Before you put an AI on your phone line, work out who it will fail

Healthwatch Rotherham published something in July that has since gone round the national press, and it is worth reading properly if you are anywhere near a decision about answering your phone with AI.

Patients at GP practices using an AI receptionist reported that it could not understand them. Not occasionally — enough for the local watchdog to write it up. The line that stayed with me was from one resident:

“I could never get it to understand me, I ended up just hanging up and not bothering to try and book an appointment.”

Read that as a business outcome rather than a health one. Somebody rang, could not be understood, and gave up. If that were your shop, you would never know it happened. There is no missed-call notification for a caller the system politely failed.

This is a known property, not a glitch

The instinct is to assume the system is early and will improve. It will improve — but the shape of the failure is structural, and worth understanding before you buy.

Speech recognition is at its best on the voices it heard most during training, and those corpora lean heavily on standard American and British English. Published research has repeatedly found meaningfully higher error rates for non-native accents than for the standard native ones, with accuracy degrading further the further a speaker sits from that centre. Strong regional accents, second-language speakers, stammers and other speech differences are all more likely to be misheard.

So “it struggles with a broad Yorkshire accent” is not a peculiarity of one product. It is the same effect that will show up on your line, in whatever accents happen to be furthest from the training data.

The uncomfortable part: it fails the wrong people

If the failures were spread evenly across your callers, this would be an annoyance. They are not.

The groups a voice system is most likely to mishear overlap heavily with the groups least able to route around it: older callers, people with speech impairments, people whose first language is not English, people who are not confident doing things online. Healthwatch Rotherham raised exactly that alongside the accent problem — digital exclusion, and accessibility for disabled patients.

Which makes the standard fallback — you can always use the website — close to useless. You have offered the alternative channel specifically to the people least likely to use it.

And in the UK there is a duty attached

Briefly, and as general information rather than legal advice.

Under the Equality Act 2010, the duty on service providers to make reasonable adjustments for disabled people is anticipatory. You are expected to think ahead about what adjustments people might need, rather than waiting until somebody is disadvantaged and complains. It applies to businesses providing services to the public, not only to public bodies.

Put a system on your phone line that a portion of disabled callers cannot use, with no easy alternative, and you have not simply made a service choice. Healthwatch's own point was the same one: practices keep their obligation to provide reasonable adjustments and other ways of getting in touch.

How to test one properly, in an afternoon

Ring it yourself, from a mobile, in the accents of the people who actually ring you. Then get five people who sound genuinely different to do the same — not five colleagues who all sound like you.

Add the real conditions. A van with the engine running. A shop floor. A hands-free kit. Most demos are recorded in a quiet room and most customers are not in one.

Try to break it on purpose. Mumble. Hesitate. Talk over it. Change your mind halfway through a sentence, the way people do.

Time the escape route. How many seconds, and how many steps, to reach a human? Do it as a caller who has already failed once and is getting annoyed.

Write down every failure and what came next. Silence, a loop, or a person. If the honest answer is a loop, it is not ready, whatever the demo showed.

The one setting that actually matters

An obvious, early, no-penalty route to a human being.

Not after three failed attempts. Not hidden behind “I didn’t catch that, let’s try again” on a loop. Offered plainly, near the start, in a way that does not make the caller feel they have failed a test. We wrote the general version of that argument in when an AI agent should call a human, and a phone line is the sharpest case of it: the person on the other end is already talking to you, and already trying.

What argues the other way

Honestly? Quite a lot, and I would be selling you something if I pretended otherwise.

The alternative in most small businesses is not a warm, capable receptionist. It is a phone that rings out at lunchtime, a voicemail nobody listens to, and a customer who rings the next name on the list. A system that answers every call at eight in the evening and hands the hard ones to a person on Monday may be a real improvement for most callers — and for the ones it does mishear, it is not obviously worse than nobody answering at all.

The developer in the Rotherham case says its system is designed to handle a range of dialects, that callers can reach a human at any point, and that feedback across practices nationally has been overwhelmingly positive — a figure of its own, so take it as the company's account rather than an independent finding. Both things can be true: most people fine, a minority shut out.

And the watchdog was not asking for the technology to be scrapped. It asked for adjustments and alternatives. That is a reasonable ask, and a cheap one.

The short version

A voice AI does not fail randomly — it fails the callers furthest from its training data, who are often the ones who most need to talk to a person. Test it with real voices in real conditions before it goes live, give everyone a fast and obvious way to reach a human, and remember that the caller who quietly hangs up never shows up in your numbers.

Sources: Healthwatch Rotherham, “Patient concern over AI receptionist ‘Emma’ — what you need to know”, published 10 July 2026, read directly — the resident’s quotation, the accent and speech-impediment concerns, the digital-exclusion point and the reasonable-adjustments obligation are all from that piece; national coverage followed in mid-August. The developer’s response is from that later press coverage and is the company’s own account. On speech recognition, the pattern of higher error rates for accents further from the standard English that dominates training data is a repeated finding across published evaluations rather than a single study. The Equality Act 2010 position — that the reasonable-adjustments duty on service providers is anticipatory and applies to private as well as public providers — is long-established; this post is general information and not legal advice.