A real-time assistant for salespeople on live calls. It transcribes as they talk, recognizes what's being discussed, pulls the matching inventory and policy, and puts the answer on screen at the moment it's needed. The agent stops searching and goes back to talking.
What an agent held open before, and what replaced it.
Nothing here waits for the conversation to end — except the last step.
Audio streams in as the call happens.
Speech becomes live, readable text.
Vehicles, intents and objections get named.
Matching stock and policy are fetched.
The one relevant answer appears on cue.
Hang-up writes the notes and follow-ups.
Step five is the whole product. Retrieval is commodity; choosing the single thing worth showing at second forty-three of a live conversation is not.
The arrangement arrived at after two rejected alternatives.
Fixed sides, live centre — the reason the screen stays legible at conversation speed.
Two explorations were built and killed. Four rules survived them.
| What was tried | Why it broke | The rule it produced |
|---|---|---|
| A chatbot — one clean answer, nothing else | The feed moved at talking speed with nothing to anchor to | Fast is not usable. Give the eye a fixed point. |
| Dropping the customer profile to save room | The agent had to hold the context in their head instead | Never make the human the memory. |
| Vehicle options as rich cards | Ate the screen and fell apart past a handful of trims | Dense and scannable beats pretty and paginated. |
| Trusting the model's output as final | A wrong read mid-call had no escape hatch | If it can be wrong, it must be correctable — live. |
Each surface exists because a different moment demanded it.
The three-panel call screen. Inventory, policy and action items assemble themselves while the agent talks.
Auto-summary at hang-upVoicemails, unhappy customers, same-day appointments. Managers choose which alerts reach which agent.
75+ types · 1-hour callback windowThe same listening, untethered. Open, find or create the customer, record the in-person conversation.
Hands the summary to financeEach organization runs a different playbook, so policies are configurable: trigger, wording, scope, on or off.
Per-organization rulesRecording competitive, autonomous salespeople for the first time.
Managers get review screens framed as helping the agent win, not catching them out.
Framing decides adoptionSolve today's annoyance before asking for tomorrow's data. Value has to land before trust is requested.
Utility precedes buy-inThe right sentence at the right second, not a dashboard of everything that might apply.
One answer beats tenDistilled from a single public product-design case study published in 2026, covering roughly eight months of work on a vertical AI assistant for automotive retail. Product and company names are omitted by request. Figures — call volume, seat count, organization count, alert types, and the workload ranges — are the case study's own; the pipeline, the rules table, and the three principles are my compression of its narrative, not labels it uses.