What it is, and what it is not
An AI receptionist is not an interactive menu with better recorded prompts. There are no numbered options and no fixed tree. The caller speaks normally, the system understands what was asked, retrieves the answer from your own material, and either completes the task or hands the call to a person. It answers on the first ring, at three in the morning, and it can hold twenty of those conversations at once without a queue forming.
It is also not a replacement for a front of house team in any business where the call itself is the service. The correct way to think about it is as cover: it takes the calls nobody is currently taking, and the routine ones that consume a trained person's day.
What happens on a call, step by step
It answers and identifies itself
On the first ring, with your greeting. It should confirm it is an automated assistant when asked, and in several markets there is a regulatory expectation that it does.
It works out what the caller wants
From what the caller says rather than from a menu. A caller who opens with three things at once is a normal case, and the system should hold all three rather than take the first.
It answers from your own material
Opening hours, services, prices, policies, availability, order status. The source is your documents and systems, not general knowledge, and it is configured to say it does not know rather than to guess.
It takes the action
Books the appointment into the live calendar, raises the ticket, updates the record, sends the confirmation. This is the step that separates a receptionist from an answering service.
It escalates what it should not handle
A warm transfer where somebody is available, a call-back booked into a real diary where nobody is, and a written summary either way.
It leaves a record
A transcript, the outcome, and the data written into your CRM before the next call starts. In practice this is often the largest single gain, because a human front desk rarely has time to write everything down.
What it handles well, and what should always reach a person
The right split is not a matter of technical capability. Several of the items in the second column are perfectly within the system's ability and belong to a person anyway, because the consequence of getting them wrong is borne by the caller.
| Handles well | Should always reach a person |
|---|---|
| Opening hours, location, parking, directions | A distressed, vulnerable or unwell caller |
| Availability and booking an appointment | A complaint, at the moment it becomes a complaint |
| Rescheduling and cancellations | Anything clinical, legal or financial that amounts to advice |
| Order, delivery and job status from your systems | Negotiating a price, a discount or a contract term |
| Answering questions from your published material | Safeguarding concerns and anything involving a child at risk |
| Qualifying a new enquiry and routing it | Card details and any payment taken over the telephone |
| Taking a message with the detail actually captured | An emergency, which should transfer immediately with no questions |
| Out of hours and overflow when every line is busy | A caller who asks for a person, first time of asking |
Where they fail
A balanced account has to include this, because these are the cases that decide whether callers accept the system.
- Poor line quality. A weak mobile signal or heavy background noise degrades recognition far more than anything about the caller.
- Strings of characters spoken aloud. Reference numbers, registration plates and unusual surnames are the most common source of a wrong record, which is why they should be confirmed back and, where possible, sent by message instead.
- Interruption. Callers talk over automated systems, and a receptionist that cannot be interrupted mid-sentence feels worse than a menu.
- The unanswerable question. If the material it was given does not contain the answer, the correct behaviour is to say so and transfer, and a system that instead produces a plausible guess is worse than no system.
- Emotional context. It can detect frustration in the words used. It cannot judge when a caller needs somebody to simply listen, which is exactly why the escalation rules matter more than the conversation quality.
What it costs and what it replaces
Beyond the build, an AI receptionist costs telephony plus model usage, which for typical small business volume lands in the region of USD 50 to USD 300 a month. Comparing that against a salary is the wrong comparison in most cases, because the honest alternative is usually not a second receptionist but the calls that currently go unanswered.
| What it replaces | Typical position before | After |
|---|---|---|
| Out of hours calls | Voicemail, or an unanswered ring | Answered and booked, with the record written up |
| Overflow at peak | A queue, then abandoned calls | Every line answered simultaneously |
| Routine enquiries in office hours | A trained person answering the same question repeatedly | Handled unaided, with the exceptions passed on |
| Call notes | Written up later, or not at all | A transcript and a structured record on every call |
Our own figure, marked as ours
A first receptionist build typically removes in the region of forty hours of telephone cover a month. That is a figure typical of a first build rather than a measured client result, and it varies enormously with call volume and with how much of your traffic arrives outside working hours.
Six things to settle before you buy one
- Where the escalation goes at two in the morning, and what the caller is told.
- Which numbers it answers, and whether it takes overflow only or every call.
- Who keeps the source material current, because accuracy degrades the moment prices or hours change and nobody updates the documents.
- Whether calls are recorded and transcribed, what the retention period is, and where that data is stored.
- What it says when it is asked whether it is a person.
- Who receives the alert if it stops answering, which is a question most buyers ask only after the first outage.
Frequently asked questions
Will callers know they are talking to a machine?
Most will, and the system should say so when asked. Disclosing it early tends to improve the call rather than harm it, because the caller adjusts how they speak. Attempting to pass a voice agent off as a person creates a regulatory problem in several markets and an avoidable trust problem in all of them.
What happens when the AI receptionist cannot help?
It transfers. A correctly configured agent has a named fallback for every category it cannot complete, whether that is a warm transfer to a colleague, a call-back booked into a diary, or a message with a transcript sent to the right inbox. A dead end is a configuration fault, not a limitation of the technology.
Can it take payment details over the phone?
It should not. Card details on a recorded and transcribed call create a compliance burden most organisations do not want. The correct pattern is for the agent to send a payment link by message during or immediately after the call, so the card details never enter the voice channel.
How does it know the answers to questions about our business?
It is given your own material: price lists, service pages, opening hours, policies and internal notes. It answers from that source rather than from general knowledge, and it is configured to say it does not know rather than to guess. Keeping that source current is the single largest factor in how accurate it stays.
What does an AI receptionist cost to run?
Beyond the build, the running cost is telephony plus model usage, which for typical small business volume lands in the region of USD 50 to USD 300 a month. The comparison is not with a salary alone but with the enquiries currently lost after hours and while the line is engaged.
Related answers
Disclosure, and the only sales pitch on this page
This page is published by a company that sells the thing it describes.
Oxford Crown Technologies builds voice and operations agents for organisations headquartered in London and across the Gulf. We publish these pages as reference material, including market figures that are not ours and that do not always favour us, because a buyer who understands the range negotiates better with everybody, including us. Leadership holding postgraduate degrees from the University of Manchester and the University of Oxford.
Our own builds are fixed fee, live in 14 days, built on accounts in your name, and carry a 28 day money back guarantee. We prove Return on Investment, or you do not pay.
Reviewed 25 July 2026. Market figures change; where this page quotes a range from a third party, the source is named above. Figures described as ours are our published fees or typical results for a first build, not measured results for a named client.