What can go wrong with an AI receptionist for a plumbing business

We sell an AI receptionist, so you might expect this article to say nothing can go wrong. Several things can. You should know them before you buy one, from us or anyone else.

A warning on the evidence first. I found no independent test of AI receptionists on UK trade phone lines. Most surveys on AI customer service are paid for by companies with something to sell. So this is what the wider evidence suggests, plus my judgement.

Problem 1: Some callers don't want to talk to AI

This is the main one, and the evidence is fairly consistent.

Neither Gartner nor AnswerConnect says which countries it covered, and Gartner's survey pre-dates today's voice AI.

The hang-up figure is what people say they'd do, not what they did, and it comes from a firm that sells human answering. I couldn't find any independent data on what callers really do.

Set against that, Ofcom has reportedly seen "limited evidence" of AI causing widespread harm to telecoms customers. And the fair comparison isn't AI against a friendly person. For a plumber on a job, it's AI against nobody. In the YouGov survey for Answer4u, only 6% of UK adults said they'd leave a voicemail.

How to reduce it:

• Say it's an AI at the start of the call. In a Five9 survey, 76% of UK respondents said being told is very or extremely important. That's vendor research, but it matches common sense.
• Keep the greeting short, and get to the problem quickly.
• Offer a route to you: "I can take your details and Dave will ring you back by 4."
• Listen to your first 20 calls if your provider records them, or read the summaries. If calls are recorded, tell callers.

Problem 2: It mishears names, postcodes and accents

Addresses are the worst place for an error, because a plumber at the wrong house has lost an hour.

I found no independent figure for AI receptionists mishearing addresses. There's a general warning, though. In a preprint, researchers testing speech recognition on Newcastle English recorded a word error rate of about 32% for one system's UK English setting, and they cite earlier work finding higher error rates for speakers from Northern England and Northern Ireland than for Southern British English. That study used recorded speech, not phone calls, and it doesn't measure AI receptionists. Treat it as a reason to test with your own customers.

How to reduce it:

• Have the AI read the address back.
• Send a text with the details, so the customer can correct them.
• Test it with people who sound like your customers, on a bad mobile signal.
• Get a call summary and the caller's number every time, so you can always ring back.

Problem 3: It gives a wrong answer

An AI can state something false with total confidence. And you can't necessarily blame the machine.

In a Canadian case, Moffatt v Air Canada, a tribunal held the airline responsible for wrong information from its website chatbot, and rejected the argument that the bot was a separate entity. That's Canadian law and doesn't bind a UK court, but I wouldn't expect "the AI said it" to make a good defence here either.

The risky areas for a plumber are:

• Prices. "We charge £50 for a boiler repair" is a promise you didn't make.
• Promises about time. "Someone will be there in an hour."
• Scope. "Yes, we do gas work", if you don't.
• Advice. Telling someone how to fix something could make it worse.

How to reduce it:

• Give it a written list of what it may say about prices, usually a range or "Dave will confirm".
• Tell it what you don't do.
• Ban it from giving repair advice, except for agreed safety lines.
• Review summaries each week, and correct anything that's off.

Problem 4: Real emergencies

A burst pipe at 2am can't wait for a summary in the morning, and a smell of gas needs a different answer altogether.

• For water emergencies, set rules for what counts as urgent and how it reaches you.
• For a smell of gas, it should give a fixed script based on the gas network's advice: don't touch switches, don't use flames or smoke, open doors and windows, turn the gas off at the meter if that's safe, stay out of any cellar or basement, and ring the National Gas Emergency Service on 0800 111 999 (0800 002 001 in Northern Ireland). LPG customers should use their supplier's emergency number. If someone's unwell or in immediate danger, they should ring 999.
• For suspected carbon monoxide, it should tell people to turn off the appliance if they can, get into fresh air, stay out, and ring 0800 111 999. Anyone with trouble breathing, confusion or chest pain, or who collapses, needs 999.
• Get your provider to confirm it gives these scripts every time, and test them yourself.
• Say what it'll do outside hours if you don't take emergency calls. A clear "we can't attend tonight" is kinder than silence.

Test this before you go live by ringing it yourself and saying the words a panicking customer would use.

Problem 5: Privacy and call recording

Lots of AI tools record or transcribe calls. That's personal information, so UK data protection applies.

• The ICO's guidance says you must tell callers if you're recording a call and why. That guidance sits in its section on monitoring workers and is under review, but I'd follow it.
• Other privacy details can go in a link, or on your website.
• Your privacy notice should say how AI is used, and if a third party handles transcripts.
• People can ask for recordings that hold their information.

Providers differ. ServiceM8 logs a recording. Fasthosts says its AI receptionist doesn't record calls, and only keeps a summary and basic details. Ours can record calls, so the greeting tells callers. Ask any provider what's recorded and how long it's kept.

I found no UK rule that specifically requires a business to say a caller is talking to an AI on an inbound call. But I couldn't open the latest parliamentary briefings, so check this, and tell people anyway. Rules on outbound marketing calls are much stricter: the ICO says automated marketing calls need prior consent. Don't use an AI to cold call.

Problem 6: The bill grows

Many AI receptionists charge by call or by usage. Spam calls can use up your allowance, and a busy month can cost more than you planned.

• Check what counts as a call, and if spam is charged.
• Ask what happens when you go over.
• Watch the first three months closely.

Our plans include a monthly usage credit, and any extra goes on your bill.

Problem 7: It takes details, and then nobody acts

This is the most common failure I'd expect, and it has nothing to do with AI. A receptionist that captures 20 leads is worthless if you don't ring them back.

So decide how you'll handle it before you switch it on. Set a daily time for call-backs. Add the follow-up messages from our text templates. And fix the rest of the chain, as set out in why plumbers lose work after the first call.

How ours deals with these

Here's what ours does, so you can compare:

• It says it's an AI at the start of the call.
• It answers questions from what you've told it about your services, areas, prices and hours.
• It takes the caller's details, and you get a summary.
• On its own, it doesn't book. You call back and book. The full AI Business Engine adds booking.
• It can record calls, and the greeting says so.
• You choose how emergencies reach you. You can have a separate emergency number that comes straight to you, and an enquiry number the AI answers.
• On a number the AI answers, it only knows a call is urgent if the caller says so. Then it tries to transfer the call to you. If you can't pick up, it takes their details and puts them in your summary.
• If a caller says they can smell gas or mentions carbon monoxide, it gives the safety steps above and tells them to ring the emergency number straight away.

Ours can't remove these risks, and neither can anyone else's. If you'd like to hear whether it's worth it for you, read our honest view.

David Bell, founder of Real Business Systems

David helps small UK businesses get found, answer every enquiry and keep customers coming back. We sell some of the services mentioned here, and we say so where they come up.