LISTEN TO THIS ARTICLE

The NHS Bet on AI Triage Is Bigger Than Anyone Admits

A single GP surgery in Surrey cut patient waiting times by 73% in four months. Not by hiring more doctors. Not by extending hours. By letting an AI decide who needs to be seen, when, and how urgently.

The Groves Medical Centre in South West London deployed Rapid Health's Smart Triage system in October 2023. An independent evaluation funded by Health Innovation Kent Surrey Sussex tracked the results through February 2024. Same-day appointment requests dropped from 62% to 19%. The 8am phone rush that defines the British GP experience effectively vanished, with peak-hour calls falling 47%. Face-to-face appointments actually increased from 53% to 85%, and GP consultation time grew from 10 to 15 minutes per patient.

Zero clinical incidents during the entire evaluation period.

That's one surgery. The question nobody in NHS England wants to answer honestly is what happens when you try to scale this across 6,300 GP practices serving 62 million patients, in a health system running on IT infrastructure that still relies on fax machines.

The Pilots Tell a Consistent Story

The Groves isn't an outlier. Up in York, Priory Medical Group deployed Klinik Healthcare Solutions' AI triage across its Primary Care Network. York Health Economics Consortium evaluated the results: phone contacts dropped from 99% to 30% of total contacts within two months. Reception staff, pharmacists, and clinicians saw a 20% reduction in tasks. The practice released over 300,000 pounds in capacity in the first year and delivered 18 months' worth of appointments in 12 months. Patient approval hit 87%.

These aren't Silicon Valley startups making promises. They're NHS practices running real patients through AI systems and publishing independently evaluated results.

The Tony Blair Institute pulled together data from multiple AI triage platforms in a 2025 report that projected NHS-wide benefits: 29 million additional GP appointments per year, 340 million pounds in annual productivity savings, and 28 million work hours freed up. Their analysis found that 1 in 6 GP appointments are currently unnecessary, and nearly 40% of A&E patients could have been treated in primary care with better triage. NHS 111 calls nearly doubled from 12 million to 22 million per year between 2014 and 2022, with over 200,000 calls abandoned monthly when wait times exceeded 30 seconds.

The numbers make a compelling case. But numbers made a compelling case for Babylon Health too.

The Babylon Cautionary Tale

Babylon Health is the ghost that haunts every conversation about AI in NHS primary care. Founded in 2013 by Ali Parsa, its GP at Hand service launched in November 2017 and promised to fix everything wrong with British general practice. Video consultations via an app. AI-powered symptom checking. NHS patients could register and see a GP within hours instead of weeks.

By mid-2020, over 47,000 patients had registered. Only 10% were actually from the surgery's catchment area in Hammersmith and Fulham. The rest were younger, healthier patients from across London who'd cherry-picked the digital-first option, leaving traditional practices with the complex cases and elderly patients but without the funding that follows registered patients.

The financial damage was specific and measurable. GP at Hand created a 21.6 million pound in-year deficit for Hammersmith and Fulham's Clinical Commissioning Group in 2019/20. The total health funding deficit hit 31.8 million pounds over two years. Local MPs called for government inquiries. An Ipsos Mori review concluded the funding model was "not appropriate."

Babylon went public on the NYSE in October 2021 at a 4.2 billion dollar valuation. Parsa's personal stake was worth roughly 825 million pounds. The share price collapsed from $272.50 to below $1. He later called the listing "an unbelievable, unmitigated disaster."

By September 2023, Babylon was in administration. Its assets sold for 500,000 pounds to eMed Healthcare. The CQC had flagged concerns about prescription misuse and off-label prescribing as early as 2017. Babylon tried to suppress that report by taking the CQC to the High Court. The court ruled it could be published. Babylon dropped the case and paid 11,000 pounds in legal costs.

Parsa returned in November 2024 with a new healthcare AI venture. Some lessons take longer than others.

What Patients Actually Experience

Healthwatch England's January 2026 report paints a picture that's messier than any pilot evaluation. Patients approved of AI handling administrative tasks but pushed back hard against AI making health-related decisions.

The specific complaints are telling. WhatsApp-based AI receptionists failed to transfer bookings correctly. One patient reported: "Can't get a GP appointment because I have to book through WhatsApp using AI options that don't give me options to say what I need." AI clinical summaries fabricated consultation content mid-recording. Medication request systems prescribed identical medications despite patients requesting changes.

Then there's the accuracy problem. A Nature systematic review found symptom checker diagnostic and triage accuracy "varied substantially and was generally low." Individual tool accuracy ranged from 11.5% to 90%. LLMs like ChatGPT and Gemini scored between 57.8% and 76% on symptom assessment, which is better than untrained laypeople (47.3% to 62.4%) but hardly reassuring for a clinical gatekeeper.

The systems tend to over-triage, which sounds safe until you think about what it means in practice: routing patients toward same-day urgent appointments they don't need, consuming exactly the capacity the AI was supposed to free up.

And the digital exclusion problem is real. AI chatbots proved incompatible with screen-reading software for blind users. Systems couldn't understand patients with autism, learning difficulties, or speech issues. GPs in more deprived areas were less likely to have AI tools (27%) compared to affluent areas (35%), according to the Nuffield Trust's December 2025 survey of 2,108 practising GPs.

The Data Privacy Minefield

In 2025, researchers at University College London and King's College London built an AI model called Foresight, trained on de-identified NHS data from 57 million people in England. It could predict future health outcomes, working like "auto-complete for medical timelines."

The BMA and RCGP said they were "not aware" that GP data collected for Covid-19 research was being used to train the model. The Joint GP IT Committee called it "very surprising and extremely concerning" and raised "serious concerns about the lawfulness of the data use." NHS England paused the project pending a Data Protection Officer review.

This is the core tension. AI triage systems need data to work. The NHS has some of the richest longitudinal health data in the world. But every data-sharing arrangement carries the risk that destroyed Babylon's credibility and now threatens to undermine the entire NHS AI programme.

As BMA deputy chair David Wrigley put it: "It only takes one disaster story of leaked confidential data or data sold on for profit to result in the loss of trust in this technology and loss of trust from patients."

Imperial College London has warned that AI trained on unrepresentative data could worsen health inequities for minority ethnic groups. NHS England is piloting "algorithmic impact assessments" on healthcare datasets in response, and the STANDING Together project is developing international standards for dataset diversity. But these initiatives are early-stage while deployment is accelerating.

The Regulatory Vacuum

The MHRA launched a National Commission on AI Regulation in Healthcare in September 2025, with recommendations expected sometime in 2026. The Topol Review called for a digitally ready NHS workforce back in 2019. Seven years later, the Nuffield Trust found that 84% of GPs cite lack of regulatory oversight as a concern, and the rollout has been described as a "wild west" by GP publications.

Here's what's actually happening on the ground: 28% of GPs now use AI tools at work, but 11% are using independently obtained tools like ChatGPT rather than practice-approved systems. Nobody is auditing what prompts they're using, what patient data they're inputting, or whether the outputs meet clinical standards. The verification gap that plagues AI agent systems everywhere is alive and well in British primary care.

The MHRA commission has four working groups spanning AI specialists, clinicians, and regulators across England, Scotland, Wales, and Northern Ireland. Their call for evidence closed in February 2026. Whatever framework they produce will arrive after adoption has already outpaced regulation by at least two years.

What This Actually Changes

The evidence from Groves and York is hard to argue with. AI triage works in controlled pilots with motivated practices. The efficiency gains are real. Patients who can use the systems tend to like them.

But the NHS isn't a collection of motivated pilots. It's a creaking system where 300 deaths per week are associated with A&E delays, where GP satisfaction has halved since 2012, and where the gap between what technology can do and what infrastructure can support grows wider every funding cycle.

The smart money says AI triage will become standard in NHS primary care within five years. The honest assessment says it'll work brilliantly in practices that were already well-run and create new problems in practices that weren't. The patients who need the most help, the elderly, the digitally excluded, those with complex conditions, are precisely the ones these systems serve worst.

Babylon proved that a good demo and a compelling pitch deck aren't the same as a functioning health service. The current generation of AI triage tools is demonstrably better than what Babylon offered. Whether the NHS has learned enough from that failure to deploy them equitably is a different question entirely.

Sources

Research Papers:

Industry / Case Studies:

Data Privacy & Regulation:

Babylon Health:

Related Swarm Signal Coverage:

Keep reading

Join the Swarm Signal newsletter

Get the Freelance Command Center on Payhip