AI Receptionist for Small Business: A Buyer's Checklist
An AI receptionist demo is designed to impress. The voice is warm, the agent answers instantly, and it books a test appointment in under a minute. What a demo cannot show you is what happens on call number four hundred, when your calendar is full of old time-off blocks, a caller reads out a ZIP code, and the automation database behind the agent has been quietly filling up for weeks.
I audited and migrated a live AI receptionist that answers the phone and books repair appointments for a home service company. It takes a few hundred calls a month on five phone numbers, runs on Retell AI with an ElevenLabs voice on top of a large language model, and has 15 conversation states and 16 custom tools behind it. Everything in this checklist comes from that agent and the problems it had.
What is an AI receptionist for a small business?
An AI receptionist is a voice agent that answers your business line, finds out what the caller needs, checks who is available and books the appointment into your calendar. When it cannot book, it should hand the caller to a person with a short summary.
That last part is the difference between an AI receptionist and an answering service. An answering service takes a message. An AI receptionist is supposed to finish the job on the call, which means it reads your calendar and CRM on every call. That is where the money is, and it is also where almost every serious failure lives. The setup, testing and monitoring I do on these agents is described on AI receptionist for home service businesses.
How much does an AI receptionist cost?
An AI receptionist has two separate costs: the platform's per-minute charge for every call, and the one-time work to set up, test and connect the agent to your calendar and CRM. Some vendors bundle both into one monthly price, so ask for them separately.
For a sense of the platform side, Retell AI's pricing page lists AI voice agents at $0.07 to $0.31 per minute depending on the voice, the language model and the telephony you choose, and phone numbers at $2 a month. Your bill depends on call volume and average call length, so multiply your monthly call minutes by the per-minute rate before you compare offers.
The setup side is where quality is decided. My own work on this is priced as a fixed project: an audit of an existing agent from $700, and a new build priced as an operations automation build, from $1,800. I do not resell minutes or charge a percentage of call volume. The minutes are paid to the platform directly, on your account.
Which AI receptionist is best for a small business?
The best AI receptionist for a small business is the one that reads your real calendar and CRM correctly on every call and lets you check what it did. The voice and the platform matter less than how the agent is connected to your data and how its outputs are monitored.
When you compare products, judge them on things you can verify: whether it works with the calendar and CRM you already use, whether you can export recordings and transcripts, whether you own the account, and whether you can see what it booked against what your calendar allowed. A beautiful voice on top of a broken calendar check books jobs on technicians who are off.
What should you ask an AI receptionist vendor before signing?
Ask questions that show how the agent handles your data, what happens when it fails, and who owns it. A good vendor answers these in writing without hesitation:
- Which platform, voice and language model does it run on? You want to know whose deadlines and price changes you depend on.
- Whose account is it on? If the agent, the phone numbers and the call history live in the vendor's account, you lose them when you leave.
- How does it read my calendar? Does it check every event, or only a recent page of results? Ask what happens when the calendar has thousands of events.
- What happens when it cannot book? Who gets the handoff, in what form, and what is the caller told?
- Can I get recordings, transcripts and word timings for every call? Without them you cannot check what callers actually heard.
- What does it do with data a tool returns? An agent should never read a customer record aloud.
- What settings are switched on by default? Voicemail hang-up, contact memory and a fixed timezone came switched on in a new agent version I migrated, while the old agent had all three off.
- What platform deadlines are coming? According to Retell's deprecation notice, new multi-prompt agents cannot be created from December 15, 2026, and existing ones cannot be edited from February 1, 2027.
- How will you know if it starts booking wrong? "We watch the dashboard" is not an answer.
What test calls should you make before going live?
Before an AI receptionist answers real customers, call it yourself with the situations that break agents, not the easy booking from the demo. Each test should end with a check of what was booked in the calendar and what the audio sounded like, not only what the transcript says.
The test calls I would make on any agent:
- A booking in a busy week, on a technician who has an old time-off block, not only a clean test week.
- A price question. Does it give a useful answer or stall?
- A caller who reads out a ZIP code and phone number. Listen for the agent repeating data back in a way a customer would find strange.
- A caller who asks for a person. The handoff should be fast, and the team should get the name, number and problem.
- A wrong number or an off-topic call. Agents often say their oddest sentences here.
- An after-hours call, if the agent covers nights and weekends.
- An outbound callback, if the agent calls customers back. Its first line often comes from outside the agent, from an automation platform.
Then check the word timings, not only the transcript. In the agent I audited, normal words ran 0.2 to 0.4 seconds, so any single word longer than 2.5 seconds is a red flag. That check is how I found one dash in the agent's text producing 3.4 seconds of garbled sound that a tester described as a foreign language, while the transcript read as clean English. The full story is in how one em dash broke the voice.
What failures only show up after an AI receptionist goes live?
The most expensive AI receptionist failures appear weeks or months after launch, when your data grows past an assumption nobody knew had been made. They do not throw errors, so every call is logged as a success while something is wrong.
Four examples from the same live agent:
- It booked jobs on technicians who were off. The scheduling system held 7,801 calendar events, and the agent only ever read the 600 most recently touched ones. At launch there were fewer than 600 events, so the bug did not exist yet. It grew one booking at a time. The anatomy of it is in the 600 of 7,801 bug.
- The phone bot went down at 3 AM. The automation database behind the agent filled its 5 GB disk with saved run history. Downtime was about 20 minutes, and the settings that prevent it are a few lines. What filled the disk is in the 3 AM outage write-up.
- Callbacks could go to the wrong agent. A callback workflow treated agent version 0 as missing and silently fell back to an old version.
- The vendor's convert button produced an empty agent. The one-click conversion from the old agent type produced a template with 2 nodes and 0 tools, against 15 states and 16 tools in the real agent. I rebuilt it with a script and compared the result with the original before anything went live.
None of these looked broken from the outside. Calls connected, the voice sounded fine and the dashboard showed green.
How do you monitor an AI receptionist after launch?
Monitor what the agent outputs, not only whether it is running. That means comparing the bookings it made with what the calendar allowed, and comparing booked calls with jobs in your CRM, on a schedule.
A simple monitoring set for a small business:
- Bookings against the calendar, so a booking on an unavailable technician shows up as a number, not as an angry customer.
- Booked calls against CRM jobs, so a drop in booking rate is visible within days.
- The word-timing check on every call, which flags audio glitches without anyone listening.
- A disk or storage alert on whatever database sits behind the automation.
- A test call or webhook check that expects a valid answer, so the agent being down is caught whatever the cause.
Should you audit your current AI receptionist or replace it?
If you already have an agent taking calls, an audit is usually the better start, because most problems sit in the setup and the data, not in the platform. Replacing the agent can carry the same problems into a new system.
During the migration I tested both versions with the same simulated calls. The new agent passed 8 of 10 scenarios, and the old one failed the same 2 scenarios word for word. Those failures lived in the prompt, not in the move. The dash glitch was the same story: the old version had produced the same noise days earlier. If I had rolled back, the bug would have stayed live and the migration would have taken the blame.
When is an AI receptionist the wrong choice?
An AI receptionist is the wrong choice if most of your calls are complaints, warranty disputes or long diagnostic conversations, or if your team does not keep the calendar and CRM up to date. In those cases a person will do better, or the agent will book on bad data no matter how good the prompt is.
If your real problem is losing after-hours calls, start smaller. Forward missed calls to a person or a simple answering service, measure how many calls you were actually missing, and then decide whether full booking is worth building.
How I audit, test and monitor these agents is on AI receptionist for home service businesses, and the fixed price for an AI receptionist audit is on pricing. If your agent is live and you are not sure what callers actually hear, start with the free written diagnostic.
Related: