Getting a call in
The cheapest transcript is the one somebody else already made.
Most meeting platforms write a transcript automatically and then throw it away into a folder nobody opens. Paying to transcribe the audio again, to get the same words, is the sort of cost that only survives because nobody looked.
So one of the four doors here costs nothing at all.
The doors
A file, a transcript, a dialler, or a bot.
Which route a call takes is decided by what you hand over rather than by a setting: audio goes to transcription, a transcript file is read directly.
Plain text is accepted too, and is deliberately honest about being worse — it is split into paragraphs with invented timings and one speaker throughout. Extraction still runs, quality drops, and the product says so rather than pretending the result is the same.
- Subtitle files keep their speakers —
The voice tags a meeting platform writes are read, so who said what survives the import.
- Two meeting platforms' own formats are handled natively —
Rather than being coerced through a generic parser that loses the structure.
- Consent is checked before anything is stored —
Not after the upload, not as a banner. The refusal happens before a single byte reaches disk, and the person, the address and the time are recorded with the file.
Dialling
Two ways to make the call, and one of them never touches your microphone.
A phone inside the browser, or a bridge: the system rings your own mobile, you answer, and it connects you to the prospect. The second is what most field teams actually want — the call quality of a phone, with the recording and the notes of a softphone.
Four providers are supported, configured with your own credentials so the relationship and the rates stay yours. Saving a new key clears the verified mark, so a rotated credential cannot inherit a stale tick, and sending an empty key means keep the one you have — a secret never travels back to the browser.
- Matching the call to a person tolerates how numbers are stored —
The exact number first, then the last ten digits — because numbers get saved without a country code and a strict match would lose half of them.
- Three of the four providers do not sign their callbacks —
So the platform mints a secret per workspace and carries it in the address it hands them. The fourth does sign, and the secret is kept as a fallback for workspaces that turn signing off.
- Every refusal is audible —
An unrecognised number, a disabled dialler, a failed signature check — each answers with a spoken message rather than dead air.
›Does the browser phone work with every provider?
One of the four today. Bridged calling and recording capture work with all of them — worth knowing before choosing a provider on the strength of the in-browser phone.
›What happens to a call from a number we do not recognise?
It is discarded rather than queued. There is no triage inbox for unattributed calls, so an inbound from an unknown number leaves no record here.
›Are inbound calls captured the same way?
They are recorded and processed, but they do not reach the timeline, because who made the call is what attributes the activity and that only travels on the outbound path.
The bot
It joins as a participant, with a name, in the list.
Not a silent tap on the line. It appears in the meeting the way a person does, so consent is a thing everyone can see rather than a clause somebody signed.
It records, uploads, and hands the file over through the same route a person would use. Because it watched which participant was speaking, the transcript comes back with real names rather than speaker one.
- Opt in per booking link —
And only offered where the meeting is on a platform it can actually join.
- It gives up after three failures —
An administrator can still retry by hand. Automation that keeps trying to join is worse than automation that stops.
- A booking with no meeting link is skipped silently —
There is nothing to join, and a loud failure would train people to ignore the failures that matter.
How the names get onto the transcript
Taken from who was on screen, not guessed from the audio.