What happens on a call

The call is an actual phone call

There is no app, no browser tab, no headset. The rep dials from their own phone, over the ordinary phone network, and the audio arrives as a live stream. That constraint drives every decision after it: a phone call has no loading spinner, no “thinking…” indicator, and no second chances. Whatever the system does, it has to do while someone is holding a phone to their ear. The same line runs both directions: the rep can place the call to a prospect, or answer a customer who calls in.

Nothing is transcribed while you talk

The usual way to build a voice AI is a relay: transcribe the speech to text, send the text to a language model, read the reply back with a synthetic voice. Three systems, three delays, and the pauses stack up into that distinctive walkie-talkie rhythm.

This does not do that. Audio goes to a speech-to-speech model and speech comes back. Nothing becomes text in between. Avoiding the transcription-and-synthesis relay reduces delay and preserves details text can lose. Tone survives. Hesitation survives. Talking faster because you are nervous survives, and the prospect can react to it, because it is hearing you rather than reading a transcript of you.

It is in character before it says hello

The prospect’s identity travels with the call as it connects, and the system waits for that to arrive before anyone speaks. It sounds like a small detail. It is not: the alternative is a neutral assistant saying “Hello, how can I help you?” and then becoming a character, which tells every rep in the first two seconds that they are talking to software. The first word is already in character. When the persona is a customer calling in, that first word is the rep’s: the prospect waits for the rep to answer, then responds.

Interruption is the default, both directions

Talk over it and it stops, because that is what happens on a phone. The harder prospects also cut the rep off when they lose patience. Neither behaviour is a feature that was added; both are consequences of running a live audio stream rather than taking turns.

The prospect is a document, not code

Each prospect is a description written in plain English — the business, the mood, what earns their attention, what ends the call. Editing that description changes who answers the phone. No engineering, no retraining, no redeploy. A company running four prospects from “will hear you out” to “is not going to buy” is running one system with four documents.

Where the standards come from

A company that selects a listed industry starts with a ready-made pack of callers and objections, alongside practice stages and language rules. That setup is applied directly, without an AI drafting step, and the owner can edit it. For “Something else,” general call stages and language rules are available immediately. AI drafts callers and objections for your business in the background. Every rule arrives as coaching; the owner decides which become strict.

What the call leaves behind

While the call runs, the system marks objections as they are raised and stamps each with its offset into the conversation, so the record knows not just that an objection came up but when. At hangup the transcript, the objections, and the timing are written together, and the recording follows shortly after — captured with the two voices on separate channels, so rep and prospect can be told apart later without guesswork.

Feedback on what was said and how it sounded

The prospect never breaks character to coach. After the call, grading reviews what the rep said and an audio analysis measures how it sounded. These are two parts of the feedback, not a count of grading passes.

The first asks what was said: did the rep answer the objections the way this company answers them, hit the pitch stages in order, respect the language rules. That is judged against the company’s own written standards. By default, the content graders each run three times and combine their findings to keep feedback consistent.

The second asks how it sounded, and needs the audio. A separate transcription pass produces a verbatim record — filler words and all, which normal transcription politely deletes — with a timestamp on every word. From that comes filler words per minute, the longest hesitation, speaking pace, and how many seconds it took to answer each objection. Audio analysis is the reason the recording matters: none of it can be recovered from text alone.

Not every rule carries the same weight, and you set which is which. A rule you mark strict is a compliance line — break it and the call’s score drops. A rule you mark coaching only advises: the rep sees the note, the number does not move. A required disclosure the rep must say and a phrase you would merely prefer are both just rules you write; the severity you give each is what separates a hard line from a nudge.

Graded against a version, not a moving target

Standards change. When a company rewrites an objection answer, the old version is not overwritten — it is superseded, and the new one lands as the next version. Every score records the exact version it was graded against. A call from March keeps the score it earned under March’s standards, and re-grading is a matter of choosing which version to grade against.

Training changes have a signed history

A change to standards, a practice persona, or company information takes a new training revision and writes a signed event in the same database transaction. The event records the actor, action, time, and revision. Each event includes the previous event's digest, linking the sequence so alterations to recorded events can be detected through signature and chain verification.

This captures content changes from the September 18, 2026 rollout onward. It does not make all application data immutable or reconstruct earlier events. Owners and training managers can review this history in their account, and the owner can record or withdraw an approval of the current content. Practice calls and their grades leave signed records of their own, from September 20, 2026 onward. Owners can download this history with the rest of their records and check it against our published keys.

What this means for compliance support →

What it does not do

English only. No CRM, no calendar, no pipeline. It does not watch a rep’s real calls or listen to real customers. It is a prospect on the phone and a grader afterwards — nothing else, and everything that implies.