A quarter of England’s population drives 82% of hospital cost, and the healthy share is shrinking. Every developed health system faces a version of this problem. For investors, the questions are which buyer pays for AI, in what order adoption happens, and whether the value holds to exit.

Ben Richardson, Managing Partner, Carnall Farrar. Based on his opening keynote at the CF and Odgers Berndtson evening on private equity in healthcare, 5 November 2026.

Multimorbidity, not ageing, is the fastest-growing cost

CF has linked seven years of patient-level data for the whole of England in HealthStrata: about 4 billion records, roughly 70 per person. Three findings frame everything else.

1. 25% of the population accounts for 82% of hospital cost and 90% of bed days. The generally healthy three quarters account for 18% of cost.
2. 12% of the population has multimorbidity and accounts for 31% of hospital cost and 33% of bed days. It is the largest single segment by cost.
3. 2% of the population lives with frailty or dementia and accounts for 17% of hospital cost and 22% of bed days.

The generally healthy share of the population fell from 79% to 75% between 2022/23 and 2025/26, as 2.5 million more people came to live with long-term conditions. Ageing explains part of that. The faster-moving component is multimorbidity, and about half of its growth traces back to obesity. Among people with chronic kidney disease, 99% have at least two other long-term conditions. England is not unusual here; it is simply the system with the data to show it.

Money has followed the sickest. Acute and community cost rose by £15 billion in real terms over three years, from £100 billion to £115 billion, and three quarters of the rise came from multimorbidity, frailty and dementia, and end-of-life care.

Source: CF Health Strata© and National Cost Collection, 2025/26, CF analysis. HealthStrata© Carnall Farrar Ltd all rights reserved. May not be reproduced or distributed without express written consent.

The spending has not reversed the problem. The acute sector’s share of NHS spend rose from 47% in 2006 to 58% in 2023, about £18 billion a year that policy had promised to move out of hospitals. Since 2013/14, hospital output per pound, per doctor and per nurse has fallen 10 to 14%, worth £14 to £17 billion, and the decline began before Covid. That productivity gap is what AI is being bought to fix. The cost of leaving it unfixed reaches beyond the NHS: economic inactivity due to ill health is at its highest level in data going back to 1995.

Source: CF Health Strata© Carnall Farrar Ltd all rights reserved. May not be reproduced or distributed without express written consent.

Patients are paying to skip the queue. Self-pay admissions to private hospitals reached a record of almost 77,000 in the first quarter of 2026, up 7.7% on a year earlier, and now fund 31% of private admissions. Insured admissions fell 2.6%, the first fall in five years (PHIN, September 2026). The growth is concentrated rather than market-wide: cataracts account for almost half of it. That gap between need and NHS capacity is the private-sector opportunity in miniature.

Source: PHIN September 2026

Where AI can act, and by how much

Health is behind other sectors in how it meets the consumer. Banking moved from rows of tellers to an app in a decade. Outpatients still looks like a waiting room. CF’s analysis suggests roughly half of outpatient appointments produce no change in care, and around 70% of patients conveyed by ambulance are discharged without follow-up.

CF’s work for Lord Darzi’s Future State programme, with Narayana Health and Nesta, sized the opportunity in three buckets. The ranges are in-principle estimates and assume clinicians and patients use the tools.

1. 20 to 40% lower cost and better outcomes from earlier presentation: bringing forward cancer, chronic and rare-disease diagnosis (a rare-disease diagnosis takes five to 10 years on average) and closing care gaps where the average NICE guideline reaches only about half of eligible patients.
2. 20 to 70% reduction in demand through digital-first care. A CF clinical panel put the share of chronic care contacts that could be avoided or resolved digitally at 43 to 66% in A&E and outpatients and 47 to 73% in general practice. The opportunity is in clinics, not wards: only around a quarter of admissions could be avoided.
3. 20 to 40% productivity gain: ambient voice for clinicians, fewer whole-time equivalents per unit of administrative output, and optimisation of theatres, beds and scarce staff.

Measured productivity gains so far are modest and sit in documentation rather than clinical decisions:

1. 8% shorter appointments from ambient scribes in the GOSH-led NHS evaluation across nine London sites and more than 17,000 patient encounters (2025), and 13% more patients seen per shift in A&E.
2. 0.8 extra encounters a week per physician, a 2.8% increase, from ambient scribes at UCSF across 1.2 million visits (JAMA Network Open, 2026).
3. £834 million a year is the headline national figure attached to the NHS evaluation: £176 million of documentation time and £658 million of extra A&E capacity, assuming every A&E clinician sees one more patient per shift. It is modelled, not banked.

Source: CF Health Strata© and National Cost Collection, 2025/26. HealthStrata© Carnall Farrar Ltd all rights reserved. May not be reproduced or distributed without express written consent.

At specialty level, elective growth is concentrated. Eight specialties account for about 65% of the increase in elective cost, led by trauma and orthopaedics at 22% and gastroenterology at 10%, alongside general surgery, cardiac, ophthalmology, urology, gynaecology and clinical haematology. The remaining 35% is spread across around 150 other specialties. Outpatient growth is more evenly spread. Each specialty has grown roughly in line with its size, so no small group of specialties dominates the increase. That makes outpatients a capacity problem across the whole system, rather than a few pathways under strain.

Source: CF, Narayana Health and Nesta (2025), Future State Programme. CF Clinical Research Panel.

The trade-off is that the largest bucket, demand reduction, depends most on behaviour change and is the one payers have least practice in paying for. Productivity tools have the shortest path to a contract and the smallest ceiling.

What others report: cost, not revenue

The published evidence across sectors points the same way as the ambient voice results. Where returns exist they sit in back-office cost functions, and reported revenue impact is limited. MIT’s NANDA study (2025) found that 95% of enterprise generative AI pilots had no measurable impact on profit and loss, and that the largest returns came from back-office automation, even though more than half of AI budgets went to sales and marketing.

This fits the adoption sequence set out in the next section. The gains that have been measured come from the second and third stages, where tools touch administration and clinical documentation rather than clinical decisions. Revenue depends on the fourth stage, and almost nobody is there yet. For a seller, that means the credible number at exit today is a cost number.

Adoption is the constraint, and evidence is the gate

The conversation has moved from experimentation to evidence, and not only among investors. Regulators need evidence to authorise a tool at all, and they are now asking whether the model was trained on a population that resembles theirs. Health systems, life sciences companies and insurers need evidence that it will deliver a measurable benefit before they buy. Providers need it to deploy safely. The supply of that evidence is thin: only 5% of 519 published evaluations of large language models in healthcare between 2022 and 2024 used real patient data (Bedi et al., JAMA, 2024, cited in CF, Narayana Health and Nesta, 2025). Much of what is being deployed has not been tested on the population it is used on.

That puts a premium on a data environment that can show two things: a representative population and replicable results. Cradle-to-grave, record-level data with universal coverage is one of the few that can, and the MHRA and NICE are already asking for evidence of this kind. For cost-saving and productivity tools the test is representative data, replicable results and demonstrably safe use. For clinical tools the bar is the same and the record of getting over it is worse.

Radiology shows both halves of the problem. The capacity gain is real and available now. Tools already on the market, set to high sensitivity, can identify 20 to 40% of patients scanned for suspected brain haemorrhage as normal with confidence, so they can go home sooner. One supplier offered this to an NHS organisation free of charge. The organisation couldn’t take it up because it didn’t have the contracting, procurement or implementation capacity to do so. Even at no cost, the tool went unused. Mammography tells the same story at scale. Kheiron’s Mia found 12% more cancers in the 10,889-patient GEMINI study, and DeepMind’s model, developed with Imperial, outperformed radiologists. The UK National Screening Committee still judged the evidence insufficient. Kheiron was acquired by DeepHealth, part of US-based RadNet, in 2024, and about 35% of US imaging centres now use AI for detection (CF, Narayana Health and Nesta, 2025).

The irony is that the UK has permissive regulators and cautious systems. The MHRA will clear a tool the NHS then cannot buy, in a service where there is no penalty for saying no and every penalty for a mistake. That gap is the opening for private operators and insurers: adopt first, and build a superior ability to embed AI in the systems and processes that either enable NHS services or compete with them. The condition is a clear understanding of where each tool sits in procurement requirements and in regulation; that is where most private deployments stall. The trade-off is that moving first means carrying the liability risk the NHS is waiting out.

Narayana Health’s experience, which informed the Future State work, suggests adoption then follows a sequence. Investors should expect it and price it.

1. Paperless first. Universal electronic records on cloud infrastructure come before any digital or AI layer. Without reliable baseline data, AI has nothing to work on.
2. Safe management operations next. Reporting from integrated data, then administrative tasks for patients and staff. Low regulatory risk, fast proof of benefit, and a place to build trust before touching clinical decisions.
3. Productivity tools third. Convenience and labour productivity tools that need no regulatory approval bring 65% or more of cost into scope. Natural language processing for notes and queries offers large gains but needs roughly a fourfold increase in adoption to deliver them. Supply productivity gains of around 20% are a realistic target.
4. Clinical deployment last. Once trust and regulatory approval exist, up to 90% of patient interaction comes into scope. Negligence risk is real. The prize is resolving up to 80% of demand digitally, depending on the setting.

Source: CF, Narayana Health and Nesta (2025), Future State Programme.

Five enablers run across all four stages: a bold aspiration that redesigns practice rather than digitising it; change management that keeps clinicians and managers inside the process; investment in skills as well as technology, in non-technical roles as much as technical ones; measured and reported benefits in productivity, cost, satisfaction and outcomes; and governance for safe decision-making and risk. A portfolio company that cannot show measured benefits has nothing to sell at exit.

Three buyers, three theses

Across these three buyers we see four investable segments, each winning on something different: AI-enabled operators win on cost, with the risk that consultants they do not employ cannot be told to adopt; health system software wins on adoption, with the risk of pilots that never scale; direct-to-patient digital wins on consumer demand, with the risk that free public services absorb the simplest contacts; and life sciences AI wins on data, with the risk of access and governance delays.

Markets have started to price the doubt

The Nasdaq 100 closed in correction on 2 November, 10.4% below its 6 October record. Three things rattled it: slowing growth in Google’s search advertising revenue as AI answers change how search makes money; the scale of AI data-centre spending now being disclosed, in Anthropic’s IPO filing and Oracle’s $20 billion debt and equity raise among others; and a voter backlash against data centres that the US midterms confirmed the following day. Markets are asking the question health investors should already be asking: where is the revenue?

Set against the size of the opportunity and the difficulty of adoption, that market mood sharpens the questions every investor should ask.

Six questions before investing, one for the exit

1. What unmet need does this address, and how large is it in the data rather than the pitch deck? AI changes the contact, not the need, and demand only rises in multimorbidity, cardiometabolic disease, musculoskeletal conditions and mental health.
2. How aligned is it with policy, and how durable is that alignment through a change of government?
3. Who pays, and have they demonstrably paid for something like it before: the health system, an insurer or the patient? A buyer wants before-and-after unit cost against tariff, measured rather than modelled.
4. Is there a sustainable margin once the cost of inference and of compliance is counted? Margin widens only where the operator owns the infrastructure and employs the workforce.
5. Can the value be captured and held for the length of the hold, or does it leak to the platform providers?
6. Does the capital match the asset? Long-lived assets such as dementia care need capital of matching duration; a five-year hold on a 30-year asset lies behind many of the sector’s distressed exits.

The exit question follows from all of this. A buyer in 2028 or 2029 will not pay for an AI narrative. They will ask for cost impact visible in the P&L, for revenue uplift only where a payer has actually paid, and for evidence of which adoption stage the company has reached. Those that start measuring now will have a defensible number at exit. Those that do not will be selling a story into a market that has just learned, at some expense, to discount stories.

FAQs

Sources

• CF HealthStrata linked record-level data for England and National Cost Collection, 2025/26, as presented on 5 November 2026.
• Carnall Farrar, Narayana Health and Nesta (2025) Future State Programme: the impact of data, digital and AI in healthcare, June 2025, available at https://www.carnallfarrar.com/future-state-programme-the-impact-of-data-digital-and-ai-in-healthcare/.
• Private Healthcare Information Network (2026) Private healthcare market update, September 2026, available at https://www.phin.org.uk/press-releases/private-market-update-sept-2026-uk. Bedi, S. et al. (2024)
• Testing and evaluation of health care applications of large language models: a systematic review, JAMA, doi:10.1001/jama.2024.21700. MIT NANDA (2025)
• The GenAI divide: state of AI in business 2025.
• Great Ormond Street Hospital (2025) GOSH-led trial of AI scribe technology across London, 4 September 2025, with economic modelling by York Health Economics Consortium.
• UCSF (2026) Ambient artificial intelligence scribes and physician financial productivity, JAMA Network Open. Nasdaq 100 closing levels, 6 October and 2 November 2026.
• Mammography evidence (DeepMind, Kheiron GEMINI and iCAIRD, UK National Screening Committee) as summarised in the Future State report.
• Acute share and productivity figures are CF analysis of NHS England and ONS data.

About CF

CF is a leading consultancy dedicated to making an enduring impact on health and healthcare. We work with leaders and frontline teams to improve health, transform healthcare, embed life science innovation and boost growth through investment. With unmatched access to UK healthcare data and award-winning data science expertise, our team are a driving force for delivering positive and meaningful change.

About the authors

Ben Richardson

Ben Richardson is a Managing Partner at CF, leading Life Sciences and Data Innovation. With two decades of experience, he has worked with health systems and life sciences companies globally, focusing on strategy, transformation, and development. Ben has contributed to primary care, diabetes, cardiovascular, cancer, mental health, and population health management. Since 2014, he has helped CF become an award-winning healthcare company in management consulting and data services.

Ellen Teesdale

Ellen is the Practice Manager for Health Investing, working to expand CF’s work and expertise into the private sector. Ellen manages the health investing practice, working with a diverse range of clients to facilitate the successful integration of innovation into healthcare systems.

Christina Rashbrook

Nina works across business development for CF’s health investing and life sciences practices. She brings more than seven years of frontline healthcare experience and has just completed an MSc in Global Health Management at Imperial College London. For the health investing team, she runs pipeline reporting, shapes proposals and case studies, and turns CF’s events into published insight.

Vernon Baxter

Vernon is an Associate Partner at CF, connecting leaders in private equity, healthcare and banking with the team at CF. Vernon brings extensive market intelligence to the team, who advise clients on commercial and operational due diligence, value creation and growth strategy.