Healthcare has unusually large room to gain from AI, and the evidence now shows where those gains are real and what limits how far they spread. Health data makes up an estimated 30% of all the data produced across industries, yet the sector uses less than 5% of it for decisions, according to the OECD's 2026 report on scaling AI in health.[1]
The gains measured so far sit in narrow tasks that can be checked, such as billing and claims, breast cancer screening and eye checks moved into routine clinic visits. Beyond those tasks, adoption slows. Clinicians are unsure who is to blame when AI is wrong, and medical device rules were largely designed for products that change little once approved.[2] Bain & Company and Bessemer Venture Partners describe clinical AI as "much earlier in adoption as it requires trust, clarity of liability, regulatory and reimbursement pathways".[3]
Exhibit 1. Measured AI gains come from narrow tasks in one pathway or organisation, and only three of these eight results come from randomised trials.
Read together, the 2026 evidence points to a sequence for health leaders. Save first in the back office, where staff already trust the tools. Prove outcomes next in screening and specialist checks, where a narrow task can be measured against a clear result. Scale into wider clinical use only with governance that keeps watching a tool after launch, settles who is liable and tells patients when AI is involved. Each step pays for and protects the next, and skipping one spends trust that is slow to rebuild.
Healthcare's AI upside is large because its data sits idle and its staff are short
The case for AI in healthcare rests on the gaps it could close more than on the power of any one model. The OECD calls health the largest and fastest-growing source of data of any sector, yet its records, scans and notes are fragmented within and across health systems.[1] Health has also invested less in the means to use that data. In 2024 the OECD estimated that health spends about half the share of its budget on technology that the financial industry does.[4]
The second gap is people, and it is widening as health workers age alongside their patients. The World Health Organization (WHO) projects a global shortage of 11.1 million health workers by 2030.[5] It also finds that nearly one doctor in four is nearing retirement age.[5] A system that cannot use most of its data and cannot hire enough staff has more to gain from automation than one that already runs lean.
The same gaps explain why the gain arrives slowly, because every use of AI has to pass through clinicians who are short of time and patients who are cautious. The OECD argues that health is not unique and can learn from industries such as finance that handle similarly sensitive data.[4] What sets health apart is the combination of a regulated product, a clinician who is professionally accountable and a patient who carries the risk, which is why the UK's National Commission into the Regulation of AI in Healthcare finds that AI products present regulatory challenges that differ from those of traditional medical devices.[2]
Non-clinical AI is cutting healthcare costs first
The fastest returns from AI in healthcare come from work that never touches a diagnosis. Bain and Bessemer put the cost of administering US healthcare at around a trillion dollars a year.[3] They estimate US$260 billion of it is waste.[3] Their 2026 survey of payer, provider and pharmaceutical executives finds the gains in adoption concentrated in administrative work, such as payer contracting, claims and the provider back office.[3] A 2026 review in BMJ Digital Health & AI likewise finds that the most mature real-world deployments of AI in healthcare are still administrative.[6]
Staff confidence helps explain why the back office moves first in healthcare. A Health Foundation survey cited by the Commission found that 86% of health staff consider AI safe for administrative purposes, against 48% for patient care.[2] The UK Commission recommends clarifying when software for administrative purposes is not a medical device, and expects that much of it is not.[2] A back-office tool can therefore be judged on the time and cost it saves, without first proving a clinical result.
Health systems then have to decide what to do with the time saved. Roughly half of the organisations in the Bain and Bessemer survey have cut or plan to cut headcount because of AI, mostly in back-office roles.[3] The same organisations plan to put less than a tenth of next year's additional AI investment into governance, risk and compliance.[3] Health systems that bank every back-office saving as a job cut leave the clinical step underfunded. A share of those savings belongs in the monitoring and training that clinical AI will need.
Health insurers have been slower to scale AI than providers, even though much of their work is administrative. Bain and Bessemer find payers still lagging and focused on generating ideas and testing them, while providers and pharmaceutical companies have moved on to scaling.[3] Stanford HAI reports that many insurers do not document the accuracy of their prior authorisation tools or test them for bias.[7] Fewer than a quarter tell providers when they have used AI.[7] The late start helps insurers only if they use it to build that testing and disclosure in before they scale.
The next wave of AI in healthcare makes measuring savings more important. In McKinsey's survey of US healthcare leaders, carried out in late 2025, about one organisation in five already uses agentic AI, software that carries out a whole chain of tasks, and many more are testing it.[8] The leaders Bain and Bessemer surveyed report an average return of 3.5 times on their AI investments, but these are their own estimates.[3] Measured time savings on clinical notes range from seconds per note in one review to most of the documentation time in another study.[6][1] Each health system has to measure its own saving.
In screening, a large randomised trial shows AI finds more cancers without more false alarms
Screening is where AI has its clearest outcome evidence, because a trial can count cancers found and missed. The MASAI trial in Sweden randomly assigned more than a hundred thousand women to AI-supported mammography or to standard reading by two radiologists. AI support found 80.5% of cancers against 73.8% with standard reading.[9] The rate of false alarms was the same.[9] Cancers appearing between screenings were no more common, and fewer of them were invasive.[9] The investigators also reported a lighter reading workload for radiologists.[9]
Real-world programmes point the same way when the task is narrow. At a dozen screening sites in Germany, radiologists who chose to use AI support detected nearly a fifth more breast cancers without a significant rise in recalls.[6] At more than two dozen NHS hospitals using stroke software, rates of clot removal doubled at the AI-enabled sites, a rise significantly greater than in hospitals without AI.[6] Both are observational results. Both also show the same pattern of a checkable task inside an established pathway.
That narrowness also limits what the evidence can show about AI in clinical care. The BMJ Digital Health & AI review behind those results finds that fewer than 5% of the AI devices cleared by the US Food and Drug Administration have undergone peer-reviewed evaluation.[6] A June 2026 scoping review of human and AI teamwork found that many evaluations stop at whether the decision was correct and rarely follow the patient.[10] Screening is therefore the place to prove outcomes before scaling. A health system that skips this step scales tools whose effect on patients nobody has measured.
AI widens access by moving specialist checks to the first visit
The strongest access evidence comes from moving a check that usually needs a referral into a routine visit. In the ACCESS randomised trial, published in 2024 and run at two paediatric diabetes clinics in Baltimore, every young person offered an autonomous AI eye exam at the clinic completed it, against 22% of those referred to an eye specialist.[11] Of those whose AI exam was abnormal, 64% went on to see an eye care provider.[11] The trial was small and limited to young people. It shows how much care is lost between a referral and the visit that should follow.
Remote monitoring extends the same idea beyond the clinic. It is still at an early stage. WHO Europe's survey of its member states finds AI-assisted remote monitoring in use in some member states, and established practice in very few.[12] Access gained this way still needs a clinician behind the result. The risk is that some patients already use AI in place of a clinician.
A King's College London survey, carried out in March 2026, found that 15% of UK adults have used an AI chatbot for health advice instead of contacting a GP or other NHS service.[13] About one in five of those who sought health advice from AI report that something a chatbot told them led them not to seek professional care.[13] General-purpose AI models are usually not intended for medical purposes and are not regulated as medical devices, as the UK Commission notes.[2] Health systems that offer no trusted route of their own leave patients to those tools.
Quality gains so far come from fewer clinical errors
The clearest gains in quality of care come from AI that checks a clinician's work as it happens. Penda Health and OpenAI tested such a tool in a 2025 quality improvement study across fifteen clinic sites in Kenya.[6] It covered nearly forty thousand visits.[6] AI decision support was followed by 16% fewer diagnostic errors and fewer treatment errors, as rated by physicians, according to the BMJ Digital Health & AI review.[6] Patient-reported outcomes, however, did not differ significantly.[6]
That distance between fewer errors on paper and better health in practice is the quality problem in miniature. Clinicians working with AI do not automatically do better than the AI alone. The same review finds that clinician and AI teams often fail to outperform the model on its own. It also describes a study in which a system built from individually superior parts underperformed a less optimised design, because information did not flow well between them.[6] Quality depends on how the tool, the record and the team work together.
Clinicians themselves are clear about what helps AI improve their work. A February 2026 systematic review of qualitative studies found hands-on training, early involvement in design and clear explanations of how a tool reaches its answer to be what helps most.[14] Gaps in skills, worries about legal accountability and poor fit with existing workflows were the main obstacles.[14] Quality is earned hospital by hospital, so a tool proven in one setting needs its own checks in the next.
Health AI governance must cover tools that can change after approval and the people who over-trust them
A one-time approval cannot govern health AI, because the tool, the people using it and the setting can all change after launch. The UK's National Commission, whose recommendations the government accepted in full in October 2026, warned that AI tools "may iterate rapidly, perform differently in different settings and depend on the data, workflows, people and organisations around them".[2][15] Its answer is oversight across the whole life of a tool, with approval widening in stages as evidence builds.[2]
Human oversight, the safeguard most health systems rely on, is weaker than it looks. In one randomised trial cited by the BMJ Digital Health & AI review, physicians already trained in AI scored 14 percentage points lower in diagnostic accuracy than a control group when the AI gave them wrong recommendations.[6] WHO warned in its 2024 guidance on large multi-modal models that such tools can encourage "automation bias" in professionals and patients alike.[16]
Training can improve how clinicians work with AI, yet this result shows that it leaves the risk in place. Oversight has to be designed into the work and its effect measured like any other clinical control. Governance as a whole receives less than a tenth of planned new AI investment, so this work competes for a small budget.[3]
Liability is the next gap, and it tends to fall on clinicians. "When something goes wrong with AI, responsibility is often placed on clinicians, even where they have limited control over how AI tools are introduced," said Professor Graham Lord, Executive Director of King's Health Partners.[17] In the Bain and Bessemer survey, 58% of providers say the treating clinician bears primary accountability for clinical decisions influenced by AI.[3]
That uncertainty slows the move to scale as well as worrying clinicians. Half of the same providers call liability a barrier to moving AI from pilot to full deployment.[3] Fewer than one in ten member states in the WHO Europe survey have liability standards for AI in health.[12]
Patients are the last link, and they are rarely told when AI shapes their care. Only 10% of UK patients report being clearly informed, according to The Patients Association, as cited by King's College London.[13]
Clear legal standards and honest disclosure are what turn oversight into trust. "Without clear legal standards, clinicians may be reluctant to rely on AI tools and patients may have no clear path for recourse if something goes wrong," said Dr David Novillo Ortiz, Regional Advisor on Data, Artificial Intelligence and Digital Health at the WHO Regional Office for Europe.[18]
Health systems in the Gulf Cooperation Council (GCC) can write these safeguards in before clinical AI spreads, and Abu Dhabi has begun to do so. Its Department of Health standard, in force since October 2025, requires healthcare AI systems to inform patients and clinicians when AI contributes to clinical or operational decisions, and says clinicians should retain ultimate decision-making authority.[19] "If you add governance at the end, it becomes a bottleneck," said Majid Sultan AlMheiri, Chief AI Officer and Director of IT at the Dubai Health Authority.[20]
In Europe, legal uncertainty and cost hold health AI back most, and clear rules help most
When national experts across Europe rate the barriers, law and money come first. In the WHO Europe survey of member states, legal uncertainty was rated a major barrier by 48% and affordability by 46%, ahead of data quality and capacity, with infrastructure and trust further down.[12] Clear liability rules, guidance on transparency and explainability, and monitoring of AI after it reaches the market top their list of enablers, each rated a major or moderate help by at least nine in ten.[12]
Exhibit 2. National experts across Europe rank legal uncertainty and cost as the main barriers to health AI, and rules on liability, transparency and monitoring as the biggest help.
No hospital can buy any of these enablers, because governments and regulators have to write them. The most useful work health leaders can do here is to press for them early, alongside the integration and skills work they control themselves.
The same gap between use and rules shows across OECD member countries. Every responding member country already uses AI in health administration, yet fewer than one in five reports a national strategy or an oversight body for AI in health, and only about one in ten has updated its procurement guidelines for AI.[1] Health systems are buying tools faster than their countries are building the rules to scale them.
Exhibit 3. Most of these ten health systems have adapted device approval for AI, but few have an AI oversight body, procurement rules for AI or a health AI law.
In the US, the leading barrier for healthcare leaders is more practical. Integration with existing systems is the barrier they cite most often in McKinsey's survey, ahead of concerns about risk and gaps in capability.[8] Every tool has to work with records and workflows that were not built for it. The provider pays for that integration. Its cost is a reason to set common standards for how new tools connect before buying many of them.
Skills are the enabler that most health systems have yet to provide at scale. Only a quarter of the member states in the WHO Europe survey train health workers in AI while they are in post.[12] France has gone further, mandating AI and digital health training across all health professional programmes, backed by EUR 119 million to train half a million professionals over five years.[1] Given the automation bias result, training is the minimum for safe use. Most of the protection has to come from checks built into the work.
The gains go to health systems that build trust before they scale
The sequence of save, prove and scale also settles when liability belongs on the agenda. Many back-office tools are likely to fall outside medical device rules,[2] so providers can bank those savings now. Wider clinical use should not run ahead of the law. Liability rules take time to settle. Health leaders therefore need to press for clear liability and transparency rules from the first step, long before the clinical tools that depend on them are ready.
For health system leaders, the agenda follows from the evidence. Put a fixed share of each back-office saving into monitoring and training. Prove each clinical tool on a measured outcome in a narrow pathway before widening it, and name the person accountable when it is wrong before it goes live. Tell patients when AI shapes their care, since few say they are told.
Insurers have a narrower priority. They need to test the accuracy of their AI tools for claims and prior authorisation, and disclose their use, before those tools scale, because claims and authorisation decisions determine whether care happens. Insurers also set the pace for providers' clinical AI, since Bain and Bessemer find payers still unwilling to reimburse AI that acts without a clinician's involvement.[3]
For boards, the practical test of any AI proposal is whether it names the outcome it will measure, the person accountable when it is wrong and how patients will be told. Health systems that answer those three questions before they scale will turn healthcare's large potential gain into lower costs, wider access and safer care.
Frequently asked questions
What is the use of AI in healthcare?
AI in healthcare is used for administrative work such as claims, coding, scheduling and clinical notes, for screening and diagnosis support, for triage and remote monitoring, and for drug discovery.[3][6] The strongest outcome evidence so far is in breast screening, where the Swedish MASAI trial found AI support detected 80.5% of cancers against 73.8% with standard reading.[9]
Which industry will benefit the most from AI?
There is no agreed ranking of industries by AI benefit, but healthcare has unusually large room to gain. Health data is an estimated 30% of all industry data, yet the sector uses under 5% of it for decisions.[1] WHO projects a shortage of 11.1 million health workers by 2030.[5] Capturing that gain depends on trust.
What are the downsides of AI in healthcare?
The main risks are missed errors, tools that behave differently by hospital, unclear liability and uninformed patients. In one randomised trial, physicians trained in AI scored 14 percentage points lower in diagnostic accuracy than a control group when the AI was wrong.[6] Fewer than one in ten WHO Europe member states surveyed have liability standards for AI in health.[12]
What are medical AI chatbots trained on?
Many are general models trained on large datasets, such as the internet, which WHO warns "may be rife with misinformation and bias".[16] Others are trained on medical case reports, X-rays or millions of health records.[16] In one study, Google's AMIE research system ran on Gemini 2.5 Pro without medical fine-tuning.[21] Training data is often undisclosed, so real-world testing matters.[16]
How will AI change healthcare in the future?
AI will take on more administrative work, extend screening and specialist checks to more patients, and support clinical decisions under human oversight. In the UK, regulation is shifting from one-time approval to oversight throughout a tool's use, a change the government accepted in October 2026.[15] WHO Europe warns that without clear legal standards, clinicians may hesitate to rely on AI.[18]
