Call notes written for you, once we can do it properly

It will join or receive audio from a recruiter's call or interview, transcribe it, and produce two artefacts: a summary written in the shape recruiters actually use, and a set of action items with owners. Both will be filed against the candidate, the requisition and the client contact the conversation belongs to, in the same activity stream as everything else.

The summary format will follow the recruitment conversation rather than a generic meeting template. A candidate screen summary is not the same document as a client intake summary: one needs motivation, notice, expectation and objections, the other needs requirements, budget, process and decision makers. Both will populate structured fields where the answer belongs in a field, so a notice period discussed on a call is searchable afterwards.

A notetaker is not a transcription problem. Transcription is a solved commodity. The hard part is that a notetaker records other people, frequently people who are not your customers and have not agreed to anything, in a conversation that may cross two jurisdictions with different rules about who has to consent.

We are running that problem once, on the voice interview beta, where the calls are outbound, the question set is fixed, the jurisdiction gating is explicit and the volume is controlled. What we learn there about disclosure, retention, deletion and the operational reality of an administrator switch is what a notetaker will be built on. Doing both at once would mean doing both badly.

Four things, and none of them are model quality. Consent capture has to work at the point of the conversation, not as a checkbox somebody ticked at signup, and it has to record who consented, when and how. Jurisdictional recording rules have to be enforced by the product rather than left to the recruiter's memory, including the all-party-consent case, which we currently handle by disabling recording rather than assuming agreement.

Retention has to cover the audio, the transcript, the summary and the extracted fields as one policy, with erasure removing all four together. And meeting-platform permissions have to be real: each platform has its own rules about bots joining calls, what an application may capture and what it must display to participants, and those are integration terms we have to satisfy properly rather than route around.

Notes are typed, and they go into the activity stream on the record rather than into a notebook. That sounds like a downgrade because it is one, but the structure around it does most of the work a notetaker would.

Interview kits give the interviewer the question set in front of them, so notes are taken against the right prompts. Candidate scorecards capture structured feedback that can be compared across a panel instead of a paragraph per interviewer. The activity stream keeps every note with the candidate rather than in a document nobody can find. And the AI interview calls beta already produces a transcript and structured answers for the screening stage it covers, which is the stage most desks would have pointed a notetaker at first.

The same way the voice features were: behind a flag, on a specific tier, enabled per account rather than switched on globally, with a spoken or displayed disclosure before recording, a per-tenant retention setting and an administrator switch that turns it off for everyone.

It will also ship with the same boundary. A notetaker will summarise what was said and extract what belongs in a field. It will not score a candidate on their delivery, infer confidence or sentiment from a voice, or produce a recommendation. Those things are easy to build and hard to defend, and they are not what the feature is for.

Changes are posted on the shipping page as they happen, and the changelog carries the dated releases. The roadmap page shows what sits in which wave and moves items between waves publicly rather than quietly.

If automatic note-taking is a real requirement for your desk, say so during a demo. Wave 2 ordering is influenced by what buyers actually block on, and a feature that three agencies are blocked on moves ahead of one that reads well in a comparison table.