Work Case study · AI voice agent
The phone line that never puts anyone on hold
A private medical centre with five branches was missing almost a third of its calls every morning. We built an AI phone agent that answers instantly, day and night, books channelling appointments, and hands anything clinical or complicated to a person.
- 2%of calls missed, down from 31%
- 0 swaiting to be answered
- 19%of bookings made after hours
- −64%routine calls reaching the front desk
Where they started
Engaged tones from seven to ten
Channelling bookings for the week open early, and every morning the same thing happened: all five branches’ lines rang at once while receptionists checked in patients at the counter. Callers heard an engaged tone or waited, and the longer people wait on a phone line, the more of them hang up and try somewhere else1. Almost a third of calls were never answered.
Most of those calls were simple. Which days does Dr. Fernando see patients? Is there a number free on Tuesday? Where is the Nugegoda branch? After 8 pm there was nobody to answer them at all.
- Client
- Private medical centre, Western Province (name withheld)
- Scale
- 5 branches, about 60 visiting consultants
- Calls
- About 1,800 a day at peak
- Timeline
- 10 weeks to the first live branch
01 A call
Listens, checks, books, confirms
The call’s audio streams live from the phone network2 into speech recognition that transcribes as the caller speaks3. A language model decides what to say, and when it needs facts or wants to act, it can only use the tools we give it: look up a doctor’s clinic days, check free numbers, book, cancel, send an SMS4. It cannot invent a slot that doesn’t exist.
We tuned recognition to Sri Lankan English, doctors’ names and branch names, and tested it with real local voices, because speech recognition accuracy varies a lot between speakers5.
Incoming call · Nugegoda lineLIVE 01:24
Hello, I want to channel Dr. Fernando, the cardiologist, for next Tuesday.
Of course. Dr. Fernando sees patients at our Nugegoda branch on Tuesdays from 4:30 pm. Would that branch work for you?
Yes. What’s the earliest number?
Checking Dr. Fernando · Tue · Nugegoda
Number 12 is free, at about 5:15 pm. Shall I book it? May I have the patient’s name?
Yes, for my mother, Kamala Perera.
Booked · No. 12 · SMS sent
Done: Kamala Perera, number 12 with Dr. Fernando, Tuesday at Nugegoda. You’ll get an SMS with the details now.
Sounding human
Under a second,
or it feels wrong
In conversation, people answer each other after a gap of about a fifth of a second, across languages and cultures6. A phone agent that pauses for three seconds sounds broken, and callers start talking over it.
So nothing waits for anything else. Transcription streams word by word, the reply starts generating before the caller has quite finished, and speech begins playing as soon as the first words are ready. A typical turn takes under a second.
About 0.8 seconds from the caller’s last word to the agent’s first.
02 Where the agent stops
It knows what it shouldn’t do
This is a medical centre, so the rules came first, following the WHO’s guidance on AI in health: protect people’s safety, be transparent, and keep humans in charge of decisions about care7.
- No medical advice, ever. Questions about symptoms or medicines go to a nurse or doctor.
- Emergencies first. A caller describing an emergency is told to call 1990, Sri Lanka’s free ambulance service8, before anything else.
- A person whenever it’s needed. During opening hours the agent transfers difficult calls to the front desk; after hours it books a callback.
- Sinhala and Tamil callers are offered a callback from a staff member who speaks their language.
- Private by default. Recordings and transcripts are kept for 90 days, visible only to authorised staff, in line with Sri Lanka’s Personal Data Protection Act9.
Results
Three months live
- 2%
Of calls missed
Down from 31% at the morning peak, across all five branches.
- 19%
Booked after hours
Appointments made between 8 pm and 7 am, calls that used to go unanswered.
- −64%
Routine calls at the desk
Receptionists now spend that time with the patients standing in front of them.
- 100%
Of clinical questions to people
Every question about symptoms or medicines reached a nurse or doctor.
References
The research behind it
Every method on this page builds on published research, standards or official documentation. Numbers in the text point here.
- Gans, N., Koole, G., & Mandelbaum, A. (2003). Telephone call centers: Tutorial, review, and research prospects. Manufacturing & Service Operations Management, 5(2), 79–141. doi.org/10.1287/msom.5.2.79.16071
- Twilio. Media Streams. Twilio Docs. www.twilio.com/docs/voice/media-streams
- Deepgram. Streaming audio transcription. Deepgram Docs. developers.deepgram.com/docs/live-streaming-audio
- Anthropic. Tool use with Claude. Claude Developer Platform documentation. platform.claude.com/docs/en/agents-and-tools/tool-use/overview
- Koenecke, A., Nam, A., Lake, E., Nudell, J., et al. (2020). Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14), 7684–7689. doi.org/10.1073/pnas.1915768117
- Stivers, T., Enfield, N. J., Brown, P., Englert, C., et al. (2009). Universals and cultural variation in turn-taking in conversation. Proceedings of the National Academy of Sciences, 106(26), 10587–10592. doi.org/10.1073/pnas.0903616106
- World Health Organization. (2021). Ethics and governance of artificial intelligence for health: WHO guidance. www.who.int/publications/i/item/9789240029200
- 1990 Suwa Seriya Foundation. 1990 Suwa Seriya: free island-wide pre-hospital emergency care. www.1990.lk/
- Parliament of Sri Lanka. (2022). Personal Data Protection Act, No. 9 of 2022 (as amended). www.dpa.gov.lk/guidelines.php
How many calls is your front desk missing?
Tell us how your phone lines work today. The first conversation about an AI voice agent is free.