Voice AI
128 CRM updates from 46 calls. One word put the quarter over target.
Sotto, an agent that sits in on every sales call and keeps the CRM current. Every field it writes comes with the clip it heard it in. It writes the small things itself and holds the one update a person has to hear.
- Scope
- Flow, product, how the AI behaves
- Surfaces
- Live call, the week on the web, the CRO's phone
- Hard part
- Knowing which words a person must hear
- Deliverable
- Coded product on Prodular UI, recorded flows
00Overview
A forecast that goes to the board on Monday at nine.
Copperline sells dock scheduling software to warehouses. Six reps, a week of calls, and on Friday afternoon a CRM that has to be right before the CRO shows it to the board.
Sales calls Sotto sat in on, in one week
CRM updates it proposed from what was said
Written on its own, each with its clip
Held for a person to listen to
01One word
Procurement has. Procurement hasn't.
At 31:12 on the Marlow call Ana and Daniel talk at the same time. The transcript caught “has”. Written down, it moves $412,000 into Commit and the quarter past its $2.6M target. Sotto is 38% sure of that word, so it holds it.
Separated, the one recording becomes a lane per voice
As recorded: 2.4 seconds, two voices at once
What the quarter would say if the word were right
02The week
From a live call on Tuesday to the board on Monday.
Every number matches across screens: 46 calls, 128 updates, 121 written by Sotto, 6 confirmed by reps, 1 held.
03Three people
Priya talks. Leah closes the week. Martin asks.
The rep stays in the conversation while Sotto writes beside it. The CRO doesn't open the CRM: he asks out loud, and every sentence of the answer plays its source.
Priya, on the call: the fields fill in while she talks
Martin, CRO: asks before the board call, then sends
04The system
Prodular UI underneath. A voice layer on top.
The shell, the tables, the messages and the voice input come from Prodular UI, our design system. What makes Sotto feel like listening was built for it, on the same tokens.
From Prodular UI
- App shell
- Message
- Bubble
- Voice input
- Table
- Stats
- Property list
- Badge
- Search input
- Avatar
- Switch
- Tooltip
Built for Sotto
- OrbSotto's presence: listens, thinks and answers
- Call waveform46 minutes, coloured by who was talking
- Clip playerOne track that splits into a lane per voice
- Clip chipThe seconds a field came from, one tap away
- Week waveformEvery call of the week, every update above it
- Live captionsThe line being said, word by word
05The impact
The first month with Sotto.
Copperline's six reps in the quarter before Sotto and in the month with it. Nobody types up a call, and nothing reaches the forecast without a clip.
Friday CRM updates, per rep
Commit deals with no quote from the buyer
Forecast lines the CRO can trace to a call
Updates Sotto wrote without asking: next steps, contacts, dates, budgets said out loud
Stage changes a rep confirmed with one tap
Update held for a person: the word that would have moved $412,000
From the product’s scenario: Copperline, Q3 2026 before Sotto and the first month with it.
06Decisions
Three calls made before a screen was drawn.
Each one has a price, and the price is written down.
- 01
Every field carries the clip it came from
Budget, term, next step: each one plays the seconds where somebody said it. The cost: if nobody said it out loud, Sotto doesn't write it, and the rep types it in.
- 02
Sotto writes the small things and holds what moves the forecast
Next steps, contacts and dates go straight in. A deal moving into Commit on audio Sotto is less than 80% sure of waits for a person. The cost: a rep still listens to a clip on a Friday afternoon.
- 03
The rep who was on the call decides
Not RevOps, not the CRO. Priya was there, so the clip goes to her, with the voices pulled apart. The cost: the forecast waits for the one person who heard it live.
07What we test next
Three assumptions that need evidence.
With account executives, RevOps leads and CROs, on real calls with names removed, with tasks rather than opinions.
- Assumption 1
The hold
Will a rep trust 121 updates written without asking, if the one that matters waits for them?
- Assumption 2
The voices
Does pulling the voices apart settle a word faster than reading the transcript?
- Assumption 3
The answer
Will a CRO ask out loud instead of opening the CRM, if every sentence plays its source?
What does your AI write down that nobody checks?
We design the moment a person hears, reads or approves what the model produced, and hand it back designed in five days.