Sales calls

One reading of the transcript yields up to 30 signals. Each is one of nine kinds, and each carries a strength: explicit (said outright), implied (the customer only confirmed the rep's paraphrase) or blocking (stated as a condition of buying or staying).

Each signal also gets the existing feedback type (bug, request, praise, complaint, question), sentiment, emotions and language, so the Requests and Bugs views, Reports and Ask work on calls with no new filters to learn.

The transcript is read once and never kept. Not stored, anywhere: the transcript body, raw or redacted; the notetaker's summary and notes; the rep's own words (a rep saying "our export is fast" is not feedback); and anyone's name, on either side of the call. An attendee is identified by email only, so erasure by email covers calls the same way it covers everything else.

What is kept per call: the source id and a link back to the call in the source tool, the title with personal names removed, when it happened, how long it was, the word count, the external attendee's email domain, the rep's email, the folder and the language. What is kept per signal: the customer's quote, redacted before storage, plus a short redacted context window of the sentence before and after it, the offset into the call, the kind and the strength.

Calls are counted as items by length: one item per 500 words of transcript, both speakers included, with a minimum of 2 per call. A 45-minute call runs to about 6,750 spoken words, so it counts as about 14 items. The count is fixed at the door, before the model runs, from the word count in the response.

A call counts even when nothing is found in it, because the reading is the cost. That includes a call the model reads and then judges internal (skip reason internal_call). A call the door's own filters stop before any reading (too short, too long, no customer speech, no external attendee) is recorded with its reason and not counted. Sending the same external_id twice never counts twice; the second send returns the first row unchanged.

Uploaded and backfilled calls are history: they draw from the one-time 1,000-item import allowance first, shared with CSV imports, and then from the plan's monthly volume. If the plan has no room, the call is recorded as plan_blocked and analysed when it does.

Under Sources → Sales calls, choose a file. Three formats are read:

Speaker mapping is the same for every format: me and them are literal; any other label that matches your name, your first name or the local part of your email is you; every other label is the customer. A file with no speaker labels at all is recorded as skipped, reason no_quotable_speech: nothing is guessed.

Uploads are history (see below). A file that cannot be read is refused whole with the reason and what was found, for example the column names of a CSV with no transcript column. Nothing is sliced or partially imported.

POST /api/v1/transcripts takes one call per request and answers 202 before any model runs. It accepts secret keys only, 60 requests per minute per key, bodies up to 2 MB. The full field list, every response and every error are on the Transcripts API page; the shape is:

status is queued when the call will be analysed, skipped with a skip_reason when a filter stopped it (under 800 words, over about three hours, no customer speech, no external attendee), or plan_blocked. Extraction runs asynchronously; poll GET /api/v1/transcripts/{id} for the signal count, and for internal_call when the model found it was not a customer call after all.

Connect Granola once and every call it finishes arrives on its own. This is the only notetaker Sentriment connects to natively; everything else comes through the upload door or the API above.

Which calls are read is up to you, under Filters on the connector row. All four are decided from a call's metadata before any transcript is fetched or anything is counted:

Every skipped call is still recorded with its reason, so the card can tell you "Skipped 14 calls in the last 7 days: 12 internal, 2 under 800 words" rather than leaving you to wonder where a call went. A skipped call is never counted. Change the filters whenever you like and check them against 20 recent meetings again; changes apply to new calls, not to calls already read.

A call is live when it reaches Sentriment within 48 hours of started_at (a missing started_at counts as now). Everything else is history: uploads, backfills and API calls older than 48 hours. History joins themes, counts in totals and scores health, but never fires an alert, never registers as a spike and never reopens a resolved theme. A quarter of old calls sent in one go changes what you know, not what pages you.

Who was on the call decides what a signal can do, not where it came from. If the external attendee's email matches a customer you have identified, it is a customer call: the signals land on that profile, count towards revenue at stake, and a churn-risk signal on a renewal call feeds the at-risk rule exactly as a ticket would. If the email is unknown, the person becomes a prospect: scored and visible in themes today, with the profile identified by email. Three things are still to come: the prospect badge in Users, the exclusion of prospects from health aggregates, and a theme's separate pipeline figure from a deal_value trait passed through identify, shown beside revenue at stake and never added to it.

Recording consent is the customer's responsibility: the party running the notetaker obtains it, under the rules that apply where the call takes place. Granola provides in-meeting notices for this. Sentriment receives a transcript that already exists, reads it once, and keeps only the redacted quotes described above.