What the voice agent owes the person answering

Most survey tooling treats the respondent as a source of rows. The conversation is the product here, so the person on the other end can interrupt it, question it and correct it — and each of these is an action the agent can actually take, not a tone of voice it was asked to adopt.

• “It answers your questions too.”

  Who sees my answers? How long will this take? Who is actually asking? The agent answers from the study’s own versioned record — the disclosure the client committed to before anyone was invited — not from whatever the language model considers plausible. If the fact is not on the record, the agent says so rather than inventing one.

• “Change your mind, mid-call.”

  Go back and correct an earlier answer, decline a question without ending the conversation, ask for the question in simpler words, or ask the agent to slow down. Each of those is a real action with a record behind it: a revised answer is stored as revised, and a declined question is stored as declined rather than quietly counted as unanswered.

• “Complaints become numbered commitments.”

  When someone raises a real problem mid-conversation, the agent registers a numbered complaint, speaks the reference number back to them, and states the turnaround the organisation has committed to. The number is minted by the platform before it is spoken, so it is a commitment on a clock and not a courtesy.

The respondent-facing promises above are not a separate goodwill exercise. A person who can decline a question does not invent an answer to it, and a person who can correct themselves leaves you a better record — so the conduct and the data quality are one mechanism seen from two sides.

• “Quality enforced during the interview, not filtered after.”

  An answer the agent could not actually hear is re-asked in the moment instead of being cleaned up in a spreadsheet later. Confidence, length and speech-duration floors decide whether something counts as an answer at all; coverage of the mandatory questions is tracked as the conversation runs; abuse is flagged as it happens.

• And the gate that decides billing is the same one you can audit.

  One record per conversation holds the inputs and the verdict: consent, duration, coverage of mandatory questions, quality, duplicate detection. Only those hard rules decide what counts against what you bought — the softer quality scores used for internal review have no path into that decision.

• Every one of these actions leaves a typed event behind it.

  A courtesy nobody can prove is a marketing line. Each action above writes a typed event on the conversation and an audit entry, so “the respondent declined question 4” and “a complaint was raised in this call” are facts in the record rather than an impression of the transcript.

These are on the current work, they have no date on this page, and nothing here is available to a study today. They are named because a page that lists four live promises and quietly omits two reads, to anyone who later sees the full list, as a page that was hiding them.

• “It knows when to stop.” — coming soon

  Today the agent can flag a safety disclosure as a typed event and hand the conversation to a human. Spoken helpline resources, pausing the interview outright, and escalating to a named person on duty are next. We are building it deliberately slowly, because the human side comes first — who answers, in which hours, with which helplines — and a duty of care is not something software may invent for itself.

• “Probing you can configure.” — coming soon

  The agent already follows up on a thin answer. Coming soon: a per-question follow-up policy you set — never, only on a short answer, or always, with a maximum number of follow-ups — so depth becomes a setting on the survey instead of a hope about the conversation.

• Attention checks and mid-call consent rechecks — coming soon

  Two more quality tools are on the same work: a polite check question when the signals suggest someone has stopped reading, and a re-confirmation before a sensitive section begins. Until they arrive, this page will not describe them as things your study can use.

If a distress disclosure is a live risk in your study — a hospital, a college, an exit interview — say so before fieldwork starts, and we will scope the human escalation with you rather than leave it to the agent.

Each behaviour on this page is a named action the agent invokes, with a schema, a timeout and a fallback — not an instruction in a prompt that a longer conversation can wander away from. That is the whole difference between “the agent should tell people who sees their answers” and “the agent reads the disclosure this campaign actually committed to, and can be shown to have done it”.