Ambient agent · B2B sales

Real-time sales copilot: AI that assists the salesperson during the video call

Our newest system: a copilot that listens to the salesperson’s video call and shows them — only them — the fact, the number or the right question at the exact moment. It does not dictate scripts: it whispers glanceable cues, fed by the company’s own knowledge.

B2B Sales · Ambient agentIn pilot with real users

Problem

Technical salespeople need specs, prices and account history while they talk — and searching mid-call breaks the conversation.

Result

A copilot that listens to the call and whispers the right fact or question on screen, visible only to the salesperson.

The context

In a technical sale, the salesperson converses while juggling memory: specifications, prices, account history, the objection this client raised three months ago. Stopping to search breaks the rhythm of the conversation; not searching means answering halfway and ending with "I’ll confirm and get back to you."

The knowledge that wins those conversations exists — in product sheets, in battle cards against competitors, in each account’s history — but it lives scattered across documents and, above all, in the heads of the two or three experts who cannot join every call.

How it works

The copilot captures audio on two channels — the microphone is the salesperson, the system audio is the other party — and transcribes in streaming while the conversation flows. On top of that transcript, a two-tier judge decides at every moment one of three things: search, suggest, or stay quiet. When it acts, the result appears as a glanceable card — a verb and a few words — on an overlay the other party never sees, even during screen sharing.

The difference from a generic assistant is where the answers come from: before touching the web, the copilot consults three drawers of private knowledge — the live facts of the domain, the product sheets and battle cards, and the account file with its history. The web only complements. A copilot that answers with what anyone can google is not a competitive advantage.

Engineering decisions

Decision 01

Intentions, not scripts

The copilot suggests "→ ask about last season’s losses," never the verbatim phrase to read. A salesperson reading sounds like a robot; a salesperson reminded of the right point sounds like an expert. The rule came from testing both: scripts show.

Decision 02

Two model tiers, conversation-grade latency

A fast model acts as gatekeeper — is it worth intervening right now? — and only when the answer is yes does a more capable model write the cue. The system operates at the pace of human conversation, not of a chatbot that thinks for five seconds.

Decision 03

Staying quiet is a feature, not a failure

The worst copilot is the one that interrupts constantly. Success is measured in trigger precision — the right card, on time, without spam — and the salesperson trains the system with one click: useful, give me another, or remove. That feedback tunes when to speak and when not to.

Decision 04

Invisible on the shared screen

The overlay is excluded from screen capture at the operating system level: the salesperson can share their full screen and the copilot remains theirs alone. Without that, the tool would die in its first client demo.

What changes for the business

Every salesperson joins the call with the memory of the whole company: the specs, the account history and the best salesperson’s arguments, available the second they are needed.

  • Cues arrive in ~3 seconds when the fact lives in the private knowledge base — conversation speed, not chatbot speed.
  • The salesperson answers in the moment — no more "I’ll confirm later."
  • The expert’s knowledge scales to the entire sales team.
  • Sales onboarding shortens: the copilot rides along from the first call.
  • Minutes and commitments are ready on hang-up, audio included.

Where else it applies

Technical B2B sales

Products with complex specs, regulations and comparisons.

Brokers & advisors

Insurance and finance: precise data with the client on the line.

Customer support

Support that answers with the knowledge base in its ear.

Recruiting

Interviews with the candidate’s file and the right questions in sight.

FAQ

Frequently asked questions about this case

Does the client notice the salesperson is using a copilot?

No. The overlay is invisible in screen sharing — excluded from capture at the operating system level — and the cues are so brief they are read at a glance, without breaking eye contact or the rhythm of the conversation.

Does this record the calls?

The copilot transcribes to assist live and to generate the minutes at the end. What is kept and for how long is defined by each company’s policy, including whatever consent its legal framework requires.

Where do the answers come from?

First from your own knowledge — product sheets, battle cards, account history — loaded into a private vector base. The web only complements. That is the difference from a generic assistant: it answers with what only your company knows.

Does it work in Spanish?

Yes — it was born in Spanish, industry vocabulary included. The transcription engine and the models also handle English and mixed conversations.

Have a similar challenge?

Tell us your problem. We'll tell you how to solve it — and how long it takes.