How to Create an AI Receptionist for a Small Business
A missed call is a lost job. Not always, but often enough that for most trades and service businesses it’s the largest single leak in the operation — and unlike marketing, fixing it doesn’t require finding new customers, just answering the ones already ringing.
An AI receptionist answers every call, gathers the details, books straightforward jobs and passes on the rest. Setting one up takes an afternoon. Setting one up that doesn’t annoy people takes a bit more thought, because a phone call is higher-stakes than a website chat: people ring when something is urgent or complicated.
This is the full build, including the two things trades in particular must get right — emergency handling and the legal position on recorded calls.
Step 1: Work out what missed calls are costing you
The number that decides whether this is worth it
Do this before shopping for a tool, because it either justifies the subscription immediately or tells you not to bother.
For one week, log every missed or unreturned call — your phone already has the data. Then apply three figures you know: how many missed calls were genuine enquiries, what proportion of enquiries you normally convert, and what an average job is worth to you in profit.
Multiply it out and annualise. Most trades doing this exercise find the answer runs to thousands, and it’s usually the strongest business case for anything on this site. If the answer is small — because you rarely miss calls, or because most callers ring back — then say so honestly and skip the rest of this guide.
Step 2: Decide what it’s actually allowed to do
Four levels, and most businesses want the second
How much you let it handle
| 1. Take a message | Answers, identifies the caller and what they want, promises a callback. Simplest, lowest risk, and already recovers most of the lost value. |
| 2. Qualify and route | Also asks the questions you’d ask — what’s the job, where, when, is it urgent — so you ring back informed. Usually the sweet spot. |
| 3. Book appointments | Writes into your live calendar. Genuinely useful, but only if your availability is genuinely reliable and the job types are standard. |
| 4. Quote and take payment | Don’t, for a small business. Pricing needs judgement and taking payment over an automated call adds risk you don’t need. |
Start at level one or two for the first month. You can always widen the remit once you’ve read a fortnight of transcripts and know what callers actually ask — widening is easy, and recovering from an assistant that booked jobs wrongly is not.
Step 3: Write the instructions
The block that goes into the configuration
Every AI answering service has a setup where you describe your business and its behaviour. Most people write three lines. This is what it should contain.
FIRST THING YOU SAY
“Good [morning/afternoon], [business name]. You’re speaking to an automated assistant — I can take your details and get [owner name] to call you back. How can I help?”
WHAT YOU DO ON EVERY CALL
– Get their name, number, and the postcode or area of the job
– Get a plain description of what they need
– Ask whether it’s urgent
– Confirm the number back to them, digit by digit
– Tell them when someone will actually call back: [your real timescale]
WHAT YOU KNOW
– We do: [list services]
– We do NOT do: [list — so you never accept work we can’t deliver]
– Areas covered: [list by name]
– Typical prices: [only the ones you’re happy stated on a call]
– Normal hours: [times]
YOU MUST NEVER
– Give a price for anything not on the list above
– Estimate how long a job will take or when we could attend
– Promise a same-day or emergency visit
– Discuss another customer’s job
– Give any advice about a technical, medical or safety problem
– Argue with anyone or defend the business against a complaint
IF THEY DESCRIBE AN EMERGENCY
[Insert your emergency script — see Step 4. This must come before everything else.]
IF THEY ASK FOR A PERSON
Take their details immediately and say someone will call back within [timescale]. Do not ask further questions first.
TONE
Calm, brief, practical. No enthusiasm, no “absolutely”, no “fantastic”. Say “I’ll pass that straight on” rather than “I’d be delighted to assist you today”.
Step 4: Write the emergency script before you go live
The step that matters most, and takes ten minutes
If you’re a plumber, a gas engineer, an electrician, a locksmith, a vet or anything with a genuine emergency category, this is non-negotiable. An automated assistant cheerfully taking details while someone describes a gas smell or water coming through a ceiling is a serious failure.
Write out, explicitly: the words and situations that constitute an emergency in your trade; exactly what the assistant should say when it hears them; and where it should direct the caller — an emergency number, your mobile, or the relevant national service.
For gas, the assistant should tell the caller to leave the property and ring the National Gas Emergency Service on 0800 111 999. For anything involving injury or fire, 999. For a burst main, the water company. Put the specific numbers in the configuration rather than leaving it to interpret.
Then test it. Ring your own line and describe an emergency in a few different ways, including vaguely — “there’s a funny smell” as well as “I can smell gas”. If it doesn’t escalate correctly on every attempt, don’t launch.
Recording and monitoring calls. If your provider records or transcribes calls — most do, and that’s how you get transcripts — callers must be informed, and you need a lawful basis under UK data protection law for holding those recordings. In practice: state it in the opening greeting, keep recordings only as long as you actually need them, set a retention period rather than keeping everything indefinitely, and be able to delete a caller’s data on request. Check whether your provider will sign a data-processing agreement.
Be honest that it’s automated. Say so in the first sentence. People generally accept an automated assistant that takes their details efficiently; they react badly to discovering mid-conversation that the warm, chatty “receptionist” was synthetic — and that reaction lands on your business, not the vendor’s.
Step 5: Set the greeting and pick the voice sensibly
Plain beats impressive
Two practical decisions. The greeting should name the business, disclose that it’s automated, and say what will happen — all inside about ten seconds. Callers on a mobile with a problem have no patience for a preamble.
The voice should be clear and unremarkable rather than characterful. Choose a plain accent your customers will find easy, keep the pace slightly slower than natural, and resist anything that sounds like a marketing video. If your provider offers a regional voice matching your area, that’s usually worth taking — it reads as local rather than as a call centre.
Do not clone your own voice for this. It sounds like you, which means callers reasonably believe they’re speaking to you, and that’s a deception you’ll have to unpick every time it comes up.
Step 6: Ring your own number, repeatedly
Twenty test calls before a single real one
Test from a phone that isn’t yours, ideally over a mobile connection with some background noise, because that’s how real calls arrive.
Work through all of these: a straightforward enquiry; a request for a price you haven’t listed; a job outside your area; a service you don’t offer; an emergency, described three different ways; an angry caller complaining about existing work; someone asking to speak to you directly; someone with a strong accent or speaking quickly; a caller who gives their number too fast; someone asking a technical question; and a caller who says almost nothing and waits.
Check the outcome of each: did the details reach you accurately, was the number captured correctly, did it escalate when it should, and did it avoid committing to anything?
The single most common real-world fault is a mis-transcribed phone number. Confirming the number back digit by digit — in the instructions above — fixes most of it, and it’s worth verifying that it does.
Step 7: Read the transcripts, then adjust
A fortnight of reading, then monthly
Read every transcript for the first two weeks. You’re looking for questions it couldn’t answer, points where callers got frustrated or hung up, and anything it said that you wouldn’t have.
Then use the transcripts as free market research, which is the underrated benefit here. The questions people ask on the phone are frequently different from what they’d email, and they tell you exactly what your website should explain better.
Tell me:
1. Questions it couldn’t answer — these need adding to its knowledge
2. Any point where a caller sounded confused or frustrated, and what caused it
3. Anything it said that it shouldn’t have — a price, a timescale, a commitment
4. Any call that should have been escalated as urgent and wasn’t
5. The questions asked most often, which I should answer on my website
6. Whether callers seem to be hanging up at a particular point
Be specific and quote the wording. Flag point 4 first if it happened.
Before you go live
- The greeting discloses it’s automated in the first sentence.
- Emergency script written and tested with three different phrasings.
- Specific emergency numbers in the configuration, not left to interpretation.
- It refuses prices, timescales and attendance promises — tested.
- “Speak to a person” works instantly, with no qualifying questions.
- Numbers confirmed back digit by digit — tested on a fast talker.
- Call recording disclosed, retention period set, DPA in place.
- Your real callback time stated, and one you’ll actually meet.
- Twenty test calls done from a phone that isn’t yours.
Frequently asked questions
Will customers hang up on an AI receptionist?
Some will, and the honest comparison isn’t against a human answering — it’s against your voicemail or an unanswered ring, which loses the call entirely. Against that baseline, an assistant that takes details reliably is a clear improvement, and most callers with a practical problem will happily give their name and number to something that sounds organised. Where you lose people is when the assistant is slow, chatty, or tries to handle something it can’t. Three things reduce hang-ups measurably: disclose it’s automated immediately so nobody feels tricked, get to “how can I help” within about ten seconds, and make the callback promise specific rather than vague. If you’re anxious about it, a reasonable middle path is to route calls to the assistant only when you genuinely can’t answer — out of hours, or after four rings — so anyone who catches you live still gets you.
Should it book jobs straight into my diary?
Only once you trust it, and only for genuinely standard work. Direct booking is the biggest time-saver available here, but it’s also where the mistakes hurt: a job booked in a slot you can’t make, two jobs an hour apart at opposite ends of your area, or a job accepted that turns out to be twice the size described. Sensible progression: run message-taking or qualification for a month, read the transcripts, and only then enable booking for your most standardised job types. When you do, set hard guardrails — buffer time between appointments, a maximum number per day, no bookings within a certain notice period, and geographic sanity limits if your provider supports them. Keep anything non-standard, large or unusual as a callback. The rule of thumb: if you’d want to ask a question before committing to the slot, the assistant shouldn’t be committing either.
What happens if it mishears a phone number?
You lose the job, and this is the most common practical failure rather than anything dramatic. Phone numbers are the weak point because they’re spoken quickly, often from memory, frequently on a poor connection. Three defences, all worth implementing. First, have the assistant read the number back digit by digit and ask the caller to confirm — that alone catches most errors and takes eight seconds. Second, if your provider captures the incoming caller ID, use that as a cross-check against what was dictated and flag any mismatch to yourself. Third, ask for a second contact method where the call warrants it. When you review transcripts, specifically check numbers against caller ID for a sample, because a systematic transcription problem is worth knowing about early. If your provider handles this badly and won’t fix it, that’s a legitimate reason to change provider.
Is an AI receptionist better than a human answering service?
It depends on your call pattern and what you need from the call. AI wins on cost at moderate to high volume, on consistency, on availability at three in the morning, and on producing a searchable transcript of every call. A human service wins on handling anything unusual, on genuine judgement about urgency, on callers who are distressed or confused, and on the impression it creates for higher-value work — if your average job is worth several thousand pounds, a real person answering may pay for itself in conversion alone. Cost comparison depends on structure: human services often charge per call or per minute, which is cheaper at low volume and more expensive as calls increase, while AI is usually a flat monthly fee. Some businesses run both — AI out of hours, a human during the day. Work out your volume first, then compare on the same basis rather than headline price.
The bottom line
Count what missed calls cost you before buying anything — for most trades that figure alone settles it, and it’s the strongest business case going. Then start narrow: take messages and qualify, rather than booking jobs or quoting prices, and widen only after you’ve read a fortnight of transcripts. Write the emergency script before you go live and test it with vague phrasing as well as obvious, because that’s the failure that genuinely matters. Disclose that it’s automated in the first sentence, confirm numbers digit by digit, and make sure “I’d like to speak to a person” works instantly. Then read the transcripts — they’ll tell you what your website should have said.
Pricing is indicative and set by each provider — always confirm current pricing and what’s included. Recording or transcribing calls means processing personal data: callers must be informed, you need a lawful basis, and you should set a retention period and obtain a data-processing agreement. Emergency numbers and procedures should be verified against current official guidance for your trade. Nothing here is legal or data protection advice.