An AI co-pilot for every automotive sales agent

My role: Account AI-CTO; architected the solution, which the team built.

Executive Context#

A major Japanese automaker’s dealer partners could not reliably match the type of lead to the right salesperson. They also lacked visibility into how leads moved through the sales process and where customer service could improve.

The pilot covered 10 cities and approximately 50 franchises.

The Actual Problem#

The stated problem was lead matching.

The real problem was context lost before it could be used.

The process relied on end-of-day summaries of salesperson inputs, written after the interaction was over. Context that should shape the match and the next conversation could be lost before anyone used it.

Diagnostic Approach#

• AI Necessity: Matching depended on the buyer’s context, interactions, and the behaviour of locally available salespeople. Sales teams needed guidance during the interaction.
• Data Sufficiency: First-touch buyer inputs, interaction history, behaviour, and the chosen car model supplied context. Personal agents captured salesperson inputs during the work instead of relying on end-of-day summaries.
• Methodological Fit: Lead matching and LLM assistance served different decisions. The system matched a lead to an available salesperson and recommended talking points. The salesperson decided how to proceed.
• Infrastructure Readiness: The team delivered Docker images through Azure Container Registry. An API route connected the application to the LLM and database in Azure. A QA agent tested each image against edge cases and deployed it.

Strategic Intervention#

We built the system around the salesperson’s decision, not around a replacement for it.

1. Match the Lead to an Available Salesperson#

We connected first-touch information about the buyer to salesperson behaviour, then recommended the best-fit agent among those available locally.

2. Capture Context During the Work#

Each salesperson had a personal LLM agent. It captured inputs as the interaction happened, preserving context that could otherwise be lost in an end-of-day summary.

3. Recommend the Next Conversation#

Using interactions, behaviour, and the chosen vehicle, the system recommended what the salesperson should discuss next. The salesperson retained the decision.

Outcome#

Metric	Before	After	Context
Client acceptance period	3 months	15 days	The pilot worked well enough for the client to shorten acceptance
Pilot scope	N/A	10 cities, ~50 franchises	Dealer-partner pilot
Context capture	End-of-day summary	Live, per salesperson	Personal LLM agent captured inputs during the interaction

That is a pilot acceptance result. I left after the pilot, so this case does not report long-term conversion or revenue outcomes. The existing stage-wise automotive conversion case is a separate engagement.

Strategic Takeaway#

The system supported a decision while the salesperson could still act. Live context capture, matching, and conversation guidance worked together, with the salesperson responsible for the final action.