← All work
Agentic AI · Telematics

Agentic Service Co-Pilot: CAN-bus signals → CRM tickets

Drove 0-to-1 strategy for an agentic AI co-pilot that turns live CAN-bus signals into auto-generated CRM tickets, so the dealership can call the customer before the customer calls them. Cycle times −35%.

Role
Founding Product Manager
Company
AI Growth Vector
Timeline
2024 — 2025
Industry
Connected Vehicles
−35%
service cycle time
0.85
anomaly probability threshold (brake)
Closed-loop
with concept-drift monitoring
PM-set
precision/recall trade-offs

Context

Dealership service is reactive by default. A customer calls after the brake squeals start. Modern vehicles already broadcast the signal long before that: CAN-bus tells you, at sub-millisecond cadence, that something is drifting out of normal. The product question is whether you can act on it without annoying the customer or missing a real safety event.

The architecture I scoped

  • Ingestion: CAN-bus signals (brake wear, voltage, brake-squeal frequency) streamed to Azure Databricks.
  • Medallion data layers: Bronze raw, Silver clean, Gold features (squeal frequency, vibration intensity).
  • An anomaly detection model produces a probability score (e.g. 0.85 brake-failure risk).
  • Once the score crosses threshold, an LLM agent takes over and runs a plan.
  • Action layer: check parts inventory, check technician availability, retrieve customer preferences, send personalized SMS, write the CRM ticket.
  • Feedback loop: inspection outcomes feed model retraining; concept-drift monitoring catches when sensor patterns shift.

The PM decisions only I could make

Data scientists built the model. As PM I defined how it should behave:

  • Brake safety → high recall. Don't miss a failure. False alarms are tolerable.
  • Oil-change recommendations → high precision. Only alert when confident. Don't annoy the customer.
  • Drift monitoring shown to the user, not buried in MLOps.
  • Customer-facing message tone and the booking-confirmation copy ("Reply YES"). Tiny UX details that decide whether the loop closes.

Outcome

  • −35% service cycle time on the pilot dealership cohort.
  • End-to-end loop in production: model detects, agent acts, technician verifies, drift monitor retrains.
  • Repeatable template now applied to other vehicle subsystems beyond brakes.

Reflections

My GM OnStar years taught me that connected-vehicle data is only as useful as the human workflow you can wire it into. The CAN-bus signal has been there for two decades. What's new is having an agent layer reliable enough to act on it without a human in the middle of every loop.

The PM call that mattered: treating model precision and recall as a per-subsystem business decision, not a single knob. Brakes and oil changes have very different costs of being wrong. Encoding that asymmetry into the product was what made dealerships trust the agent enough to let it write tickets on their behalf.

Stack

Azure DatabricksMedallion (Bronze/Silver/Gold)Anomaly DetectionLLM Agent (GPT-4o)Salesforce / CRM APIs

References

Read next