Should a small business use an AI receptionist?

The product is usually pitched as a replacement and is almost always worth buying as a supplement, and getting that backwards is what produces the disappointing outcomes. Compared against a competent person answering promptly, an automated receptionist is worse at nearly everything that matters — reading tone, handling the unusual case, exercising discretion. Compared against a phone ringing out, it wins by a distance, because the alternative is nothing at all.

So the size of the opportunity is not a property of the product; it is a property of your current unanswered-call rate. A business whose phone is answered promptly by a person throughout its published hours has a small opportunity here and should be sceptical of a large quote. A business whose calls ring out at lunch, in the evening, on Saturdays, or whenever both people are on jobs has a large one, and it is quantifiable before anyone is contacted.

Quantifying it takes a fortnight and needs nothing from any supplier. Call your own published numbers across a spread of times and record what happens — a person, a queue, voicemail, or nothing. Voicemail counts as unanswered, because a caller comparing three businesses treats it that way. The output is an answered-call rate by hour band, and it converts the sales conversation from a discussion about features into a discussion about a number that already exists.

What a well-configured system should be evaluated on is narrower than what gets demonstrated. It should answer immediately at every hour, identify what the caller wants at the level of routing rather than resolution, capture a name and a number reliably, offer a callback commitment somebody will meet, and hand off to a person when the call is outside its scope. Anything beyond that — quoting, diagnosing, committing to appointments — is scope creep into the territory where errors are expensive and irreversible.

The behaviour to test before purchase is what it does when it does not understand. A system that says so, takes the details and escalates is safe. A system that improvises a plausible answer creates a customer who acts on wrong information, and the business learns about it later, from the consequence. Every supplier will demonstrate the happy path; the informative demonstration is the one where you ask it something outside its configuration and watch.

The two practical failure modes are worth naming because they are common and both are avoidable. The first is the callback promise nobody staffed — the system commits to an hour that is not covered, and a waiting customer becomes a let-down customer at a documented time. The second is the return path: if the system captures a mobile number and sends a text from a number nobody monitors, the customer's reply goes nowhere, and the business has proved it is contactable and then that it is not.