Guide · AI & automation

AI receptionist for a small business: buy one first, and when building your own makes sense

AI receptionist for small business, honestly: 2026 pricing from Smith.ai, Goodcall and Ruby, the real cost of building one, and HIPAA rules for medical offices.

By Akash Bansal, founding engineerUpdated 5 November 20266 min read

If you run a small business and you're missing calls, my advice is to buy an AI receptionist this month and not to hire anyone to build one. That is an odd thing for someone who builds voice AI to say, so here's why, and the few situations where building your own is the better call.

My background on this: I built NextMantra solo, an AI recruiter that runs live voice interviews. The interview agent runs on Python and FastAPI on Pipecat, with Daily for real-time audio and Groq for fast inference, and multi-LLM scoring against the role's rubric. A receptionist is a simpler conversation than an interview, but the hard parts are the same: latency, interruptions, handover and knowing when the AI shouldn't answer.

What does an AI receptionist actually do?

An AI receptionist answers your business phone with a synthetic voice, understands what the caller wants, and does one of four things: answers a question, books or changes an appointment, takes a message, or transfers the call to a person. Good ones also send you a summary and a transcript after each call.

Under the hood it is a loop: speech-to-text turns the caller's words into text, a language model decides what to say or do, text-to-speech says it, and a telephony provider carries the call. Every step adds delay, which is why the products that feel natural are the ones that have tuned that loop hard. If your customers mostly type rather than call, the same buy-or-build question applies to an AI chatbot for your website, with much lower running costs per conversation.

How much does an AI receptionist cost in 2026?

Published plans run from about $79 to $500 a month for a small business, which is several times cheaper per call than a human answering service. These are the vendors' own pricing pages as of November 2026; check them before you sign, because prices in this market change often.

Do the maths on your own call log, not the plan names. A business with 300 calls a month averaging three minutes uses 900 receptionist minutes. That is beyond Ruby's top listed plan, and around $500 a month on Smith.ai's per-call pricing. Goodcall counts unique callers, so repeat callers are cheaper there.

Should you buy or build an AI receptionist?

Buy, unless one of three things is true. Off-the-shelf receptionists already integrate with common calendars and booking tools, come with a phone number and a dashboard, and have spent years on the latency and interruption problems you would be paying to solve again.

Best for buying: clinics, salons, trades, agencies, restaurants, any business whose calls are mostly routine. Best for building: businesses with unusual booking logic, high call volume, or a startup selling voice AI as the product.

What does it cost to build your own AI receptionist?

Building has two costs: the build itself, and per-minute running costs that never go away.

The build. My starting price for an AI feature or agent is $8,000, on a fixed quote after scoping. A receptionist that books into an existing calendar and hands over to staff usually sits at or a little above that. Integration with an unusual internal system, multiple languages or several locations pushes it higher. To see how those extras move the number for your own setup, try the free app cost calculator.

Running costs. On a hosted voice-agent platform such as Retell, the published pay-as-you-go range is $0.07 to $0.31 a minute, made of separate parts: $0.055 a minute for the voice infrastructure, then the language model, the voice and telephony on top (Retell pricing, vendor page). Its own example configuration comes to about $0.11 a minute.

Assumptions: three-minute average calls, a mid-priced model and standard voice, US telephony. Compare that with roughly $500 a month for 300 calls on a per-call plan. At 300 calls a month, the build cost takes years to pay back once you count maintenance. At 3,000 calls a month, with rules no product supports, the numbers start to favour building. That is the honest break-even: volume plus unusual logic, not either alone.

What about an AI receptionist for a medical office?

A medical or dental office can use an AI receptionist, but HIPAA changes the vendor checklist. Callers give names, dates of birth and reasons for visits, which is protected health information. HHS guidance says a cloud provider that creates, receives, maintains or transmits that information for you is a business associate, and you need a business associate agreement (BAA) with it, even if it only stores encrypted data (HHS, HIPAA and cloud computing).

• Ask for the BAA before you start a trial with real patients. Neither Goodcall's nor Smith.ai's pricing page mentions HIPAA, so ask sales directly. Retell lists HIPAA and BAA options in its plan comparison.
• Find out where call recordings and transcripts are stored, for how long, and who at the vendor can read them.
• Collect the minimum on the call: name, callback number, reason in a few words. Leave clinical detail for the clinic.
• Route anything that sounds urgent to a person or an emergency message, every time.

This is general information from an engineer, not legal advice. Have someone qualified review your setup. I've written more on patient-data rules in the dental practice automation story, and on Australia's health privacy rules in the physio clinic booking story.

What usually goes wrong with AI receptionists

• No clean way out. The caller wants a person and the AI keeps trying to help. Every setup needs a phrase like "speak to someone" that always works.
• Confident wrong answers. The model invents a price or a policy. Give it a short, approved list of facts and tell it to take a message for anything else.
• Talking over people. Interruption handling ("barge-in") is the difference between a natural call and an annoying one. Test it with real callers, including older ones and people on bad lines.
• Booking conflicts. The AI books a slot a human just booked. Booking must check the calendar at the moment of confirmation, not at the start of the call.
• Nobody reads the summaries. The receptionist takes 40 messages and nobody calls back. Decide who owns follow-up before launch.

Can the AI receptionist call people back?

Inbound is the easy case. Outbound is regulated. In 2024 the FCC ruled that AI-generated voices count as "artificial or prerecorded voice" under the TCPA, so automated outbound calls with an AI voice generally need the person's prior express consent (FCC). Start with answering calls. Add callbacks only to people who asked for one, and get advice on consent first.

How to choose and launch in two weeks

1. Export a month of call logs. Count calls, average length, after-hours share and the top ten reasons people call.
2. Write the answers to those ten reasons in plain sentences. This becomes the receptionist's knowledge.
3. Trial two vendors on a forwarded number, after hours only at first.
4. Listen to 30 real recordings from each. Count wrong answers, failed handovers and hang-ups.
5. Pick one, move it to daytime overflow, then to first answer if the numbers hold.
6. Price a custom build only if you hit a limit the products can't solve.

If you do hit that limit (rules no product supports, a voice agent that is your product, or volume where per-call fees hurt), that is the kind of AI build I do, from $8,000 on a fixed quote. Here's how I work. For the general question of whether you need an agent at all, see hiring an AI agent developer.