Paris, in Sao Paulo, and in Tokyo, on the same day and a year apart.