An AI receptionist that books appointments on the call

Handoff is designed first, because the calls that matter most are the ones the agent should not finish. In the first stage of every engagement, Map, we pull a week of call logs or recordings and sort every call into what the agent can handle and what it must pass on. It is the phone version of a human-in-the-loop design: the agent does the routine part, and a person takes every call that needs judgment.

During opening hours, transfers are warm where your phone system allows it. The agent tells the caller it is connecting them, rings the staff number set for that intent and speaks a one-line summary first ("New patient, wants a cleaning next week, prefers mornings"). If nobody answers within the time you set, the call returns to the agent, which apologizes, takes a message and texts the callback request to whoever is on duty.

After hours there is usually nobody to transfer to. The agent books what it can, takes messages for the rest and sends anything marked urgent to an on-call number. It never promises a callback time your team has not agreed.

Every transfer is logged with its reason. When the same question keeps causing transfers, that is a gap in the knowledge base, and it goes on the weekly fix list. The agent sits inside a voice AI build, and its knowledge base follows the same rules as our RAG knowledge systems.

The build is a fixed price between $12,000 and $30,000. The low end covers one location, one calendar and a knowledge base of about 50 questions. Several locations, a practice or property management system with a harder API, bilingual calls or outbound reminders push it toward the top. Most builds take 5 to 7 weeks, including 2 weeks of supervised live calls.

That price covers the full receptionist: answers from the knowledge base, booking, messages and transfers. If you need only one call flow, such as after-hours booking, that is a smaller voice AI build at $6,000 to $15,000, and it can go live in 3 to 4 weeks.

Usage is billed per minute and passed through at cost on your accounts. Retell AI publishes pay-as-you-go pricing of $0.07 to $0.31 a minute, depending on the model and voice (Retell AI pricing (opens in a new tab), checked October 2026). Other platforms, Vapi among them, bill the platform and the speech, model and phone providers as separate lines, so compare all-in figures.

A worked example: 1,500 call minutes a month at $0.12 a minute is $180 in usage, plus Twilio number and SMS fees.

The larger running cost is attention. Someone should read the weekly report of failed and transferred calls for the first 2 months. If nobody on your team has the time, an AI ops monitoring plan at $1,500 to $5,000 a month covers call review, knowledge base updates and model changes.

If your phone rings fewer than about 10 times a day and someone is always there to answer, keep the person. The agent will work, but the savings will take years to cover the build.

It is also wrong for calls where the conversation is the service. Clinical triage, a tenant with a burst pipe at 2am, a legal intake that needs a conflict check: these need a trained human, and the agent's only job is a fast, clean transfer. If most of your calls look like that, an answering service with trained staff is the better buy.

Two technical blockers come up in scoping. Booking needs an API. If your scheduling system has none, the agent can only take a request for a person to confirm, which removes much of the benefit. And if nearly every call is the same request, such as booking a viewing, a short phone menu that texts the caller a booking link costs far less. We will say so in the scoping call and build that instead.

This blueprint covers inbound calls. Outbound reminders are a separate decision: the FCC's February 2024 ruling brought AI-generated voices under the Telephone Consumer Protection Act (FCC-24-17 (opens in a new tab)), so we add them only after your counsel reviews consent. For the wider choice between agents, workflows and chatbots, read AI agents vs automation vs chatbots.