Voice receptionists and chatbot
AI that talks to the public has to be safe, cheap and useful. I built a voice receptionist and a website assistant for a counselling clinic, and ClientRelay, my own voice receptionist product.
CallerHi, I would like to book a first appointment for my son.
ReceptionistI can help with that. Would you like to meet in person or by video?
CallerVideo, please. Weekday evenings if you have them.
ReceptionistThursday at 6 pm is open. I will text you the intake form now.
The problem
A clinic’s first contact is often a phone call or a question on its website, sometimes from a person in crisis.
The assistant must answer real questions, never give clinical advice, and send anyone at risk to real help at once.
It also has to cost little enough to run on every page, and the phone agent must hand callers to the clinic’s own workflow.
What it does
- A voice receptionist that routes intake by phone
- A website assistant tested on 110 written safety cases
- A crisis card, rate limits and a kill switch
- Structured call data sent to the clinic’s workflow
- ClientRelay: voice receptionists as a product
My role
I built the assistant and its safety controls, wired the voice agent into the clinic’s tools, and chose the model on measured results.
ClientRelay is my own product: all 236 commits are mine.
| Model | Score | Cost | Crisis and privacy |
|---|---|---|---|
| Claude Haiku | 96.7% | 1× | All passed, 3 runs |
| Larger model | 97.3% | about 12× | Not needed |
Decisions
-
Choose the model on measured results
Claude Haiku scored 96.7% on the 110-case safety set. A larger model scored 97.3% at about 12 times the cost. I shipped Haiku, at about US$0.0006 per answer.
A 0.6-point lower score, which I accepted because Haiku passed every crisis and privacy case in three runs.
-
Safety outside the model
A crisis card puts help lines in front of anyone who needs them. Rate limits and a kill switch let the clinic stop the assistant at once.
Some answers are less flexible by design.
96.7%- Test set
- 110 written cases
- Model
- Claude Haiku
- Scope
- Clinic website questions
-
Redesign the call when callers hang up
Early callers kept hanging up on the voice receptionist, so I redesigned the call flow to stop the hang-ups.
Call flows need testing with real callers, not only with scripts.
-
Send structured data, not transcripts
At the end of each call the agent posts structured JSON by webhook to the clinic’s automation, so staff get a clean intake record, not a recording to listen to.
A schema to keep in step with the clinic’s workflow.
{ "type": "new_client_intake", "for": "child", "format": "video", "preferred_times": ["weekday evening"], "booked": "Thu 18:00", "risk_words": false, "intake_form": "sent_by_text" }
Results
- 96.7%on a 110-case safety test set, with every crisis and privacy case passed in three runs
- 12×cheaper than the larger model, for a 0.6-point lower score
- $0.0006in US dollars per answer from the website assistant
- 236commits in ClientRelay, all mine
Tech
- Models
- Claude Haiku through the Claude API
- Voice
- Vapi, Twilio, ElevenLabs
- Automation
- Webhooks and Make.com
- Product
- ClientRelay on Supabase with row-level security, and Stripe
- Safety
- A 110-case written safety test set
Next case study
A 533-page clinic website moved from WordPress to Astro on Cloudflare, with every old URL crawled first.
Hiring for a remote role in Canada? Tell me about it.
This opens your email app with your note, addressed to me.
- ekamxjot@gmail.com