When AI Answers Your Business Phone, Who Is Speaking?

A plumber misses eleven calls on a Tuesday because he is under a sink. A boutique misses four because two people walked in at once. Neither of them has a receptionist problem they can solve by hiring, and both of them can now put a voice on the line that answers in one ring, knows the opening hours, and books an appointment into the same calendar a human would have used.

That is a genuinely useful thing, and it is also the point at which a business quietly acquires a spokesperson. The questions worth answering before you switch one on are not technical. They are about what the voice may commit you to, what it has to tell the caller, and where it must stop.

Which calls are actually regulated?

The ones your system places, not the ones it receives. This distinction does most of the work in this subject and it is the one that gets collapsed in nearly every vendor pitch.

In February 2024 the Federal Communications Commission adopted a declaratory ruling, FCC 24-17, treating a voice generated by AI as an artificial or prerecorded voice for the purposes of the Telephone Consumer Protection Act. Mayer Brown's analysis of the ruling records the operative standard: a caller needs prior express consent from the called party before making a call that uses an artificial or prerecorded voice simulated or generated through AI technologies. The statute restricts calls made to residential lines and mobile devices without that consent.

Read the direction of travel in that sentence. The Act is about somebody being called. A customer who dials your shop number has initiated the contact, and the consent question does not arise in the same way. So an AI that answers your phone sits in a much quieter part of the law than an AI that rings a list of numbers.

The trap is the feature that sounds most helpful. Voice products routinely offer to call the customer back when the part arrives, confirm an appointment the day before, or chase a quote that went unanswered. Every one of those is an outbound call using a synthetic voice, and every one of them lands squarely inside the framework above. If you want that capability, get consent explicitly, record when and how you got it, and keep the option to deliver the same message by text instead.

What does the voice have to tell the caller?

That it is a machine, in most of the world, and increasingly by law rather than by courtesy. The EU AI Act's transparency obligation became applicable on 2 August 2026, and it is written in a way that covers a telephone agent squarely.

Article 50(1) requires providers to design systems meant for direct interaction with people so that the person is told, one way or another, that an AI system is on the other end. A carve out applies where the fact would already be obvious to a reasonably observant person in the circumstances. A synthetic voice that answers a shop line and holds a natural conversation is precisely the case where that carve out cannot be relied on, because the entire product claim is that it does not sound obvious.

The practical version is one sentence at the top of the call, before anything else happens. Name the business, say that an automated assistant is answering, and offer the route to a person. Doing it first is better than doing it well: a disclosure that arrives after the caller has explained their problem reads as a trick even when it was not meant as one. We went through the same question for text channels in our piece on what a chatbot has to disclose and when, and the answer is consistent across both.

What can it close, and what must it hand over?

The dividing line is whether the outcome is already determined by rules you published. If the answer sits in your opening hours, your price list or your booking policy, the machine is reading something out. If the answer requires judgment about this particular customer, it is making a decision, and it should not be.

What if the assistant promises something you do not offer?

You are likely to be held to it. This is not a theoretical risk and it has already been litigated in the text channel, where a company argued unsuccessfully that its own chatbot was a separate entity responsible for its own statements.

The reasoning transfers directly to voice, and arguably more forcefully, because a caller has even less opportunity to check a claim against a published page while they are listening. The defensive measure is not a disclaimer read at speed. It is a narrow scope: the assistant quotes only from a price list it reads live, refuses to negotiate, and hands over the moment a caller pushes past what it can see. We set out how that liability works and what a business can do about it in the piece on what happens when your bot makes a promise.

Setting one up without creating three new problems

The configuration that works is unglamorous and takes an afternoon. Most of it is deciding what the thing may not do.

Start with the greeting and the disclosure, in that order, and keep it under six seconds. Callers abandon long menus and they abandon long introductions for the same reason.

Give it exactly one data source for anything factual. If stock comes from your storefront database and the calendar comes from your booking system, connect both rather than pasting a summary into a prompt, because a pasted summary is out of date the day after you write it. This is much easier when the booking a caller makes lands in the same place as the orders from your website, which is one of the arguments for running the storefront on infrastructure you control, as our approach to building the shop itself assumes.

Write the escalation list before the script. Four categories belong to a person: complaints, refunds, anything priced outside your published list, and any caller who sounds distressed or confused. That last one is a judgment the model will not make reliably, so build the exit on the caller's request rather than on detection. The phrase a caller is most likely to use is a plain request to speak to someone, and it should work at any point, every time, without a loop.

Does it actually pay for itself?

For an appointment business the answer is usually yes, and the arithmetic is simple enough to do on the back of an envelope before you sign anything. For a shop where calls are mostly questions rather than bookings, the answer is much less obvious and the honest test is a month of measurement.

Work it out this way. Count the calls you currently miss in a normal week, not a busy one. Estimate how many of those would have become a booking or a sale, which for most trades is a lower fraction than owners guess, somewhere between one in four and one in ten. Multiply by your average job value, then subtract the monthly cost of the service and the hour a week you will spend reading transcripts in the first month. If the result is not clearly positive on conservative numbers, it is not a good buy yet.

The failure mode to price in is the call that the assistant handles badly and that you never hear about. A missed call at least leaves a number in your log. A mishandled call leaves a customer who quietly went elsewhere, and unless you read the transcripts you will not know it happened. Budget the reading time honestly. It is the part everyone skips and it is the only mechanism you have for finding out whether the thing works.

How do you test one before it touches a real customer?

Call it yourself, twenty times, badly. That is the whole method, and it finds more problems in an hour than a month of live traffic because you can deliberately produce the situations a real caller produces by accident.

Run a fixed list. Ask a question it should answer, and check the answer against the source rather than against your memory. Ask for something you do not sell, and listen for whether it invents an alternative. Ask for a discount twice over, and see whether the second refusal is as firm as the first. Interrupt it mid sentence and change your mind about the date. Give a name it will struggle to spell and check what lands in the calendar. Say you want to complain, and time how long it takes to reach a person. Call from a car with the window down. Call while someone else is talking nearby.

Write down what happens in each case, because the useful output of this exercise is not a pass or fail, it is the list of scenarios that go to a human. Every failure you find is a line in the escalation rules rather than a reason to abandon the project. The one result that should stop you is a confident wrong answer about price, stock or availability, because that is the failure a caller cannot detect and will act on.

The transcripts are the part with lasting value

An overlooked consequence of putting a machine on the phone is that you suddenly have a written record of what customers ask, in their own words, at volume. Most small businesses have never had this. It is more useful than the booking feature.

Three things fall out of a month of transcripts. The first is the list of questions you answer constantly, which is your missing website content, written for you in the phrasing real buyers use rather than the phrasing you would have guessed. The second is the vocabulary gap: the words customers use for your products, which are frequently not the words on your product pages, and which are the words they also type into a search box. The third is the pattern of what people ring about at particular hours, which tells you when it is worth being reachable and when it genuinely is not.

Acting on the first of those is usually the highest return of the whole exercise. A question asked forty times a month is a page you should have, and publishing it reduces the calls rather than automating them. Reducing the calls beats answering them faster, and it is the outcome a voice product will never suggest to you because it is paid by the conversation. That is not a criticism of the vendors so much as a reason to read your own transcripts rather than their dashboard, which counts calls handled and has no column for calls that should never have happened. The same reasoning applies to appointment reminders and the practical ways of getting a booked customer to actually turn up, which we covered separately in the piece on cutting the no show rate.

What still does not work well

Accents and noise remain the real limit, and no amount of prompting fixes them. A caller on a building site, or one using a speakerphone in a moving car, produces the transcription errors that turn into wrong bookings, and they are common in exactly the trades where missed calls cost the most.

Interruption handling is the second weak point. Humans talk over each other constantly and repair the conversation without noticing. Voice agents are better at this than they were, and they still lose the thread when a caller changes their mind mid sentence, which is a normal thing for a customer to do while looking at a calendar.

The third is anything requiring memory of a relationship. A regular customer who says the usual, please expects to be understood, and an assistant that has to ask four questions to establish what the usual is has made the experience worse than voicemail. If your business runs on repeat customers who know you, that is a strong argument for using the assistant only outside opening hours.

None of these are reasons to avoid the technology. They are reasons to deploy it where the calls are simplest and to keep measuring. Somebody phoning to ask whether you are open on Sunday is well served by a machine. Somebody phoning because their order arrived broken is not, and the difference between those two callers is the entire design problem. Get that boundary right and the tool earns its place quietly. Get it wrong and you have automated the moment your business was most likely to lose someone, which is an expensive way to save a ring.

One more caution that belongs here rather than in a security article. A synthetic voice answering your line trains your customers to accept a synthetic voice from your business, which makes them fractionally easier to defraud by somebody imitating you. It is a small effect and it is not a reason to avoid the tool, but it is a reason to keep a verification habit on anything financial, of the kind we set out in the piece on checking who is really on the phone.