Voice screens nobody acts on until a person has read them

The question set is authored by you, attached to the requisition stage, and fixed for the duration of the search. Every candidate at that stage hears the same questions in the same order, which is exactly what a first-stage screen should be and rarely is when four recruiters run it by phone across a fortnight.

Answers are captured as structured fields against the questions that produced them, not only as prose. That means an availability answer lands in the availability field and a notice period answer lands in the notice period field, so the record is directly comparable across a shortlist and can be filtered without anyone re-reading twenty transcripts.

A candidate who does not want to speak to an artificial voice can say so, and the call ends. The record shows that the candidate opted for a human conversation, and a task is raised for the recruiter to make that call. Declining is not a disposition, it does not affect a score, and it does not move the candidate anywhere in the pipeline.

This is a substantive design position rather than a courtesy. A screening step that penalises the people who refuse it stops being optional in any meaningful sense, and it collects a signal about a candidate that has nothing to do with whether they can do the job.

When a call completes, the session panel presents the recording, the transcript, the structured answers and a summary side by side. Nothing has advanced. The candidate is at the stage they were at before the call, and the transcript is a document awaiting review.

A recruiter reads it, corrects anything the transcription got wrong, and then makes the stage decision themselves with a coded disposition. The review is recorded against the session, so the record shows both the machine output and the human who acted on it. There is no configuration that lets a call advance or dispose of a candidate on its own.

Recordings, transcripts and derived summaries are three artefacts, and a retention policy that deletes one and keeps the others is the failure most systems ship with. Retention here is configured per tenant, applies to all three together, and can be set shorter for the audio than for the structured answers if that is the balance you want.

When a candidate exercises an erasure request, the erasure workflow removes the audio file, the transcript, the summary and the structured answers as one operation, and records that the request was fulfilled and when. Where a placement has already been made and financial records must legally be retained, the workflow says so rather than quietly keeping the interview material as well.

The system does not infer personality, emotion, confidence or cultural fit from a voice. It does not score an accent, a speech pattern or a hesitation. It does not produce an employability rating. Those capabilities are technically available and we have chosen not to expose them, because they are not measuring what they claim to measure and because they carry obvious discrimination risk.

What is produced is a transcript, structured answers to the questions you wrote, and a summary of what was said. That is the honest boundary of what a voice screen can contribute, and it is already worth having when the alternative is four hundred unreturned calls.

Wave 1 is a narrow, well-instrumented release with a small number of accounts running real volume, because a voice product needs to be tested against genuine phone lines, background noise and people who talk over the prompt rather than against a demo script.

The changes we expect ahead of general availability are broader language coverage, better handling of interruption and clarification, all-party-consent jurisdiction support once the consent capture flow around it is properly built, and a published operation weight for call minutes. Progress is posted on the shipping page rather than announced when it is finished.