AirDrv handles 80% of inbound calls for a dental practice. The remaining 20%, escalations, ambiguous insurance questions, multi-step booking edits, is where most voice AI demos quietly fall apart. Here's what we learned shipping voice into a real receptionist workflow.
AirDrv is the voice AI product in our portfolio. It answers calls 24/7 for dental practices, books appointments, triages emergencies, verifies basic insurance, and routes anything ambiguous to a human. The deceptively simple framing hides a hard real-time engineering problem.
#Latency is the product
A receptionist who pauses for two seconds before every reply sounds broken to the caller, no matter how accurate the words are. Voice AI lives or dies on perceived latency, the time between the caller finishing a sentence and the agent starting its reply.
We target sub-700ms end-to-end. To get there, we stream partial transcripts into the LLM and stream the LLM response into the TTS, so the first audible word starts before the full reply has been generated. Anything else feels like a robocall.
#The 20% that breaks demos
Demo videos always use the happy path. "Book me a cleaning next Tuesday at 3" is easy. "My filling came out yesterday, I don't have insurance through my old employer anymore but my partner's plan should kick in next week, what do I do?" is the actual job.
- Ambiguous time references ('next Tuesday' vs 'the Tuesday after next')
- Multi-issue calls (booking + insurance question + payment plan)
- Emotional callers (post-procedure pain, billing disputes)
- Background noise (callers in cars, near children)
- Caller corrections mid-sentence ('actually, make that Thursday')
#Escalation is a feature
A voice AI that pretends it understood when it didn't will burn the practice's reputation faster than any missed call. We treat escalation as a first-class product surface, the agent has explicit confidence thresholds, and below them it offers to take a message or route to the practice owner's mobile.
#Our voice stack today
- Real-time ASR via a streaming provider (we benchmark quarterly)
- An LLM router that picks between Claude, GPT, and a fine-tuned small model based on intent
- A streaming TTS layer with voice cloning for practice-specific branding
- A dialog state machine that owns booking, insurance, and emergency flows separately
- An observability layer that records every call for evals and disputes
A receptionist isn't paid to transcribe. They're paid to make the patient feel heard. The benchmark for voice AI is not WER, it's whether the patient called back.
If you're curious how voice fits into the broader Nexobe lineup, see the company page. For AirDrv pilot details, the product site is the entry point.