AI on the desk: summaries, briefs, digests, risk and QA
What the on-server AI does with recordings and CRM data, where each result appears, who sees it, and what it never does.
The rule
Recordings, transcripts and client data never leave your server. A worker on an isolated network sends calls over 45 seconds to a speech-to-text model (Whisper large-v3-turbo) on your own GPU, and a 27B-parameter open-weight language model (Qwen3.6) on the same box writes the rest, with a CPU model on the main server as automatic fallback. Nothing is sent to any outside AI provider. Every feature below is one switch in Settings → AI; the whole thing has a master switch there too.
Per call
| What | Where it appears | Who |
|---|---|---|
| Summary, key points, suggested outcome, compliance flags | Call review page and the lead/client timeline | agents see their own calls; managers all |
| Voicemail detection on short answered calls | outcome set to VOICEMAIL automatically | everyone |
| Heard on the call: callback time, language, time zone, next step, objections | timeline line '🤖 Heard on the call'; a callback task is created only when the model is confident, the time is in the future and no callback is already open | the call's agent |
| Follow-up email draft | lead/client page → 'AI follow-up draft' → Open in composer. Never sent automatically; the normal send checks still apply | agents with email rights |
| Automatic QA score against your QA criteria | call review page next to the manual score; Quality report shows averages and the lowest calls | managers |
| Card numbers, passwords and one-time codes | masked in the transcript and muted on playback (the stored audio file is unchanged) | everyone who can play the call |
| Language from the first call | the lead's language is set when it was empty and the transcript makes it obvious | everyone |
Your day
Today opens with a brief written for you: an honest review of yesterday — callbacks missed or made late, conversations without a logged outcome, statements compliance flagged, targets hit or missed — then today's plan in priority order (first callback, who to call back first, what to fix, what to log). Sales agents get their leads, retention agents their book and the clients at risk, managers the desk's figures against the day before and the agents who drifted, then their own items. It is written for everyone each morning before the shift, so it is already there when you log in, and refreshed every few hours during the day.
AI summaries can be filtered by date, agent, dialer (Call Center, Relay, Parley, click-to-call), lead or client, logged outcome, what the AI thinks the call sounded like, sentiment, language, flags (any, high, none), call length and QA score, and sorted by newest, oldest, longest, most flags or lowest QA. The search box also matches names and numbers.
Before you dial
The lead page shows a pre-call brief: one paragraph on where the lead came from, what happened on earlier calls, what was promised and what to avoid, plus an opening line. It is refreshed after every call and rebuilt for callbacks due in the next 24 hours. Relay and Parley receive the same brief with the lead. A 'Best 10–12' chip next to the number is the local window this person actually answered in (no language model involved).
Retention
Each client carries an AI risk score (0–30 healthy, 31–69 watch, 70+ at risk) with the reason, refreshed weekly and after every call; the book can filter and sort by it, and the owner is told when a client crosses 70. The client page also has a relationship summary: the whole history in one paragraph with key facts, promises made and things to watch.
Managers
- Report digests: a plain-English note at the top of every report (Desk, Daily, Weekly, Leads, Clients, FTD, Sources, Carrier, Commission, Compliance, Quality, Outcomes, Voom) for the period you picked — what changed, who drifted, three things to do. Written every morning for the period each report opens with, refreshed hourly while the period is live (the last note stays on screen meanwhile); managers get a notification when the daily one is written.
- Coaching cards: one per agent per week on the agent's page — what they do well, where conversations die, objections they lose to.
- Training: the Training page turns the last two weeks of flags, QA scores, objections, short calls, drop-offs, owed outcomes and missed callbacks into ranked signals with the figures behind them, points each at a short module from the catalogue, and writes a one-week plan per agent (exercises, rewrites of the agent's own sentences, role-play lines). Every module also has a 3-minute click-only drill with a score (see Training). Managers assign a module as a task due in a week; passing its drill closes the task.
- Alert explanations: on Reports → Alerts, each alert shows a short 'why' with context.
- Why they said no: on Reports → Sources, objection themes per source and partner with counts and one example each.
- Partner week in plain English: an internal narrative per partner, and a neutral partner-facing text with a Copy button — read it before sending.
- Form check: on the lead page, 'looks genuine / questionable / suspicious' judged from the form alone (name, email, country coherence), never from who the person is. Averaged per partner on the Sources report.
- Quality report: winning and losing phrases of the week and where agents drift from the script. Compliance report: flagged statements clustered into patterns, and internal chat reviewed for the same risk phrases as calls.
- Email templates: a '3 variants' button writes alternative subject lines, preheaders and openings to copy; the template itself is never changed.
Reading a narrative
Every narrative is written from the numbers on the same page and may only repeat them. A sentence that cites a figure not in the data is removed before you see it, and the footer says how many were removed. Treat suggested actions as a starting point, not a verdict.
Nothing is sent to any outside AI provider. Every feature can be switched off on its own in Settings → AI; switching the master off stops the lot.