The problem
The client is a group of seven car dealerships. Their service departments generate a large share of profit, and most service appointments start with a phone call. Call tracking showed that a large share of calls went unanswered during the morning rush, every call after closing went to voicemail, and very few voicemails turned into appointments. The group wanted to capture that demand without hiring more service advisors just to answer phones.
What we actually did
A real-time voice pipeline
The agent runs on a streaming pipeline: telephony audio in, streaming speech recognition, a language model that decides the next step, and streaming text-to-speech out. We tuned for latency at every stage because callers hang up on pauses: partial transcripts start processing before the caller finishes, and the agent starts speaking as soon as the first words of a response are generated. Callers can interrupt the agent naturally, and it stops talking when they do.
Connected to the systems that matter
The agent looks up the caller by phone number in the dealer management system, confirms the vehicle on file, checks real appointment availability for the requested service, and books directly into the scheduler. It knows each store's hours, loaner and shuttle policies, and standard service durations. It can also check open recalls for a vehicle by VIN and offer to schedule the recall work in the same appointment.
Clean handoffs
Anything outside the booking workflow, such as pricing disputes, warranty questions, or an upset customer, transfers to a service advisor during business hours with a short summary on the advisor's screen, so the customer doesn't have to repeat themselves. After hours, the agent books a callback slot instead of taking a voicemail.
Challenges and tradeoffs
- Vehicle names and VINs over the phone. Speech recognition struggles with model names and alphanumeric strings. We added the dealer's vehicle catalog as recognition hints and confirm VINs back to the caller in short groups.
- Noisy calls. Many customers call from their car. We tuned voice activity detection and added a graceful "I didn't catch that" path that switches to a text message link after two failed attempts.
- Disclosure. The agent identifies itself as an automated assistant at the start of every call, and callers can ask for a person at any time.
Results
Across the group, booked service appointments rose 38% over the same quarter the previous year, with the largest gains after hours and during the morning rush. Every call is now answered, 61% are completed by the agent without a transfer, and recall work booked through the agent added a measurable stream of warranty-reimbursed labor. Service advisors report spending less time on the phone and more with customers at the counter.
What we'd do differently
We would launch the SMS fallback on day one. Noisy calls from cars were more common than our test calls suggested, and the text-link path recovered a large share of them once it existed.