🎙️ Voice Agent Conversation Flow Architect

Voice Agent Flow: Inbound Dental Practice Reception

Platform Recommendation

Vapi. Reasons specific to your use case:

• Best-in-class latency for the 800ms budget you'll need (Vapi typically lands 600-900ms).
• First-class Twilio handoff (matches your existing Twilio setup).
• Native HIPAA-compliant deployment (Retell is HIPAA-capable but requires more config).
• Their function-calling for Dentrix API integration is more robust than Bland's at this maturity.

Retell would also work well; it's a close second. Bland is better suited for outbound. ElevenLabs Conversational AI has the best voices but weaker function-calling for your Dentrix integration.

The Conversation Flow

[GREET] "Thanks for calling Dr. [name]'s office. This is [voice agent name]. I can help you book, reschedule, or cancel an appointment, or check on emergency availability. What can I help with today?"
  ↓
[INTENT] Listen for intent. Branches:
  - 'book' or 'new appointment' → BOOK
  - 'reschedule' or 'change' → RESCHEDULE
  - 'cancel' → CANCEL
  - 'emergency' or 'pain' → EMERGENCY (immediate escalation)
  - 'human', 'agent', 'receptionist' → HANDOFF
  - 'insurance' → HANDOFF (out of scope)
  - unclear → CLARIFY (max 2 attempts then HANDOFF)
  ↓

[BOOK]
  → ask first/last name (read-back confirm)
  → ask DOB (read-back confirm — HIPAA flag: stored briefly, transmitted only to Dentrix)
  → ask reason for visit (cleaning / consultation / pain / specific procedure)
  → query Dentrix for existing patient OR new
  → if existing: confirm phone on file or update
  → if new: capture phone (read-back confirm)
  → query available slots based on reason (cleanings vs ops use different blocks)
  → offer 2 slots: 'I have Tuesday at 10am or Thursday at 2pm. Which works?'
  → if neither: 'I have several options later this week. Could you tell me a day that works?'
  → confirm booking back: 'So that's [name], on [day], at [time], for [reason]. Right?'
  → on yes: book in Dentrix, deliver confirmation: 'You're booked. We'll text a reminder. Anything else?'
  → on no: re-prompt for the disagreed field

[RESCHEDULE]
  → ask name + DOB to look up
  → confirm existing appointment: 'I see you on [day] at [time]. Is that the one?'
  → ask preferred new time
  → check availability, offer 2 options
  → confirm + execute reschedule in Dentrix
  → confirmation message + text reminder

[CANCEL]
  → ask name + DOB to look up
  → confirm existing appointment
  → ask cancellation reason (optional, for our records)
  → execute cancel in Dentrix
  → 'Got it, [name]. Your appointment on [day] is cancelled. Would you like to reschedule now or call back?'

[EMERGENCY]
  → 'I'm transferring you to a real person right now. Please stay on the line.'
  → immediate Twilio handoff to receptionist
  → if outside hours: 'Our office is closed but our emergency line is [number]. For dental pain, please call there now.'

[HANDOFF]
  → 'Let me transfer you to one of our team. One moment.'
  → Twilio handoff with context tag (current state, what's been collected)
  → if outside hours: 'We're closed right now. I'll have someone call you back. Can I take your name and number?' (capture, read-back, log to ticket queue, end call)

[CLARIFY]
  → 'I didn't quite catch that. Are you calling to book, reschedule, cancel, or about an emergency?'
  → 2 attempts max → HANDOFF

Turn-Taking & Interruption Handling

Barge-in policy: Always allow user interruption EXCEPT during the read-back confirmation steps (where interrupting changes which field gets confirmed). On other turns, immediately yield to user when they speak.

Interruption acknowledgment: When user interrupts mid-sentence, agent stops, processes the interruption, and responds. Do NOT add 'sorry, I was speaking' — it adds latency and feels stilted.

Filler tokens during processing: When function-calling Dentrix takes >800ms, Vapi can play a filler ('Let me look that up, one second'). Configure this for any function call expected to exceed 800ms.

Yield priority: If user starts speaking while agent is mid-sentence, yield within 100ms. This is the platform default for Vapi; verify it's enabled.

Latency Budget Allocation

Total target: <900ms perceived response.

• ASR (Vapi default Deepgram): 200-300ms
• LLM call (Claude Sonnet via Vapi, with cached system prompt): 300-500ms
• TTS (ElevenLabs streaming through Vapi): 150-250ms (first audio chunk)
• Network overhead: 50-100ms

Critical optimizations:

• Use prompt caching on the system prompt + flow definition (reduces every turn's LLM time by 200-400ms)
• Use streaming TTS (don't wait for full sentence to start playback)
• Parallel function calls when possible (e.g., 'lookup patient' AND 'check availability' fire simultaneously when both are determinable)
• For Dentrix API calls expected to take >500ms, play filler token while waiting

The 6 Failure Mode Defenses

1. No-speech-detected (caller says nothing for >5s):

• Trigger: Vapi VAD reports no audio above threshold for 5s
• Recovery: 'Are you still there? I can help with booking, rescheduling, or cancelling.'
• 2nd timeout (after 5 more seconds): 'I'll have to disconnect. Please call back when you're ready.'
• End call gracefully

2. Partial-transcript (ASR confidence below threshold):

• Trigger: ASR confidence <0.6 on the user's response
• Recovery: 'I'm sorry, the line cut out. Could you repeat what you just said?'
• 2nd low-confidence in same state: HANDOFF

3. Accent-misrecognition (caller has strong accent, ASR producing nonsense):

• Trigger: 2 consecutive turns where the LLM cannot map the transcript to a valid intent
• Recovery: 'I'm having a little trouble understanding. Let me transfer you to one of our team.' → HANDOFF immediately. Do not loop trying.

4. Talk-over-user (agent kept talking while user tried to interrupt):

• Trigger: detection of significant overlap in audio streams (Vapi can flag this)
• Recovery: 'Sorry — go ahead, what were you saying?' — agent yields fully.

5. Dead-air (silence on both sides for >4s mid-conversation):

• Trigger: VAD reports no audio from caller for 4s, agent has nothing queued
• Recovery: 'Just to make sure I have you — what's a good day for the appointment?' (re-anchor on current state's question)
• 2nd dead-air: 'Are you still with me?'
• 3rd: graceful end-call with offer to call back

6. Escalation-trigger (caller says key phrases):

• Trigger: 'human', 'agent', 'receptionist', 'transfer me', 'real person', 'manager', 'I want to speak to', 'this is ridiculous', 'I'm frustrated'
• Also: keyword 'emergency' or 'pain' immediately escalates regardless of state.
• Action: 'Of course, let me get you to a real person.' → HANDOFF

Number Sequence Handling

Phone numbers:

• Capture: 'What's the best phone number for you?'
• Read-back: 'So that's [break each digit pair: 5-5-5, 1-2-3, 4-5-6-7]. Did I get that right?'
• On 'no': capture again, do NOT assume which digits were wrong. Re-capture full number.

DOB:

• Capture: 'And your date of birth?'
• Parse to ISO date for Dentrix lookup, but read back in spoken form: 'June 14th, 1985 — is that right?'

Appointment times:

• Always read back time + day + reason: 'Tuesday April 28th at 10am for a cleaning. Does that work?'
• On 'wait, what time?': re-state slowly.

Compliance & Disclosures

HIPAA:

• Recording disclosure (assuming you record): in greeting, after 'thanks for calling': 'This call may be recorded for quality.' (one short sentence — keep greeting brief)
• Data minimization: collect only name, DOB, reason for visit, phone. Do NOT capture insurance details, SSN, payment.
• If caller volunteers insurance: 'I'd love to help with that, but for insurance questions our team is best. Let me transfer you.' → HANDOFF

Recording consent:

• One-party-consent states (most US): the disclosure is sufficient.
• Two-party-consent states (CA, FL, IL, others): explicit verbal consent required. Add: 'I record calls to improve service. Is that okay?' Wait for affirmative.

TCPA (if you ever do callbacks):

• Outbound callbacks must respect time-of-day restrictions and prior consent.

Escalation Paths

During business hours (Mon-Fri 8-5):

• Twilio warm-transfer to main reception line
• Pass context as Twilio call attribute: state at handoff, fields collected, reason for escalation
• Reception agent sees a brief summary popup before picking up

Outside business hours:

• Capture name + callback number
• Log to ticket queue (your CRM)
• Promise callback by '9am tomorrow' (or 'Monday morning' on weekends)
• Read back: 'I have [name], number [digits]. We'll call by [time]. Anything else?'

Escalation if no human available (system error):

• Promise callback within 1 hour
• Page on-call manager via PagerDuty (this should be rare)

Implementation Skeleton (Vapi)

assistant:
  name: 'Dental Front Desk'
  voice: { provider: '11labs', voice_id: 'rachel' }  # warm, mid-tone, US accent
  model: { provider: 'anthropic', model: 'claude-sonnet-4', temperature: 0.4 }
  systemPrompt: <full state-machine flow + brand voice>
  firstMessage: '<GREET line>'
  responseDelaySeconds: 0.4
  endCallPhrases: ['goodbye', 'bye']
  silenceTimeoutSeconds: 5
  numWordsToInterruptAssistant: 2
  functions:
    - lookup_patient (params: name, dob)
    - check_availability (params: reason, date_range)
    - book_appointment (params: patient_id, slot_id, reason)
    - reschedule_appointment (params: appointment_id, new_slot_id)
    - cancel_appointment (params: appointment_id, reason)
    - transfer_to_human (params: context, reason)
    - log_callback_request (params: name, phone, callback_window)
  twilioPhoneNumber: <your number>
  recordingEnabled: true

Test Plan

Required test calls (manual):

1. Happy path: book new appointment, no interruptions

2. Existing patient reschedule

3. Cancel

4. Emergency: 'I'm having severe pain' → verify immediate escalation

5. 'I want to speak to a human' → verify immediate handoff

6. Insurance question → verify out-of-scope handoff

7. Caller goes silent mid-call (test dead-air)

8. Caller interrupts greeting (test barge-in)

9. Caller speaks with strong non-US accent (test accent failure path)

10. Caller from a noisy environment (test ASR-confidence fallback)

11. Outside business hours scenario

12. Caller says 'never mind' mid-flow (test graceful end)

Required automated tests:

• 100 simulated calls with varied scripts
• Latency p95 must stay under 1100ms
• Handoff success rate >95%
• Number-sequence read-back error rate <2%

Production rollout:

• Week 1: 20% of inbound calls (low-traffic hours), monitor
• Week 2: 50% of inbound calls
• Week 3: 100% with human override always available

Voice Selection & Personality

Voice: Warm, mid-tone, US accent (ElevenLabs 'Rachel' or equivalent). Female-presenting voice is conventional for dental front desks and tests well. Pace: slightly slower than default (-10%). Formality: professional but friendly.

Why: Dental anxiety is real. The voice shouldn't sound clinical. Warm + slightly slower paces matches the human receptionist's natural style and reduces patient anxiety on the call.

Avoid: synthetic-sounding voices, uptalk, excessive filler ('um', 'uh' — TTS-generated fillers feel uncanny).

Key Takeaways

• Latency budget is 800-900ms total. Vapi + Claude Sonnet + ElevenLabs streaming + Deepgram ASR can hit this. Use prompt caching aggressively.
• Build all 6 failure modes from day one. Voice agents don't have the luxury of 'we'll add error handling in v2' — failures during a real call are immediately visible.
• Number-sequence read-back is non-negotiable. Phone numbers and DOB will be wrong on first capture some percentage of the time.
• Hard-fail to human at 3 misunderstandings. Looping forever loses callers; transferring saves them.
• HIPAA data minimization. Collect name, DOB, reason, phone — nothing more. Insurance handoff to humans.
• Test with non-native speakers AND on bad connections. Both conditions reveal flaws clean testing won't catch.