AI receptionist: what it does, how it works, what it costs

An AI receptionist answers your business phone in a natural voice, handles routine requests such as opening hours, availability and simple bookings, takes a structured message when it can't finish, and passes anything that needs judgement to a person. It earns its place when your calls are repetitive and your information is kept current. It runs on a monthly plan plus usage, and what that comes to depends on your call volume, the hours you want covered and the systems it books into.

What an AI receptionist does

Picture a Tuesday at 12.40pm. Two of the team are at lunch, the third is with a client, and the phone rings for the fourth time in ten minutes. An AI receptionist picks up on the first ring, finds out what the caller wants, and either deals with it or gets it to the right person with the details already written down.

It can sit in three places on your line. As the main answer point for every call. As overflow, catching whatever your team can't reach within a few rings. Or out of hours only, once the office has emptied. The last two are the low-risk way in: the daytime team keeps the calls it handles well, and the AI picks up the ones that currently ring out.

The same product goes by several names. An AI answering service or AI phone answering service usually means exactly this. A digital receptionist tends to mean the same system answering website chat and messages as well as calls, from one knowledge base, so a caller and a web visitor get the same answer and a booking started in chat at midnight can be finished by phone the next morning. The phone is usually where the value starts. A call that rings out leaves nothing behind, whereas a chat message at least waits in an inbox.

What happens on a call

A well-built receptionist follows the same path every time. The branching is a small agentic workflow, and each step is something you can test.

1. Answer at once. No ring-out and no hold queue, however many people call at the same moment.
2. Say what it is. A short greeting telling the caller they're speaking to an automated assistant and, if calls are recorded, that they are and why. The UK guide sets out what the ICO expects here.
3. Work out the request. Opening hours, a booking, a price question, an existing order, a complaint.
4. Handle the routine end to end. Answer from approved information, check availability, book or move an appointment.
5. Capture what it can't finish. Name, number, reason for calling and urgency, read back to the caller so mistakes are caught while they're still on the line.
6. Hand over when it should. Transfer the live call with a summary, or flag a structured message for a callback, and tell the caller which is happening.
7. Write it down. Update the calendar or CRM and send the team a summary, so nobody has to ask the caller to repeat themselves.

Speed matters more than the voice. A call feels robotic when there's a pause before every answer, not because the voice is synthetic, so test the response time on a real phone line rather than in a browser demo.

What it can handle, and what it can't

The dividing line is judgement. A known answer and a clear next step? That suits the AI. Anything that turns on emotion, risk or a situation nobody wrote down belongs with a person.

It has practical weak spots too. Heavy background noise, unusual names, a poor mobile signal and a caller who changes their mind halfway through are where transcription slips. It can't greet someone walking through your door either, so a physical front desk still needs a person wherever presence matters. The same split applies to any AI agent, as AI agents for business explains: automate the routine, keep judgement with people, and build the handover properly.

Set the rules before it answers

A badly briefed AI receptionist fails as surely as a badly built one. Before it takes a live call, write down:

• the service descriptions, locations and hours it may quote;
• the questions it may answer, and the source each answer comes from;
• what it must collect for a callback to be useful;
• which requests go straight to a person, and to which person;
• what it says when the information it needs is missing;
• what happens if the call drops or a connected system is down.

Give one named person the job of keeping that information current. An out-of-date price or a branch that has closed turns a working system into a confident source of wrong answers. And be strict about promises. If the receptionist has only emailed a request to your team, it must not tell the caller the appointment is confirmed.

Bookings and records

Booking is where an AI receptionist can save the most staff time, and where careless wiring causes the most trouble. It needs one authoritative calendar, live availability, correct time zones and a rule for changes and cancellations. It should read the details back before committing, and it needs protection against duplicate bookings when a request is retried after a dropped connection.

A CRM link needs a rule for matching the caller to a record. Caller ID isn't identity. A phone number on its own shouldn't be enough to reveal account details, so when the match is uncertain the safe move is to take a message, flag it and let a person confirm who they're dealing with before anything is disclosed or changed. The integration section of our automation services guide works through the rest.

AI or a human answering service

A traditional answering service uses trained operators shared across many client businesses. The AI version answers everything at once and never has an off day. Here's where each is stronger.

Neither column wins outright. A human operator is better with a worried caller; the AI is faster and cheaper on the routine bulk. Per-call pricing is cheap when calls are few and expensive when they're many, so price both options against last month's actual call count, not the brochure. The blend we'd recommend for most firms is AI for routine and out-of-hours calls, with sensitive ones routed to a person, either your own team or a human service kept as backup. Vendors sell that hybrid as a package too.

An AI receptionist for a small business

For a firm of five or ten people, the real comparison isn't with a call centre. It's with the first receptionist you'd otherwise hire. At the legal pay floor that person's wages come to £24,785 for a 37.5-hour week, plus £2,968 of employer National Insurance and £556 of minimum pension.

One line in that sum often disappears. If your business qualifies, Employment Allowance takes up to £10,500 off the employer National Insurance bill, which wipes out the National Insurance on a first hire. So for many small employers the honest benchmark is wages plus pension, not the full total. Price the AI against that smaller number and against the hours a person would actually cover. A hire works 37.5 hours a week; the phone rings for more than that.

You'll also see the same product sold as an AI voice agent. That's the builder's name for the technology: software that holds a spoken conversation and acts on it. Answering the front-desk line is one job a voice agent can do. Outbound reminder calls and lead qualification are others, and they bring their own rules, so ask any vendor which job its agent was built and tested for.

What it costs to run

The monthly plan is one line of several. A full costing includes:

• set-up and configuration, including writing the knowledge and escalation rules;
• the phone number and any call forwarding;
• call minutes, with a per-minute rate once the plan's allowance runs out;
• speech processing and model usage, where they're billed separately;
• calendar, CRM and messaging integrations;
• someone's time to keep the information current and read transcripts.

Ask three questions before you sign. Are transferred calls billed twice? Do long calls or repeated attempts cost more? Can you set a spending cap and watch usage as it happens?

Then judge the result per useful outcome, not per answered call. The goal might be capturing enquiries after 6pm or freeing two people during the lunchtime rush, and the maths is the cost per task lens: what each captured enquiry or completed booking costs, before and after. Don't count every answered call as a new customer. For the other side of the ledger, what a UK receptionist costs to employ once National Insurance and pension are added, see the UK guide's hire cost table.

Test it before it goes live

Build a test set from real calls: background noise, interruptions, unfamiliar names, different accents, a vague request, a caller who changes their mind. Check that it asks a useful follow-up, repeats the important details and knows when to stop talking.

Then give it a call it shouldn't be able to handle and watch. A clean transfer or a proper message is a pass. A loop, a dead end or a confident wrong answer is a fail, whatever the demo looked like.

Measure completed tasks, incorrect commitments, successful handovers and the time your team spends fixing mistakes. Transcription accuracy is not the same as the right outcome. The lowest-risk way in is to cover overflow and out-of-hours calls first, which the after-hours and 24/7 guide sizes, and move the main line only once the transcripts justify it. Contact centres with teams of agents face different choices again, covered in the AI call centre guide.

Common questions

Before you speak to any vendor, pull last month's call log and mark three things: the calls nobody answered, the hours they came in, and how many were questions your team answers the same way every time. That list is your specification. The AI FAQs cover the common set-up questions. Then bring the log to an informal scoping chat, tell us roughly how many calls arrive in a month and which systems bookings go into, and we'll send a price range for the set-up. On the contact form, pick "AI agents".