What AI Receptionists Still Get Wrong

A clinic manager sent me a call recording in August. Her AI receptionist had answered on the second ring, taken the caller's details, and dropped the appointment cleanly into the diary. The transcript was spotless.

What the caller had actually said was that she was in pain and could she come in today. The agent caught the words "root canal" and offered the next routine consult slot, four days out. Booked, logged, and wrong, and nobody noticed until the patient rang back angry.

That call is a fair answer to the question of what AI receptionists get wrong. I'd rather put the AI receptionist limitations in front of you now than let you find them in a recording three weeks after you go live.

The short answer

AI receptionists still get four things wrong: they mishear callers who don't sound like the training data, they miss urgency in the words a caller actually uses, they fail quietly instead of handing off, and they answer calls that should never have reached a machine. All four come from how the agent was set up rather than which model sits underneath it.

1. Accents, noise, and the callers who get misheard most

Speech recognition is not evenly good. The most cited work here is Koenecke and colleagues in PNAS in 2020, which tested five commercial speech systems and found an average word error rate of 0.35 for Black speakers against 0.19 for white speakers (study). Models have moved on a long way since then. The gap has not closed.

Add a builder calling from a van with the window down, or a caller in their seventies on a weak mobile signal, and accuracy drops again.

In practice this rarely looks like a dramatic failure. It's a postcode transcribed wrong, or a surname spelled three different ways across three calls, or a callback number with one digit off, which means the job is gone and you never find out why.

Here's what most people get wrong about this. They test the agent by ringing it themselves, from a quiet office, in their own accent. Of course it works. Your actual callers are not you.

2. It hears the words, not the worry

Language models are good at what was said. They're much weaker at what it meant, especially when someone is being polite about something urgent.

British callers understate constantly. "I don't want to be a bother, but the boiler's making a funny noise and there's a bit of water underneath" is a same-day job. An agent reading that literally books next Tuesday.

So we stopped trusting the model to infer urgency and started writing it out as explicit word lists per vertical. Water. Smell of gas. Swelling, bleeding, no heating with a baby in the house. Any hit and the call goes to a human on the spot, or triggers a callback inside five minutes. It's blunt, and it works better than anything cleverer we've tried.

3. Silent failure is the real risk

A human receptionist who doesn't understand something says so, puts you on hold, and fetches someone who does.

A badly configured AI agent will keep going instead. It paraphrases confidently, collects what it can, ends the call politely, and leaves you a half-wrong booking with a tidy log attached. That's worse than a missed call, because in your dashboard it looks like a win.

Every agent we deploy has a hard stop: two failed attempts to clarify the same field and it quits trying, says it's passing the caller to the team, and does exactly that. I'd take escalating 8% of calls over quietly mangling 3% any week.

If you're comparing vendors, ask what their escalation rate is. If they can't tell you a number, nobody is measuring it.

4. Some calls shouldn't go to a machine at all

Complaints. Bereavement. A patient ringing about test results. A tenant with no heating in January who has already called twice this week.

None of these are technically hard. They're just the wrong calls to automate, and the goodwill you burn costs more than the saved minutes are worth. Every client of ours has a short list of call types that route straight to a person, and we keep it short on purpose so the front desk actually remembers what's on it.

There's a compliance edge too. If you record calls in a dental or medical setting, consent, retention, and transcript access all matter, and the rules move depending on where you operate. We covered that in HIPAA and AI receptionists.

What they do better than a person

I'm not talking myself out of a business here.

They answer at 9pm, and they answer the sixth simultaneous call while your one receptionist is still stuck on the fifth, which is the whole argument in our piece on concurrency. They don't get sharper or blunter depending on what time it is on a Friday. And they log every call the same way, which for most small practices is the first usable call data they've ever had.

Speed is the part that shows up in revenue. The famous 100x contact figure comes from Dr James Oldroyd's work at MIT Sloan with InsideSales.com, published in 2007 and routinely miscredited to Harvard (original report). It's nearly two decades old, and I'd treat the exact multiples with suspicion, but the direction has held in every campaign I've run since. Answer first, win the job.

An agent that handles most calls imperfectly still beats voicemail, which handles none of them. If you want that argument with numbers attached, it's in our ROI breakdown.

How we set one up so the failures stay small

Four things, and none of them are model choices.

Start with 20 real recorded calls off your own line. Not scripts. Real ones, including the messy ones, and build the agent against those.

Run it on overflow and after hours first rather than your main line. If it breaks, it breaks on calls you were already losing.

Review every escalated call yourself in week one. That's maybe two hours of your time, and it catches more than any amount of prompt tuning.

Then set a hard review date at day 15. We run 15-day trials for that reason: long enough to hit real edge cases, short enough that nobody is stuck paying for something that isn't working.

FAQ

Can an AI receptionist understand strong regional accents?

Usually, though not reliably enough to skip testing. Feed it recordings of your own callers before launch and read the transcripts looking specifically at names, postcodes, and phone numbers. Those are the fields where an error costs you a job.

What happens when the AI doesn't understand a caller?

That depends entirely on configuration. A well-configured agent tries twice, then transfers or takes a callback. A poor one guesses and carries on. Ask any vendor what happens on the second failed attempt.

Will customers know they're talking to an AI?

Plenty will. We disclose it in the greeting for clinics and regulated services anyway, and callers mind far less than owners expect, as long as the agent either solves the problem or gets them to someone who can.

Can an AI receptionist handle emergencies?

It should recognise them and route them, not handle them. Build an explicit urgency word list for your trade and send those calls to a person on the first hit.

Do AI receptionists work for multi-location businesses?

Yes, and this is where they pull ahead of an answering service, because routing by location, service, and time of day is configuration rather than training up a new team member.

How long before an AI receptionist is actually reliable?

About two weeks of real calls with someone reviewing the escalations. Day one performance isn't the number to judge it on, and a demo isn't either.

What does an AI receptionist cost to run properly?

Less than a part-time front desk hire in every setup we've priced, though the honest answer moves with call volume and how much integration work your diary needs. Current numbers sit on our pricing page.

Is it better than a traditional answering service?

For speed, availability, and consistency, yes. For a distressed caller who needs a human being, no. Most practices end up running both, with the agent taking the volume and people taking the exceptions.

Where to start

Pull your last 20 inbound calls and listen to them with a pen in your hand. Mark the ones a machine would have handled fine, the ones it would have fumbled, and the ones it should never touch.

That gives you a build spec. Bring it to a 15-day free trial or book a demo, and we'll set the agent up against your actual calls instead of a script.

Sources

• Koenecke et al., "Racial disparities in automated speech recognition", PNAS, 2020. https://www.pnas.org/doi/10.1073/pnas.1915768117
• Oldroyd, MIT Sloan School of Management with InsideSales.com, Lead Response Management Study, 2007. https://content.marketingsherpa.com/heap/DG07SFSlides/LeadResponseManagementReport.pdf