AI receptionists for small businesses: how they work, costs & risks

An AI receptionist is the software that answers the people who contact a business when nobody is free to. The term covers several different products, it is sold with a number of claims that do not survive contact with a real business, and the honest version of it is narrower and more useful than the marketing.

What an AI receptionist is

An AI receptionist is an AI agent placed at the point where enquiries arrive: the website, the contact form, the pages ads point to, and in some setups the phone. It greets the person, answers questions about the business from material it has been given, asks the qualifying questions a member of staff would ask, and then does something with the result, which usually means booking a time or handing the conversation to a person.

The word “receptionist” is doing a lot of work there. A human receptionist exercises judgment, recognises a regular, and knows when a rule should be broken. Software does none of that. What it can do is handle the routine majority of first contacts consistently and at three in the morning, which is a smaller claim and a real one.

It is worth being precise about the category, because “AI agent”, “automation” and “chatbot” are used interchangeably and mean different things. A receptionist is an agent: it interprets an open-ended conversation and decides what to do next within limits it has been given. That distinction is the subject of AI agent, automation or chatbot.

How it works

Under the marketing, an AI receptionist is four parts wired together.

1. A source of truth about the business

Services, prices, policies, hours, service areas and the things the business does not do. This is the part that decides whether the system is useful or embarrassing, and it is the part most implementations skimp on. An agent grounded in approved business material answers narrowly and admits ignorance; one left to improvise produces confident, fluent, wrong answers.

2. A conversation the model runs

The model handles the open-ended part: understanding what the person is asking, which is rarely phrased the way a form expects, and asking the qualifying questions back. It is given a defined job and a boundary, not a free hand.

3. Connections to the systems the answer lives in

A calendar, so an offered time is a real time. An inbox or CRM, so the enquiry lands where staff already work. Without these the receptionist is a chat window that produces a transcript nobody reads.

4. A rule about what it may do on its own

Every action gets assigned to one of three modes: it happens automatically, it is prepared and a person approves it, or the system only gathers context and a person decides. Booking an appointment into a calendar is usually safe to automate. Quoting a price, promising a deadline or handling a complaint is usually not. The framework is in human-in-the-loop decision levels.

AI receptionist, answering service, or chatbot?

Three terms used interchangeably by vendors, describing three different products with different prices and different failure modes.

What it costs

There are two costs and vendors usually quote one of them. The first is building and connecting the thing. The second is running it, which is model usage plus whatever monitoring and support arrangement you have, and it continues for as long as the system does.

On the build side, a receptionist is an AI agent like any other, and MOST’s published project ranges apply: a bounded AI workflow pilot runs $4,000 to $8,000 CAD, and a controlled AI agent in production runs $8,000 to $18,000 CAD, both fixed price. Work that reaches into several connected systems is scoped from $18,000 CAD. Model usage, monitoring and any support arrangement are quoted separately, because they are a running cost and pretending otherwise makes the first number look better than it is. The full breakdown is on the pricing page, and the anatomy of the running cost is in what an AI agent actually costs.

A subscription answering service is priced differently, per month rather than per project, and the trade is ownership. That is a real choice with a defensible answer either way; it is only a bad one when nobody asks which they are buying.

Where it fails

The recurring failures are not exotic. They are the same five.

• Nobody wrote down what the business actually does. The agent has to be grounded in real services, prices and policies. Where that material does not exist, the project turns out to be a documentation project wearing an AI costume.
• It books appointments nobody can service. A calendar that does not reflect real capacity produces confirmed bookings the business then has to cancel, which is worse than a missed enquiry.
• It is confidently wrong about a price or a policy. A fluent wrong answer is believed. Anything that commits the business to something belongs behind an approval step, not in the automatic tier.
• The handover drops the context. If a person picks up the conversation and the customer has to repeat everything, the system has made the experience worse than a voicemail.
• Nobody reads what it collected. An enquiry that lands somewhere the team does not already look is an enquiry lost with extra steps.

Privacy and call recording

An AI receptionist collects personal information from customers, which puts it inside Canadian privacy law. For a private business in British Columbia that means PIPEDA and the provincial guidance from the OIPC, and the practical questions are the ordinary ones: what is collected, on what authority or consent, where it is stored, who can reach it, how long it is kept, and whether a customer can have their record exported, corrected or deleted. Access should be limited to what the job needs and important events logged. That is the subject of AI and data privacy for Canadian businesses, which also covers whether a provider trains on your data under its business terms.

Call recording is a separate question and this page is not going to answer it. Recording a conversation raises consent obligations that depend on the jurisdiction and on how the recording is used, and MOST has not published guidance on it. A voice setup that records or transcribes calls needs that question settled before it goes live, with advice, not with a vendor’s reassurance. A text-based receptionist avoids the issue entirely, which is one of the reasons web-first setups are simpler to put in.

Who it fits, and who it does not

It is worth being blunt, because this does not suit everyone.

It fits a business where

• enquiries arrive outside the hours somebody is available to answer them;
• most first contacts are routine and answerable from written material;
• the next step is usually a booking, and there is a real calendar to book into;
• services, prices and policies are written down, or somebody is willing to write them;
• the volume is high enough that consistency matters more than personality.

It does not fit a business where

• every enquiry is genuinely bespoke, and the first conversation is the consultation rather than a route to one;
• the relationship is the product, and a customer reaching software instead of a person is itself the damage;
• the enquiry volume is low enough that a person can answer everything, in which case this is a solution to a problem you do not have;
• nothing is documented and nobody has time to document it, because the agent has nothing truthful to work from;
• the work is regulated to the point where a wrong answer is a reportable event rather than an inconvenience.

Trades and home-services businesses are the clearest fit, for the ordinary reason that enquiries arrive while everyone is on a job and the missed-call arithmetic is unforgiving. That case, including the correction of a widely repeated statistic about it, is worked through in AI for trades and home services.

If you want to know what building one involves rather than what it is, MOST’s scope and pricing for this work sit under AI receptionist and call handling.