AI Receptionist That Books Appointments: Build vs Buy Costs

An AI receptionist that books appointments costs $25 to $300 a month as a subscription, or roughly $0.05 to $0.11 per minute built from parts, before you count the build fee. Both numbers are real, and both are useless without knowing the failure modes that decide whether callers actually get booked. This post breaks down the published per-minute prices of every component, works through the math at realistic call volumes, and gives you a decision rule for buying versus building.

Everything priced below comes from the vendors' own pricing pages, checked in November 2026. Where a number comes from a vendor's marketing blog rather than a pricing page, I say so.

What a booking receptionist actually has to do

The scope matters because it drives cost. A receptionist that "takes messages" is a cheap product. One that books appointments has to do all of this during a single call:

• Answer within two rings, any hour, and greet the caller in a voice that does not make them hang up.
• Understand a request phrased a hundred ways ("Can I see Dr. Maina on Thursday?", "Do you have anything next week?", "I need to move my appointment").
• Check live availability against a real calendar, not a cached copy.
• Handle the back-and-forth of picking a slot, including "no, earlier" and "what about the week after?".
• Write the booking, with the caller's name and contact details spelled back and confirmed.
• Send a confirmation by SMS or email.
• Recognize when it is out of its depth and hand off to a human with context, not a cold transfer.

That is the product. Anything less is voicemail with extra steps. Our earlier guide on voicebots for appointment booking covers the conversation design side; this post is about what it costs and how to decide who should run it.

What buying costs: the SaaS price landscape

Published AI receptionist plans in late 2026 cluster into three pricing models. The numbers below come from a pricing survey of 28 providers published in November 2026, cross-checked against individual vendor pages like this comparison.

Per-minute plans. RingCentral's AI receptionist starts at $49 a month for 100 minutes. My AI Front Desk is $99 for 200 minutes. Overage rates run $0.35 per minute and up. These plans look cheap until you map them to real volume: 100 minutes is roughly 40 short calls.

Per-call plans. Smith.ai's AI tier includes 75 calls for $150 a month, then $2.50 per extra call. At 600 calls a month that is $150 plus 525 times $2.50, or about $1,460. Per-call pricing is fine for a solo practice and punishing at scale.

Flat plans. NextPhone and similar providers charge around $199 a month for unlimited calls. This is the model that makes the SaaS route attractive for a busy single location: predictable cost, no metering anxiety.

For comparison, a human answering service runs about $1.50 to $5 per minute, and a full-time receptionist costs several thousand dollars a month. The AI options are one to two orders of magnitude cheaper. That gap is why the question is rarely "AI or human" anymore. It is "which AI, and who operates it."

What building costs: the component stack, priced from primary sources

A self-built receptionist is four metered services plus an orchestration layer. Here is each one at published pay-as-you-go rates.

Telephony. Twilio charges $0.0085 per minute to receive a call on a US local number, plus $0.0044 per minute for Media Streams to pipe the audio to your speech-to-text engine, according to the Twilio voice pricing page. Call recording adds $0.0025 per minute. Budget about $0.013 per minute all-in for the phone leg.

Speech-to-text. Deepgram's pricing page lists Nova-3 streaming at $0.0077 per minute, billed per second with no rounding. Multilingual recognition is $0.0092 per minute.

The language model. This is the line item people overestimate. On Anthropic's API pricing, Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens. A booking conversation is short turns with a growing context: a reasonable estimate is 4,000 to 6,000 input tokens and a few hundred output tokens per minute of call, which lands around $0.005 to $0.01 per minute. Newer small models push this lower still. The model is the cheapest part of the stack, which is why "just use a bigger model" is rarely the right answer to a voice agent problem. Latency is.

Text-to-speech. This varies most. Deepgram's Aura-2 voices cost $0.030 per 1,000 characters. An agent that speaks about 40 percent of call time generates roughly 300 characters per minute of call, so about $0.009 per minute. Premium voices cost more: ElevenLabs works out to roughly $0.17 per minute of generated speech on its Pro plan, which translates to around $0.07 per minute of call. This is the quality-versus-cost dial. Callers tolerate a slightly robotic voice on a confirmation line; they notice a 900-millisecond pause far more.

Orchestration. Something has to hold the call state, stream audio between services, and call your calendar API. LiveKit Cloud includes 5,000 agent minutes on its $50-a-month Ship plan, then $0.01 per agent-session minute. LiveKit's own cost calculator estimates a complete phone agent, including telephony, STT, a hosted LLM, TTS, and observability, at $0.0672 per minute.

Add the parts yourself and a do-it-yourself stack lands at roughly $0.05 to $0.11 per minute, depending mostly on the voice you choose. LiveKit's estimate sits comfortably inside that range. If you want one meter instead of five, Deepgram's bundled Voice Agent API is $0.075 per minute for its standard tier, which includes STT, a managed LLM, and TTS in one rate.

A worked example: the three-location clinic

Take a dental clinic group with three locations, 600 inbound calls a month, averaging 2.5 minutes each: 1,500 minutes.

• Flat-rate SaaS at $199 per location: $597 a month, nothing to maintain.
• Per-call SaaS (Smith.ai-style): about $1,460 a month. Already dead at this volume.
• Built stack at $0.09 per minute: $135 in usage, plus $40 to $80 for a small server and monitoring, so about $200 a month.
• Human answering service at $1.50 per minute: $2,250 a month.

The built option saves about $400 a month against the flat SaaS. Now the honest part: a competent single-location build costs $3,000 to $12,000 according to KeenCraft's cost breakdown, and a multi-location build with CRM writes runs $12,000 to $30,000. At $400 a month of savings, a $10,000 build takes 25 months to pay back, and you are carrying the maintenance and the failure modes below the whole time.

Flip the volumes and the answer flips. At 10,000 minutes a month, the flat SaaS plans either cap out or the per-minute overages bite, while the built stack costs the same $0.05 to $0.11 per minute and you control concurrency and data. KeenCraft's threshold estimate, around 10,000 minutes a month, matches the arithmetic.

Four failure modes the demo never shows

Every vendor demo books an appointment flawlessly. Production is different. These are the four failures that decide whether your receptionist books real appointments or generates expensive noise.

1. Double booking through stale availability. The agent reads open slots, the caller thinks for 30 seconds, and meanwhile a patient books the same slot on your website. The agent then writes a booking for a slot that no longer exists. The fix is to re-verify availability immediately before the write and to treat a conflict response from the calendar API as a normal path, not an error: offer the next two slots out loud and keep going. If your booking tool cannot reject a conflicting write, you do not have a booking system, you have a suggestion box.

2. Timezone confusion. Cal.com returns available slots in UTC, as this build walkthrough notes, and timezone bugs are the most common mistake in calendar integrations. A caller in Nairobi says "Tuesday at nine"; your server sits in a US region; the model has to resolve the caller's timezone, not the server's, and say the slot back in the caller's local time. Nairobi has no daylight saving shifts, which removes one class of bug, but only if the code stores offsets rather than abbreviations.

3. The misheard confirmation. Speech recognition on a noisy line will turn "March 5th" into "May 3rd" often enough to matter. The reliable pattern costs one extra conversational turn: the agent reads the full date, time, and name back and waits for an explicit yes before calling the booking tool, the same pattern Revve's appointment-booking recipe prescribes. Skipping this turn is how you get a calendar full of confident mistakes.

4. The unbookable call. A caller who is angry, hard to understand, or asking for something outside the script will burn minutes at full per-minute rates while going nowhere. Industry practitioners estimate hidden costs like silence billing and concurrency limits add 15 to 30 percent to naive per-minute estimates, per Plura's operator guide. Your agent needs a containment policy: after two failed attempts to understand, apologize and hand off to a human or take a structured message. Measure cost per booked appointment, not cost per minute. A cheaper agent that completes fewer bookings is the more expensive agent.

A practical build path

If you decide to build, this is the shortest path that survives production.

1. Start with the calendar, not the voice. Get the booking logic right as a plain function: check availability, hold, confirm, write, send SMS. Cal.com's API authenticates with a Bearer key and allows 120 requests per minute by default, per its API documentation, which is ample for a receptionist. Note that Cal.com deprecated its Platform plan for new signups in December 2025; build against the standard v2 API. Self-hosting Cal.com is also an option if data residency matters.
2. Prototype the conversation with a bundled voice API. Deepgram's Voice Agent tier or Twilio's ConversationRelay gets you a working call in days at $0.07 to $0.075 per minute. Do not optimize cost before you know the conversation works.
3. Instrument before you scale. Log every tool call with timestamps. The metrics that matter are containment rate, bookings completed per call, and cost per booked appointment.
4. Move to a component stack only when volume justifies it. Split STT, LLM, and TTS onto LiveKit or your own infrastructure past roughly 5,000 to 10,000 minutes a month.
5. Wire the follow-through. A booking without a reminder is a no-show waiting to happen; pair the agent with automated appointment reminders.

A decision checklist

Buy a subscription when most of these are true:

• Call volume is under roughly 2,000 minutes a month.
• Booking rules are simple: one calendar, one service type, no CRM writes.
• Nobody on your team will own the system week to week.
• You need it live this month.

Build (or commission a build) when most of these are true:

• Volume exceeds roughly 10,000 minutes a month, or per-call SaaS pricing exceeds $1,000 a month.
• The agent must read and write your CRM, not just a calendar.
• You handle sensitive data and want call audio and transcripts inside your own accounts.
• You need behavior a template cannot express: multi-location routing, insurance checks, deposit collection.

The middle ground, 2,000 to 10,000 minutes, is where the $199 flat plans earn their keep while you learn your real volumes and failure rates.

Common mistakes

• Comparing headline per-minute rates. A $0.05 headline that excludes telephony, TTS, and silence is more expensive than an honest $0.075 bundle. Always price the whole minute.
• Optimizing the model before the conversation. Upgrading from a fast small model to a frontier model doubles or triples your LLM line and adds latency callers hate, while doing nothing for double bookings.
• No handoff path. An agent with no way to reach a human does not reduce support load; it moves complaints to other channels.
• Skipping the read-back. Confirmed details prevent the worst failure: a confidently wrong booking.
• Ignoring recording law. Recording consent rules vary by jurisdiction. If you record calls, say so in the greeting, and budget the recording and storage line items.

Frequently asked questions

How much does an AI receptionist cost per month? Published subscription plans run $25 to $300 a month for small-business volumes, with $199 flat unlimited plans common at the top end. A self-built stack costs roughly $0.05 to $0.11 per minute of call, so $100 to $220 a month at 2,000 minutes, plus the one-time build cost.

Can an AI receptionist book into Google Calendar or Outlook? Yes, through a scheduling layer. Cal.com syncs with both Google Calendar and Outlook natively, so an agent wired to the Cal.com API checks your real availability regardless of which calendar your staff uses.

What happens when the AI cannot handle a call? A well-built agent detects repeated misunderstanding or out-of-scope requests and either warm-transfers to a human with the conversation context or takes a structured message. If a vendor cannot show you this path in the demo, assume it does not exist.

How long does it take to build one? A simple single-intent booking agent takes about 6 to 8 weeks including telephony setup, conversation design, and testing, based on RaftLabs' development cost breakdown. Multi-intent agents with deep CRM integration take longer and dominate the build budget through integration work, not model work.

Does it work in languages other than English? Multilingual speech recognition is a paid add-on everywhere: Deepgram charges $0.0092 per minute for multilingual Nova-3 streaming versus $0.0077 for English-only. Test with real callers in the actual accent and language mix before launch, because benchmark accuracy and noisy phone-line accuracy are different things.

Is per-call or per-minute pricing better? Per-call pricing punishes short calls and high volume; per-minute pricing punishes long, unresolved calls. Estimate your real call-length distribution, then run both models against it. Above a few hundred calls a month, flat plans usually win on predictability.

Where to go from here

The receptionist is the entry point, not the end state. The same stack that books appointments can qualify leads, chase invoices, and follow up on quotes, which is the shape of a full AI employee: an always-on agent that owns a job rather than a phone line. If you want the scheduling backbone run for you, we deploy and manage Cal.com and the rest of the open-source stack for clients, and we build custom voice agents when your call volume and integration needs justify one. Bring your monthly call minutes and your current answering bill to a conversation, and the build-versus-buy math takes about fifteen minutes.