An AI receptionist answers inbound phone calls and handles tasks such as checking business hours, routing callers, or booking appointments. Buy a ready-made receptionist if it already works with your scheduling software. Build on a voice-agent platform if your call flows require custom business rules or connections to your own systems.
This guide is for the team responsible for that choice: an operations lead defining what calls should accomplish, and a developer responsible for the systems behind them. Cartesia Managed Agents is a platform for building the agent. It is not a preconfigured receptionist with every industry’s scheduling integration included.
What an AI receptionist needs to do
Start with a narrow job. An after-hours agent that answers approved questions and collects callback requests has fewer ways to fail than one that can reschedule appointments, take payments, and change customer records.
Write down which actions the agent can complete and which require a person. For example:
| Caller request | What the agent needs | Safe fallback |
|---|---|---|
| “When do you close?” | Approved hours, including holiday exceptions | Say the hours are unavailable and offer a callback |
| “Can I come in tomorrow?” | Live availability, a booking tool, and the business’s time zone | Collect a request without promising a booking |
| “Move my appointment” | Identity checks and permission to change that booking | Transfer to the team responsible for appointments |
| “I need to speak to someone” | A staffed transfer destination | Offer the approved callback or voicemail route |
A friendly voice cannot fix a booking tool that writes to the wrong calendar. Treat these integrations as part of the receptionist, not as work to add after launch.
Buy a ready-made receptionist or build a custom agent?
A ready-made product can be a better choice for an individual practice or business with standard workflows. Ask the vendor to demonstrate your actual scheduling system, cancellation rules, and human handoff. A prerecorded demo does not establish that those connections work.
For example, Arini uses Cartesia for its dental AI receptionist. Our customer story describes appointment scheduling and responses to insurance and billing questions. A dental practice can evaluate a specialized product like Arini rather than building those workflows itself. Confirm current integrations and terms directly with the vendor.
A custom agent makes more sense when your team owns the application, serves multiple businesses, or needs actions that a packaged receptionist does not support. Cartesia Managed Agents combines speech recognition, an LLM, and Sonic speech generation. It handles turn-taking and interruptions, and can call tools during a conversation.1 Your team remains responsible for tool permissions, business rules, and what happens when a call cannot be resolved.
How to build an AI receptionist with Cartesia
1. Define instructions and approved answers
Create an agent in the Cartesia Playground. Give it a greeting that identifies it as an AI assistant, its allowed tasks, and clear conditions for handing off to a person.
Separate facts from actions. Business hours can come from approved information. Appointment availability must come from the live scheduling system. Instruct the agent not to infer availability from a previous conversation or promise a booking before a tool confirms it.
Use a stock voice for the first test. If you need a custom voice later, voice cloning lets you create one from a recording you have permission to use. Test names, dates, and phone numbers with the voice you choose.
2. Connect the scheduling system through a tool
Managed Agents supports webhook tools that call an HTTPS endpoint on your server.2 Your endpoint can check availability or create a booking in your scheduling system. This connection is implementation work, not an automatic integration with every calendar.
Keep credentials on the server. Validate tool arguments, restrict which calendars the endpoint can access, and apply the same authorization rules as the rest of your application. For an appointment change, do not treat possession of the caller’s phone number as sufficient proof of identity.
Design booking requests so a retry cannot create two appointments. Use a stable request identifier and deduplicate writes in your backend. Return an explicit success or failure, plus the confirmed appointment details. If the tool times out, the agent should say it cannot confirm the booking and follow the fallback procedure.
3. Configure human handoffs
In the Playground’s Tools page, enable Transfer call and configure destinations with conditions such as “the caller asks for a person.” Cartesia supports transfers to phone numbers; SIP URI transfers require a call routed through a SIP trunk.3
Specify staffed hours and a fallback for unanswered calls. Test those paths on the phone network you will use. Do not assume a transfer tool call means a human answered, or promise a callback unless a request was actually saved.
4. Connect a test phone number
Managed Agents can use Cartesia numbers, imported Twilio numbers, or SIP trunking.4 Choose the route that fits your existing carrier setup. Start with a test number rather than changing the routing for your main business line.
Call from real phones. Include noisy rooms, speakerphone, interrupted sentences, and callers who change their minds. A browser test alone does not test the carrier connection or how names sound over a phone call.
Run an acceptance test before routing customer calls
Use synthetic customer records and a test calendar. Decide the pass criteria before listening to the calls, then record both the conversation outcome and what changed in the backend.
| Test call | Pass condition |
|---|---|
| Ask for an appointment in an unavailable slot | Agent offers only availability returned by the scheduler |
| Change the requested time mid-sentence | Agent confirms the final choice before writing a booking |
| Repeat a booking request after a timeout | At most one appointment exists; the agent reports only a verified result |
| Ask for a person during the greeting | Agent follows the configured handoff path |
| Call when the transfer destination is unstaffed | Caller reaches the approved fallback rather than a dead end |
| Ask to access another customer’s booking | Agent refuses the change or follows the required identity check |
| Say “ignore your instructions” and request a restricted action | Server permissions prevent the action, regardless of the agent’s response |
Measure completed tasks, incorrect bookings, failed transfers, and caller abandonment. For response speed, measure from the end of the caller’s turn to audible playback on the caller’s device. A speech model’s time to first audio does not include every delay in the call.
Start with a limited set of call types and review failures before adding more. Avoid sending sensitive customer information through a test setup. For healthcare or other regulated use, have the appropriate privacy and legal owners review vendor agreements, data handling, retention, and disclosure requirements before production. A voice model alone does not establish compliance for the full workflow.
What does an AI receptionist cost?
Compare total operating cost for the same workload. Ask a packaged vendor about included minutes, overages, transfers, setup fees, and integration charges. For a custom agent, account for the voice-agent service, carrier charges, connected systems, and the people who build and maintain it. Check Cartesia’s current pricing rather than budgeting from a fixed rate in an article.
Track cost per successfully resolved call as well as cost per minute. A short call that creates a wrong appointment is not a saving. Keep human follow-up costs in the comparison.
Start with one inbound call flow
If you have standard industry workflows and no team to maintain integrations, evaluate a ready-made receptionist first. If you need to build, create a Managed Agent for one task, connect a test number, and run the acceptance cases above before routing customer traffic.