Skip to content
Free Consultation

Case study 05 / 06

Building an AI Phone Receptionist That Checks Availability and Books Appointments

Clinics and appointment-based businesses lose bookings every time the phone rings out. We built an AI receptionist that picks up, understands the caller, checks real availability, books the slot and takes a message when a human is needed. The same build runs Léa, the receptionist you can call from our site.

01At a glance

Client
Appointment-based businesses
Sector
Clinics and services
Challenge
Missed calls after hours and at busy times, and bookings lost between a voicemail and a callback.
What we built
A voice agent on a phone number, with a separate booking brain that computes free slots and records appointments.
Stack
  • Vapi for telephony and real-time voice
  • Deepgram speech recognition
  • Claude Haiku as the conversation model
  • Booking webhook on Vercel

Two halves: the voice and the brain

A phone agent has two very different jobs. The first is real-time conversation: answering, transcribing speech, deciding when the caller has finished, and speaking back without awkward pauses. The second is business logic: which slots are free, how long a treatment takes, what counts as a conflict.

We kept them apart. A voice platform (Vapi) handles the telephony and the turn-taking, with Deepgram for speech recognition and Claude Haiku as the conversation model, chosen for its balance of speed and quality on a live call. Whenever the agent needs to check or book something, it calls our webhook, which holds the business rules and answers in milliseconds.

Because the booking brain is separate, it can move to another voice platform without being rewritten.

What the receptionist can do

  • Answer in natural speech and understand what the caller wants
  • Check availability against opening hours and existing appointments
  • Book the slot, and refuse a conflicting one politely with alternatives
  • Recognise a closed day and say so
  • Take a message, or route the call, when a question needs a person, for example a medical question at a clinic

Tested without picking up a phone

Voice testing by hand is slow, so the project includes a local test that replays the exact requests the voice platform sends: an availability check, a booking, a conflicting booking, a message and a closed day. Every change to the rules runs through it before anyone dials the number.

What it costs to run

A phone number costs roughly 1 to 2 US dollars a month. Calls cost roughly 0.10 to 0.20 US dollars a minute all in, covering the voice platform, transcription, the model and the synthetic voice, depending on the providers chosen. Compare that with a receptionist's salary covering business hours only.

For a client deployment, two steps follow the prototype: appointments move to a permanent database, and availability syncs with the client's real calendar.

The takeaway

Keep the voice layer and the business rules apart. The voice platform can change; the rules that decide what gets booked are the part worth owning.

Next step

Want something like this for your business?

Tell us what takes up your team's time. In a free 30-minute call we will say plainly whether AI can help, and what it would take.