Missed calls are the quiet leak in every clinic. I worked through the full cost of a missed call in an earlier post on the real math of missed calls, and the short version is that a phone answered less than half the time is a phone dropping real revenue. AI phone answering fixes the answering problem: the tools are real, the price is modest, and the first NHS surgeries are already running them. The marketing overstates the rest. Here is what works, what does not, and what it actually costs.
What the tools actually do
An AI answering service sits on your phone line, picks up in the first ring or two, and handles the routine: taking a message, answering the same five questions patients ask every day, booking or confirming a slot, and routing the rest to a person. That is the entire honest job description. The value is simple: a phone that is never unanswered, even at lunchtime, even on a Tuesday afternoon when the one receptionist is with a patient.
The boundary matters more than the capability, so here it is as a table:
| What the AI handles | What it does not |
|---|---|
| Taking messages and routing calls | Triage symptoms or make clinical judgements |
| Booking, confirming or cancelling slots | Access patient records on its own |
| Answering the routine questions patients ask daily | Handle a distressed caller without a human |
| Transferring to a person on request | Explain a clinical decision |
The GP surgery that tried it
The most useful evidence is the Kirton Lindsey and Scotter Surgery in North Lincolnshire, which put an AI receptionist named Emma on the phones in November 2025. The practice, about 10,500 patients, wanted nobody to be 10th or 15th in a queue. Emma answers, collects the reason for the call, and a clinician calls most patients back within two to three hours. The practice insists AI will not replace its human receptionists, and patients can ask to be transferred to a person at any time.
The honest part is the feedback. Some patients praised the shorter waits; others told the BBC the system misheard dates of birth and left them shouting at a robot. The lesson is not that the technology fails, it is that the design of the fallback decides whether it works. A patient in pain should never have to repeat themselves twice, once to a machine and again to a human, before they get an appointment.
Patient data changes the buying decision
Your callers’ health details are special category data under UK GDPR and the Data Protection Act 2018. That changes how you buy. The AI vendor is a processor acting on your behalf, and you remain the controller, so you need a written data processing agreement, Article 28 of the GDPR, before any call is recorded. Ask four questions and take the answers in writing: where is the audio stored, is it transcribed, how long is it kept, and who can listen to it. UK or EEA storage, short retention, and access limited to your staff are the only good answers. A vendor that cannot answer these is not a vendor, it is a liability.
Human fallback is not optional
The rule should be mechanical: any caller who asks for a person gets one, instantly, with no menus and no repeats. That is how the GP trial keeps its licence to operate, and it is how a clinic keeps its reputation. The AI’s job is to catch the calls you are currently losing, not to become the new queue.
What it costs
Pricing from US and UK vendors in 2026, converted and rounded:
| Service type | Monthly cost | What you get |
|---|---|---|
| AI answering service | £40 to £240 | Instant pickup, messages, routine booking |
| Live answering service | £80 to £800+ | Human agents, message taking |
| Per minute pricing | £0.60 to £1.20 a minute | Pay for actual call time |
| Per call pricing | £0.80 to £9 a call | Short routine calls only |
| Hybrid AI plus humans | £120 to £400+ | AI front door, people for escalations |
Most AI services quote in dollars; at mid 2026 rates the typical landing zone is £50 to £250 a month for a single practice line, with overage fees if you blow through the included minutes. Integration with your clinical system, staff training and setup are often extra. Compare that against the human alternative: a receptionist costs a salary, not a subscription, but a receptionist also cannot answer a call when they are already on another one.
The test is not whether the robot sounds human. It is whether a patient who cannot work the robot reaches a person in seconds, every time, without explaining themselves twice.
What to do now
Count your missed calls for two weeks, literally tally every unanswered ring. If the number surprises you, take a one month trial with three rules written into the order: recorded calls stored in the UK, a one touch transfer to a human, and a 30 day exit with your data deleted. The trial will tell you in a month what this article took a thousand words to say.