How AI receptionists work - and what they cannot do

If you are going to sell AI receptionists to local businesses, you need to be able to explain what one is without waving your hands. Not because clients will interrogate you - most will not - but because the ones who do are usually the ones worth having, and because a client who has been oversold will find out on their own within a fortnight.

So here is the mechanism, end to end, and then the honest list of things it does badly.

What actually happens on a call

An AI receptionist is not one piece of software. It is four or five things chained together, and the chain runs on every single turn of the conversation. Understanding the chain is what lets you diagnose problems later.

1. The call arrives at a number you control

The business either forwards its existing number to a new number, or publishes the new number directly. Forwarding is by far the more common arrangement, because no local business wants to reprint its van livery. Most forward on a rule: after five rings, or when engaged, or outside opening hours. Some forward everything.

This first step is duller than the AI, and it is where most setups actually break. Forwarding is configured with the business's telephone provider, not with you, and providers differ in what they allow. Always test it with a real call from a real mobile before you tell a client they are live.

2. Speech is converted to text

The caller talks. Their audio is transcribed into text more or less as they speak. This is the part that has improved most in the last few years and it is genuinely good - but "good" is not "perfect", and the errors it makes are not random. It struggles with proper nouns, with strong regional accents, with two people talking at once, with road noise, and with anyone speaking from a building site or a moving vehicle. It will confidently render a street name as something that is not a street name.

3. A model decides what to say

The transcribed text goes to a language model that has been given a set of instructions and a body of knowledge about the business: what it does, where it covers, opening hours, what it charges for a call-out, what it will not take on, and how it should behave. That configured knowledge is the entire difference between an AI receptionist that sounds like it works there and one that sounds like a switchboard in another country.

The model also carries the conversation so far, so it can refer back to what the caller said thirty seconds ago. It decides what the reply should be, and - this is the commercially important bit - it decides what facts it has now collected.

4. The reply is spoken back

The text reply is synthesised into speech and played to the caller. Modern synthesised voices are convincing enough that a fair number of callers do not notice. They are not indistinguishable, and you should not build a pitch on the idea that nobody will ever be able to tell.

There is latency in all of this. Somewhere between half a second and a couple of seconds passes between the caller finishing a sentence and the reply starting. Humans do not usually pause that long. Good configurations fill the gap; poor ones feel like a bad video call.

5. Details are extracted and sent onwards

At the end of the call - sometimes during it - the system pulls structured fields out of the conversation. Name, number, postcode, job type, urgency, whatever you have configured it to look for. Those fields go somewhere useful: an email, a text message, a spreadsheet, or a system the business already uses.

This is the part clients actually value, though they will describe it as "answering the phone". What they are buying is that at 7pm on a Tuesday, a plumber gets a text saying Sarah, 07xxx, Didsbury, leaking boiler under the kitchen sink, wants someone tomorrow morning, instead of a voicemail notification he will look at on Thursday.

What it does well

Set expectations here and you will have far fewer difficult conversations later.

• Answering every call. It does not get busy, it does not go to lunch, and it does not mind the fourth call in ten minutes.
• Out of hours. Evenings, weekends, bank holidays. This is where most of the value sits for trades.
• Repeated questions. Opening hours, areas covered, whether you do commercial as well as domestic, whether there is a call-out charge, how long a quote takes. These make up a large share of calls to a small business and none of them need a person, so it is worth sorting the questions it can answer reliably from the ones it cannot.
• Capturing details consistently. A human receptionist has good days and bad days. An AI asks the same five questions every time, which makes the follow-up far more predictable.
• Triage. Separating "my ceiling is coming down" from "can I get a price for a new consumer unit next month" is genuinely useful, and it is a task language models are well suited to.

If you want to hear one in a normal conversation rather than read about it, the live demo is the fastest way to form your own opinion.

What it does badly

Be honest about this with clients, because they will find out anyway.

• Genuinely unusual calls. Anything far outside its configured knowledge is better handed to a person. The model will try to be helpful, and trying to be helpful with no information is exactly how you get a wrong answer delivered confidently.
• Emotional situations. An angry or distressed caller wants a human, and a cheerful AI can make it worse. A complaint call handled by an AI receptionist is usually a complaint call that has just got bigger.
• Heavy accents and poor lines. Speech recognition is good, not perfect. Building sites are noisy. Hands-free kits in vans are worse.
• Anything requiring judgement about money. Quoting a price for a non-standard job, agreeing a discount, deciding whether to waive a call-out fee. Configure it to take the details and say someone will come back with a price.
• Long, meandering calls. Elderly callers in particular often want to explain the whole history before getting to the point. An AI will stay polite indefinitely, but it will not always find the thread.

The disclosure question

Sooner or later a caller will ask directly: am I speaking to a real person?

The honest answer is no, and the system should be configured to give it. Not a deflection, not a change of subject - a straight answer, followed by an offer to take a message or get a person to ring back. Anything else is a decision you have made on your client's behalf to mislead their customers, and it is not a decision that will look good if it surfaces later.

An AI receptionist that admits what it is when asked loses almost nothing. One that dodges the question loses the caller's trust in the business, which is not yours to spend.

Things that depend on the provider, not on the AI

This trips up new agency owners more than anything else. Several of the features clients ask for first are not properties of the AI at all - they are properties of the voice platform underneath it and the way you have configured the account.

• Warm transfer to a mobile. Whether the AI can put a live caller through to a person mid-call, and how cleanly it does so, depends entirely on the provider and the plan.
• Booking straight into a calendar. Possible with some providers and some calendar systems, fiddly or unavailable with others. Never assume; check for the specific combination in front of you.
• Call recording and transcripts. Availability varies. And if recording is switched on, the business is responsible for informing callers and for complying with UK data protection law - that obligation sits with them, not with the software.
• Writing into a CRM. Depends on what the client uses and what it exposes. Some are trivial, some need a middle layer, some are effectively closed.

The safe habit is to say "that is usually possible, let me confirm it works with your setup before we promise it" - and then actually confirm it. There is more on how the pieces fit together in how it works, and the commonly asked versions of these questions are covered in the FAQ.

How to frame it to a client

An AI receptionist is not a replacement for a good human receptionist. It is a replacement for voicemail, and voicemail is what most local businesses are actually using. That is a much easier comparison to win, and a much more honest one to make.

A firm that already has a capable person answering the phone all day does not have a problem you are solving. A firm where the owner is up a ladder with his phone in the van does. Once you learn to tell the two apart in the first two minutes of a conversation, the whole job gets easier - and you stop wasting your time on the wrong prospects.

If you are weighing up whether this is a service line worth offering at all, the AI receptionist overview covers the commercial side, and what's included sets out what you get to work with.