An AI receptionist for a small business: what it handles and where it fails

Most small businesses miss calls. The plumber is under a sink, the clinic receptionist is checking in a patient, the restaurant is in the middle of a dinner rush. Some callers leave a voicemail. Many do not, and a fair share of them simply call the next business on the list.

An AI receptionist is a reasonable fix for that specific problem. It is also easy to oversell. Here is what one does well, where it breaks, and the guardrails we would put in place before letting one answer your phone.

What an AI receptionist actually is

It is a voice agent connected to your phone number, often paired with a chat window on your website. When a call comes in, it answers, listens, and replies in a natural-sounding voice. Behind it sit a telephony provider, a speech layer that turns audio into text and back, a language model, and a knowledge base you write. That knowledge base holds your hours, services, service area, the prices you are comfortable quoting, and the rules for when to hand a call to a person.

An agent can do four things with a call. It can answer a question, book an appointment, take a message or an order, or pass the call to someone on your team. A well-built one does those four reliably and declines to do anything else.

What it handles well

After-hours calls are the clearest win. Someone who calls at 9 p.m. gets a real answer and a booking or a message, instead of a voicemail box they were never going to talk into.

Routine questions come next: hours, location, parking, whether you take walk-ins, which towns you cover, and price ranges you have signed off on, phrased the way you would phrase them, with the final price confirmed by a person. These calls are repetitive for staff and quick for an agent.

Booking works well when the agent is connected to your real calendar, with rules about appointment length, buffers between jobs, and which services need which person. It should check availability live rather than guess.

Structured messages are underrated. Instead of a voicemail that says “call me back,” you get the caller’s name, number, what it is about, how urgent it is, and the best time to reach them, in the same format every time, sent to the right inbox.

For restaurants, phone orders follow the same pattern. The agent takes the items, sizes, modifiers, pickup time, name, and number, reads the order back to confirm it, and sends it to the kitchen or the ordering system.

Then there is routing. The agent can recognize “my basement is flooding” or “you charged my card twice today” and put that call through to a person, or text whoever is on call, rather than booking it for next Thursday.

Where it fails

Angry and upset callers are the first weak spot. Someone calling about a botched job, a billing dispute, or a family emergency does not want a calm synthetic voice working through its options, and even a good agent tends to make that call worse. The right behaviour is to notice early and hand off.

Anything that needs judgment or a commitment is the second. Can you fit me in today? Will you match this quote? Is this covered by the warranty? Those are business decisions. An agent that answers them is making promises in your name that you will then have to honour or walk back.

Accents, noisy lines, and poor connections are the third. Speech recognition has improved a great deal, but a caller on the highway, on a job site, or with an accent the model handles badly will get misheard. Names, street addresses, and phone numbers are where the errors cluster. Reading details back to the caller helps, though it does not catch everything.

The fourth failure is the most dangerous, because it sounds fine. Asked something outside its knowledge base, a language model may produce a plausible answer anyway. It might describe a refund policy you do not have, a discount you never offered, or a service you do not provide. Callers believe it, because they heard it on your phone line.

The guardrails

Each of these is a decision someone has to make deliberately when the agent is built.

The agent answers only from an approved knowledge base. If the answer is not in there, it says it does not know and offers to take a message. You write and sign off on that knowledge base, and you update it when prices or hours change.

A person is always on offer. At any point the caller can say “I want to talk to someone” and get a transfer or a firm callback time. No loops, no making them ask three times.

Recording and transcription come with disclosure. The greeting should tell callers they are speaking with an automated assistant, that the call is recorded and transcribed, and why. Canadian privacy law, including PIPEDA for most private-sector businesses, generally expects you to tell people when you collect personal information and what you will use it for, and to keep it only as long as you need it. Check your own obligations with an adviser, especially in a regulated field.

The agent never takes card numbers by voice. Card details spoken aloud end up in audio files and transcripts, which is exactly where they should never be. If a call needs a payment, the agent sends a secure link by text or email, or a staff member handles it.

Someone reviews transcripts every week. Twenty minutes reading a sample of calls catches wrong answers, confused callers, and gaps in the knowledge base before they turn into complaints.

It must never approve a payment

This gets its own heading for a reason. Voice cloning is now good enough that a caller can sound like your owner, your bookkeeper, or a supplier you know well. An AI receptionist should never be a route by which anyone changes banking details, approves a payment, or authorizes a refund. It takes a message, and a person verifies the request through your normal finance controls, using a phone number you already have on file. Our post on AI voice cloning and payment fraud covers those controls in detail.

What it costs, in parts

We will not quote a figure here, because the honest answer depends on your call volume and how much the agent does. The cost has four parts, and any proposal you receive should show them separately.

Usage is charged by the minute, for the phone line and for the speech and language models. It scales with how many calls the agent takes and how long they run. A business catching a few after-hours calls a night pays very little. One routing every call through the agent all day pays a good deal more.

Then there is your number. You can keep your existing number and forward unanswered or after-hours calls to the agent, which is the low-risk way to start. Or you can port the number to a new provider, which gives more control but takes longer and needs care so you do not drop calls during the move.

Setup covers writing the knowledge base, connecting the calendar or ordering system, designing the handoffs, and testing with real calls.

Ongoing tuning means transcript review, knowledge-base updates, and adjustments when the business changes. The first month always turns up questions nobody thought to write down. Be wary of any quote that leaves out setup and tuning.

How to pilot it

Start with after-hours and overflow only. Forward calls that go unanswered after a set number of rings, plus everything outside business hours. Staff keep answering during the day exactly as they do now, and the agent catches only what would otherwise be missed.

Before you switch it on, find out how many calls you miss today. Most phone systems and carriers can show missed or unanswered calls, and that number is your baseline. Without it, you will not know whether the pilot did anything.

Run it for a month and review transcripts every week. Then decide whether to widen it, perhaps to the lunch hour or your busiest afternoon, or to stop.

What to measure

Keep it to a few numbers you can track in a spreadsheet.

• Calls the agent answered that would previously have gone unanswered.
• Bookings made or orders taken, and how many turned into an actual visit or sale.
• The handoff rate, meaning the share of calls passed to a person. Very low can mean the agent is answering things it should not. Very high means the knowledge base has gaps.
• Complaints, plus any call where the agent said something wrong. Each one should lead to a fix in the knowledge base that same week.

If the recovered calls are worth more than the running cost after a month, expand carefully. If they are not, you found that out cheaply.

Getting one set up

SetKernel builds AI receptionists for small businesses in Canada and the US, with the guardrails above built in from the first call. Our AI receptionist page explains how we approach it, and AI for small business covers the wider picture. If you want to try one, send us a short written brief saying what your business does, how calls come in today, and roughly how many you think you miss. We read it and reply in writing with a recommendation and a quote.