There's a familiar comfort in blaming new things for old problems. When an AI voice system struggles to understand an accent or fails to book an appointment, it becomes the story. What rarely makes the story is what came before it: an NHS GP phone system so broken that millions of people had already stopped trying.
Before we assess whether AI is making GP access worse, we need to be honest about the baseline we're measuring it against. Because that baseline, for a very long time, has been quietly failing a significant proportion of the population.
In 2024, satisfaction with GP services continued to fall, reaching record lows in the British Social Attitudes survey. This wasn't a sudden collapse triggered by new technology. It was the endpoint of a decade-long deterioration in primary care access. Between 2015 and 2024, the ratio of GPs to patients worsened from one GP per 1,867 patients to one per 2,071. More patients, fewer doctors, and a phone system that became a pressure valve for all of it.
The 8am rush was not a quirk. It was a structural feature of a system with insufficient capacity. That daily ritual of hundreds of patients dialling simultaneously, each hoping to be one of the lucky few to get through, was the old system working exactly as designed. By 2023, 16 per cent of patients who tried to book a GP appointment didn't get one, up from 10 per cent in 2021. Of those, 12.2 per cent went to A&E instead, the equivalent of 696,000 people, a rise of 146 per cent in just two years, according to IPPR analysis of the GP Patient Survey.
These are not people who were served by the old system. These are people the old system lost.
Even for those who succeeded in reaching a receptionist, success was relative. Getting through at 8am often meant a timed window. Call too early and you'd get a busy tone. Call too late and there were no appointments left. Those who managed to speak to someone still faced being triaged, questioned, and sometimes redirected before they could book anything at all. The Care Quality Commission found that more than one in ten people who could not get a convenient GP appointment sought care via emergency departments instead.
The point here isn't that human receptionists were failing. Most were working within an impossible system, fielding hundreds of calls with inadequate staffing and no tools to match demand to capacity. The point is that the pre-AI phone system was already generating real harm: delayed diagnoses, A&E overcrowding, patients self-treating, and a growing cohort of people who had simply given up on trying.
Underneath the 8am scramble is a more intractable issue. There are not enough GP appointments to meet demand, regardless of how efficiently they are distributed. NHS England's most recent figures show GP teams carried out a record 388 million appointments in the last year, and there were 29.2 million phone calls in February 2026 alone, more than a million each working day. The system is operating at extraordinary volume. The problem is that demand exceeds even that.
AI voice systems did not create this gap. They emerged because commissioners and GP practices were looking for something that could help manage it. Whether those systems are currently performing well enough is a legitimate and important question. But it is a different question from whether the problem they were trying to solve was real. It was.
One of the first companies to operate AI voice systems at real scale inside the NHS was QuantumLoopAI, whose system EMMA has been handling patient calls across GP surgeries in England for over two years. The experience of building and running a system at that scale, through genuine NHS environments rather than controlled pilots, is precisely what separates early promise from meaningful progress.
The EMMA deployed today is substantially different from the one that first went into surgeries. When the system launched, it required patients to use the phonetic alphabet to spell their names, a workaround that reflected the limits of voice recognition technology at the time and one that rightly drew criticism from patients and clinicians alike. That requirement has since been removed. Name recognition has improved to the point where it is no longer necessary. Call handling times that once ran to twenty minutes have been brought down to five or ten. Accent recognition, particularly for regional voices that early voice AI struggled with, has improved through exposure to genuine patient calls across different parts of the country.
None of that improvement is visible in a bad patient experience. When a system fails someone, their frustration is entirely justified and the specific failure should be fixed. But a difficult experience with an early-stage technology does not tell us the technology is wrong. It tells us where the work still needs to be done, and it is exactly that kind of feedback that has driven the improvements already made.
The debate about AI receptionists tends to produce one of two reactions. There is uncritical enthusiasm from those who see technology as the answer to NHS access, and blanket rejection from those who see it as an assault on human-centred care. Both miss the point.
The right question is not AI or no AI. It is whether the system patients are using today is better than the one they were using before, and whether it is improving. For a growing number of patients, particularly those who were already failing to get through at 8am, the answer to both is yes. For others, particularly older patients, those with regional accents, and those with speech or cognitive difficulties, there are still genuine barriers that must be addressed. That work is ongoing and it is taken seriously.
The proportion of patients now describing contacting their GP practice as easy has risen to 73.7 per cent, up from two-thirds in summer 2024, according to the latest ONS figures. That improvement didn't happen by accident. It happened because practices are using better tools, and those tools are getting better.
The 8am scramble wasn't working. The answer isn't to go back to it. The answer is to build something that works for everyone, and to be honest about how much work that still requires.
Sources: