What is an AI receptionist? It is software that answers your business phone and talks with callers in a synthetic voice. It can be set up to do a narrow job: answering from information you have approved, booking, taking details, or putting the caller through to a person.
Say you are halfway through a haircut and the phone rings, with nobody else free. On a divert, an AI receptionist can be set up to answer instead. It says it is an automated assistant, offers a time from your diary, and books it.
TL;DR
An AI receptionist listens to the caller, works out a reply from instructions and material you have approved, and speaks it in a synthetic voice. It can be set up to book, take details and hand over to a person, but it can mishear, and a language model can make up answers. So ask what it may say, who it hands over to, and what happens when a step fails.
Written on 3 October 2026 from UK government guidance and platform documentation. It describes how these products are documented to work, not a record of results.
What is an AI receptionist?
We found no official UK definition. One platform's documentation describes its version as "an automated voice assistant that communicates with callers using the instructions, business information, and actions you configure". Our AI integration page calls it a voice agent.
The nearest official framing is in the Competition and Markets Authority's (CMA) research on AI agents (9 March 2026). It notes that "Definitions vary", but says AI agents "sense (perceive their environment), decide and act", unlike "today’s chatbots, which primarily generate responses". An AI receptionist that books a slot can be acting in that sense, not just replying.
It is not a phone menu, where callers "navigate a menu using keypad selections", or voicemail. Nor is it a person: the government's AI Playbook, written for the public sector, lists a "lack of critical thinking, personal experience and judgement" among generative AI's limits.
How a call flows through an AI receptionist, step by step
One call, from ring to summary, as the platforms document it.
1. The caller rings the number you already publish
Your line diverts the call to a number the assistant answers on, so your published number can stay the same. A divert can cover all calls, calls not answered in a set time, or calls while the line is engaged. One platform also lets the assistant step in only when nobody answers, and work all hours or to a set schedule.
What can go wrong. If your own voicemail picks up first, the assistant never gets the call. Hours, routing and timeout settings also decide whether it answers. Your phone provider may charge for the diverted part of each call.
Ask: Which calls does it get, and what if my voicemail answers first?
2. It answers with the words you gave it
The greeting is a setting you write.
What can go wrong. Whether it says it is automated depends on how it is set up. The CMA's guidance on AI agents says "if the fact they are dealing with AI rather than a person might affect people’s decisions then you should tell them". Our build standard: it says it is an automated assistant in its first sentence, and its opening line offers a person and says if the call is recorded.
Ask: Can I hear and approve the exact opening line?
3. Speech recognition turns the caller's words into text
Depending on the product, one model may do steps 3 to 5: one platform's speech-to-speech mode can "hear audio and respond with audio". The assistant also has to judge when the caller has finished, and what to do if they talk over it.
What can go wrong. One speech provider says its base model "might not be sufficient if the audio contains ambient noise or includes industry and domain-specific jargon". Some platforms offer settings for specialist terms and names; one also has a setting for the pauses while a caller reads out a number. If the speech service fails with no fallback set up, one platform says "your call will end with an error".
Ask: Does it turn speech into text first, or work speech to speech? How does it handle names and numbers, and does it read them back?
4. It works out what to say
Some products follow a conversation flow you design. Others have a large language model (LLM) work out each reply from your instructions and a knowledge base: your hours, services, published prices and policies. GOV.UK Chat, a government pilot chat tool, works this way; its team preferred it to fine-tuning (further training) a model "because GOV.UK content is updated regularly". So a changed price is an edit, not a retraining.
What can go wrong. The Playbook warns that these models can create content that "appears plausible but may actually be factually incorrect". On one platform, the knowledge base is searched only when the caller's words match a trigger you set. Its advice: "Tell the agent not to guess when required information is unavailable."
Ask: Does it follow a flow or write each reply? Where do its answers come from, and what if the answer is not there?
5. A synthetic voice speaks the reply
Text-to-speech turns the reply into audio, in a voice you choose from the product's catalogue.
What can go wrong. Numbers and abbreviations may not sound right. One platform advises "writing out numbers as words or spelling out abbreviations". If the voice is convincing, the step 2 disclosure matters more.
Ask: Can I hear it say my prices, dates and street names?
6. It books, saves or sends
On one platform, it can be set up to book into a connected calendar, following that calendar's rules on "availability, buffers, minimum notice, and conflict settings". It can also save details, send a text or start a follow-up.
What can go wrong. An action can fire at the wrong moment or write to the wrong place. One platform's launch checklist asks you to confirm that "Each action runs only under its intended conditions".
Ask: What can it change, and can I see a test booking land?
7. It hands the caller to a person
A transfer sends the caller to a person's phone number. Our build standard is a transfer during opening hours and a callback request outside them.
What can go wrong. Does the person hear a summary before the caller is put through? One platform offers that in a transfer mode it labels experimental. If the person is busy or unreachable, on that platform "the fallback plan determines what the customer hears and whether the call ends"; elsewhere, it depends on the product. One platform's browser test skips transfers, so test them from a real phone.
Ask: Whose phone does it ring, and what happens if nobody answers?
8. A summary arrives after the call
The call log can hold a summary, a transcript and, if calls are recorded, a recording.
What can go wrong. One platform warns: "Transcripts may be incomplete if audio quality was poor, the call ended early, or processing was interrupted." The summary is written by the AI, so check the transcript or recording before relying on a detail. Depending on the product, the caller's voice and the summary may pass to other AI companies. The CMA's advice: "Make sure there is a human in the loop".
Ask: Who reads the calls, how long is it all kept, and which companies process the audio?
What an AI receptionist can be set up to do
The documentation we read describes products that can be set up to:
- answer questions from approved information
- book into a connected calendar
- save caller details to a customer record
- send a text during the call
- put the caller through to a person
- start a follow-up after the call
- work all hours, set hours, or only when nobody answers
Some products can also make calls out, a different job with different rules. Our build standard is inbound by default.
Our AI integration service offers a voice agent as an add-on, "on a divert from the number you already publish", answering "from copy you have approved".
What not to trust it with
Anything it has not been given. The AI Playbook, written for the public sector, is blunt: "Fundamentally, generative AI models cannot be trusted to produce factual content."
Professional advice. "LLMs are not domain experts. They are not a substitute for professional advice," the Playbook adds. In a clinic or practice, our standard is that it declines treatment questions and hands them to a person.
Upset callers, complaints and emergencies. One platform's guidance is to transfer to a person when a sensitive issue needs human judgement. The Playbook says AI "should not be used on its own in high-risk areas", such as those that could harm someone's health or safety.
The CMA's guidance sums it up: "you are responsible for what an AI agent does in the same way you are responsible for what an employee does", even if someone else provides it. Our standard follows: "Treat an answer your assistant gives a member of the public as your own marketing copy."
Where it fits, and where a cheaper fix is better
| Your calls | Fit | Why |
|---|---|---|
| Questions with a settled answer: hours, parking, prices | Good fit | It can be set up to answer from approved material |
| Bookings into a calendar with clear rules | Good fit | Booking can follow the calendar's own rules |
| Calls that ring out after hours or when every line is busy | Good fit, with a person or a callback to hand over to | The caller has somewhere to go when it cannot help |
| Answers nobody has written down and approved | Not yet | It needs approved information first |
| Treatment, clinical or legal questions | Poor fit: hand to a person | It is not a substitute for professional advice |
| Upset callers, complaints and emergencies | Poor fit: hand to a person | These need human judgement |
| A handful of missed calls a month | Poor fit: try a cheaper fix first | The build can take years to pay back |
| Calls answered, but bookings still do not happen | Poor fit: fix the booking process | The leak is after the call, not in it |
Compare it with the fixes that have no model in them: a divert of unanswered calls to a mobile someone carries, a simple phone menu, a booking link on your website and listings, an automatic text after a missed call, or an out-of-hours greeting that tells callers what to do next. Our AI integration page asks: "Does the plain fix cost less and break in fewer places?" Even the government's AI Playbook asks its readers to be open to the conclusion that "sometimes, AI is not the best solution for your problem".
Questions to ask before you choose one
Take the eight step questions above to any demo, and add three more:
- Can I test every path, including the transfer, from a real phone before it goes live?
- How fast is a wrong answer fixed?
- Can it make calls out, and can that be switched off?
The law and the money, briefly
Answering the calls your customers make is the clean case: no UK law we read stops an AI from doing it. The rules turn on what else it does: taking caller details, recording (callers must be told the call is recorded, and why), keeping transcripts, hearing health details and ringing people. On the cautious reading, marketing calls by an AI voice need prior consent that specifically covers automated calls. See our guide to whether AI receptionists are legal in the UK. This is a pointer, not legal advice.
The costs are a one-off build, usage that depends on your own calls, and the ongoing work of reviewing what it says. Our guide on how to work out whether a voice agent pays for itself does the sums.
Where to start
- Find where calls are lost. Note the hours, days and kinds of call. Then work out whether a voice agent pays for itself.
- Write down what it may say. List the answers you would approve, when a person takes over, and the subjects it must never attempt. Then read what the law asks of an AI receptionist.
- Talk it through. Our AI integration service can add "an inbound voice agent built to your own systems and answer sets". Get in touch.