Proceeding report from Madrid conference on the responsible scale of AI in healthcare

2. AI in health: From potential to practice

Health is widely recognised as one of the most promising sectors for the application of artificial intelligence (AI), with potential to improve patient outcomes, clinical decision making, system efficiency, and health innovation while protecting public and population health. At the same time, healthcare has been slow to absorb and unlock the value of technological innovation at scale. As AI adoption accelerates – and generative AI is scaling three times faster than recent technology innovations (Exponential View, 2026[1]) – it remains uncertain how health systems will enable the transition from isolated pilots to a coherent ecosystem capable of delivering sustained system-wide benefits.

In its efforts to support the development of enabling foundations for AI in health, the OECD has analysed the early diffusion of AI across health systems. The evidence points to a sector at an early but pivotal inflection point, where the potential gains are significant, but system readiness remains uneven:

• OECD projections estimate that AI alone could result in a 6% productivity gain in healthcare over a ten‑year period in OECD countries (Filippucci, Gal and Schief, 2024[2]).

• Up to 97% of health data remains unusable for clinical decision making, limiting the potential for data-driven care (Anderson and Sutherland, 2024[3]).

• Health providers are not worried about AI replacing their jobs (70%) but are worried about AI being deployed without their input (72%) (Almyranti et al., 2024[4]).

• Less than one‑third of OECD countries have taken action to strengthen foundations or human capacity to use AI in health (OECD, 2026[5]).

• Seven OECD countries have published specific strategies at the intersection of AI and health, recognising the importance of health-specific actions for the adoption of AI (OECD, 2026[5])

• AI adoption remains concentrated in operational and task-specific applications, with all surveyed OECD countries using AI for administrative automation and 95% for imaging diagnostics. By contrast, only 55% report using AI for predictive analytics, risk stratification, or personalised treatment planning, where the potential to influence clinical outcomes may be greater (OECD, forthcoming[6]). Evidence from the European Union similarly suggests that AI uptake remains uneven and often localised, constrained by fragmented and non-interoperable health data, uneven digital maturity and limited real-time data access (OECD, 2026[7]).

These findings suggest that AI adoption in health is advancing, but along a fragmented and uneven trajectory that remains weakly connected to the foundations required for scale. This signals an opportunity to work together across countries and sectors to accelerate action in a way that facilitates the scale of AI.

There is urgency to act: OpenAI estimates suggest that 230 million people are already using AI for health-related queries (OpenAI, 2025[8]), and 35% of the public report using AI for health purposes before consulting a doctor (Edelman Trust Institute, 2026[9]). Yet evidence of value remains limited with 90% of firms adopting AI report no change in productivity (Yotzov et al., 2026[10]). Collective action is necessary to move health systems beyond promising pilots towards effective scale, ensuring that AI delivers measurable value for individuals and populations and remains grounded in human needs (The Holy See, 2026[11]; Anderson and Sutherland, 2024[3]).

What AI in health can do: Early evidence of benefits

Beyond AI’s growing adoption in task-specific and lower-value functions, emerging evidence shows measurable economic and human value at defined points along the care pathway. While still uneven, early deployments demonstrate impact across administration, diagnostics, system operations, population health, and research. These non-exhaustive examples illustrate both the breadth of current applications and the concentration of evidence in specific domains of healthcare delivery (see Table 1).

Table 1. Demonstrable value from reported AI in health use cases

Domain	Illustrative Use Case	Description of value
Administrative and workflow automation	Ambient Voice Technology (AI Scribes)	AI-powered documentation tools reduce administrative burden, increase clinician-facing time and improve workflow efficiency. Early evidence suggests reductions in documentation time and clinician burnout, while emerging evidence of return on investment indicate improved productivity and coding completeness (Holmgren et al., 2026[12]; Rotenstein et al., 2023[13]; Shah et al., 2024[14]; Olson et al., 2025[15])
Clinical Decision Support (diagnostics)	Medical imaging analysis (radiology, pathology, ophthalmology)	AI models trained on large imaging datasets can detect anomalies such as tumours, fractures and retinal disease with accuracy comparable to or exceeding specialist review, while processing images in seconds rather than days. Value is strongest in prioritisation and second-reader roles rather than full automation (Liu et al., 2019[16]; Hassan et al., 2024[17]).
Health system operations and capacity management	Use of health resources (bed management, demand forecasting, patient flow optimisation)	AI-driven demand forecasting and bed management tools allow hospitals to anticipate patient flows, reduce bottlenecks, and allocate staff and resources more efficiently. In some health systems, predictive scheduling has reduced average length of stay and cut avoidable readmissions by identifying patients at risk of deterioration before discharge (Alnsour et al., 2023[18]; Rampengan, Permana and Wuisan, 2025[19]).
Population health and prevention	Risk prediction and screening prioritisation	AI models integrating clinical, demographic, and behavioural data can identify population segments at elevated risk of disease before symptoms emerge. Value is highest in stratified screening and preventive care targeting, though external validity varies across systems (Rajkomar, Dean and Kohane, 2019[20]; Topol, 2019[21]; Antegenes, 2024[22]; Eisemann et al., 2025[23]).
Biomedical R&D and drug discovery	Target identification, clinical trial design	AI is compressing the drug discovery timeline from an average of 10 to 12 years to a fraction of that in specific stages. Prediction of protein structures, previously a bottleneck requiring years of laboratory work, is the most cited example. More broadly, AI tools are accelerating target identification, molecule screening, and clinical trial design, with direct implications for rare disease research where commercial incentives have historically been insufficient to sustain long development cycles (European Medicines Agency, 2026[24]; Jumper et al., 2021[25]).

These use cases demonstrate AI’s potential in providing value at both individual and population levels, but the benefits remain localised, unevenly distributed, and weakly connected to system-wide transformation. As a result, productivity gains from digitalisation in health continue to lag behind other sectors (Filippucci, Gal and Schief, 2024[2]). The core challenge is therefore not the absence of viable solutions, but the limited ability of health systems to scale them consistently across settings and populations.

The OECD AI in health Expert Group, formed in May 2024, has emphasised that the central constraint is not technological potential but system-wide diffusion at scale, including across borders. Building on this finding, the next section examines the structural barriers that limit scaling and sets out the foundations required for co‑operative action for the effective scaling of AI in health.

Why are AI advances not yet translating into better health outcomes at scale?

Despite the increasing evidence of AI’s potential in improving health outcomes, the benefits of AI for patients, healthcare providers, and health systems remain limited in scale and unevenly distributed. This is not primarily a technology problem. It reflects a more fundamental mismatch between the logic driving AI investment in health and the conditions required to generate sustained improvements in care. Addressing this mismatch requires moving beyond a simple “top-down” versus “bottom-up” view of AI governance and understanding scale as a process of multilevel orchestration across policy, institutions, developers, providers and users (Cresswell et al., 2026[26]).

Investment in AI in health has largely followed a technology-driven growth model, premised on the assumption that advancing technical innovation will generate commensurate economic and rapid productivity gains. This approach draws on experience from sectors such as ICT, telecommunications, finance and professional services, where AI exposure is relatively high and productivity levels and growth have historically been stronger. However, these experiences may not readily translate to health, where the conditions for adoption are less standardised, the pathways from deployment to impact are longer and less linear, and the organisational and human foundations needed to absorb and effectively use technology are more complex (Filippucci et al., 2024[27]).

Applied to healthcare, that model reveals a fundamental limitation: it conflates technology diffusion with system transformation, which challenges how use of technology supports productivity gains in health (World Economic Forum, 2026[28]). In health, value is unlikely to arise from inserting novel AI tools into unchanged care pathways. Productivity and outcome gains depend on whether health systems are able to redesign workflows, adapt organisational and policy settings, and build the change leadership needed to absorb innovation. This cannot be left to AI developers alone: it requires change‑ready health systems that work with developers to identify what is technically possible and translate it into new models of care that improve outcomes, experience and sustainability.

The productivity paradox in healthcare: Why more AI is not always better

The assumption underlying most AI productivity arguments is borrowed from manufacturing or financial services logic: if a tool allows the same input to generate more output, value is created. In healthcare, this translation breaks down not because the efficiency gains are absent, but because efficiency and outcome are not the same thing, and health systems have spent decades optimising for the former while struggling to measure the latter.

Among the most studied applications, Ambient Voice Technology (AI scribes) generate the most return-of-investment evidence in the literature and across countries (see Table 1). Such a tool allows clinicians to document faster and see more patients per day with expected productivity gains. However, it produces more encounters, but not necessarily better care. The relationship between volume and outcome in health is not linear: beyond a certain threshold, the literature on physician workload and diagnostic error suggests it turns negative (Rotenstein et al., 2023[13]; Slawomirski et al., 2025[29]). Throughput is improved but not necessarily health status.

This is consequential because clinician time is already highly constrained. While health encounter has been estimated to least five minutes or less on average (Irving et al., 2017[30]), AI-enabled time savings will therefore not automatically translate into better care. In many systems, clinicians are evaluated on patient experience while operating under volume‑based targets. Whether AI creates more time for patients or is absorbed into additional activity will depend on the incentives, metrics and accountability arrangements that shape how productivity gains are used.

Evidence from early adopters suggests that while AI scribes may reduce clinician burnout and after-hours documentation time (Tierney et al., 2025[31]), their system-wide financial impact remains uncertain and is increasingly associated with higher overall costs rather than savings (Peterson Health Technology Institute, 2025[32]). By reducing documentation time per encounter, AI scribes may therefore enable higher patient throughput and more detailed coding, expanding the volume and complexity of reimbursable activity. Practically, this reduction in documentation time may come from reduction in time for documentation that occurs outside of regular working hours. What appears as productivity improvement at the level of clinical workflow can thus translate into upward pressure on total system expenditure by increasing reimbursable activity while reducing after-hours documentation time.

What makes healthcare genuinely different from most sectors is not simply that outcomes are hard to measure, but what demand in healthcare actually represents. In most markets, higher demand reflects a proxy for value creation. A health system generating more consultations, more prescriptions, and more procedures may be responding to genuine need, but that need is itself partly a product of what the system failed to do upstream. More activity is not evidence of more value, but it may be evidence of accumulated failure. The right productivity measure for health is therefore not output or volume of care per clinician hour but progress towards improving health’s Quintuple Aim: better patient outcome and experience, improved provider well-being, sustainable system expenditure, and benefits accessible to all. AI has yet to demonstrate, at scale, that it moves that needle (Tierney, Lee and Liu, 2026[33]; Weeks, Lavista Ferres and Weinstein, 2024[34]). This gap reflects, in part, the time required for effects to emerge. The components of the Quintuple Aim change over different horizons: provider well-being and patient experience may shift within months, whereas patient outcomes and system expenditure may take years to change. Evaluations of AI will need to consider these time horizons, while also respecting demands for return on investment.

Beneath all of this sits a measurement problem. Health systems are considerably better at measuring what AI improves such as speed, accuracy, and volume processed rather than what it changes: patient agency, quality of life, avoidable deterioration, long-term disease trajectory. A diagnostic algorithm that detects a condition earlier will show up in detection rates. Whether that detection translates into better survival or reduced system burden years later is far harder to capture, and most health information systems are not built to do it. AI adoption gets evaluated on the metrics it improves most visibly and rapidly, which are not always the ones that matter most. Until outcome measurement catches up with AI capability – including through indicators of access, care co‑ordination, patient outcomes, and broader system performance – the productivity paradox will persist, not because AI is failing to deliver, but because the systems for determining whether it is delivering remain incomplete.

These risks do not suggest that AI is incapable of delivering better outcomes at scale. Instead, they demonstrate that the conditions for success are more demanding than the standard productivity narrative implies. The central challenge for healthcare is not whether AI can make clinicians faster, but whether health systems can convert those efficiency gains into measurable improvements in health.

In that context, a central question emerges: what practical actions and enabling conditions are needed to scale health AI in ways that improve health outcomes, well-being, and experience, while fostering innovation and managing health system expenditures. This question anchored discussions at the Madrid AI in health Conference, co‑organised by the Spanish Ministry of Health and the Organisation for Economic Co-Operation and Development (see Section 3).

Beyond technology: Why AI struggles to scale in health

Integrating AI into healthcare is increasingly less a question of technological capability than of institutional readiness. AI capabilities and investment are advancing rapidly, and adoption is expanding beyond clinical and administration functions. Yet healthcare has demonstrated that technological deployment alone does not generate system transformation, and the pathways through which AI creates value may differ across health system models and institutional contexts.

A similar pattern has been observed in the diffusion of informatics in the 1980s, where productivity gains were delayed for decades after widespread adoption, materialising only when complementary investments in skills, organisational redesign, and process innovation allowed firms to absorb and operationalise the technology (Brynjolfsson, Rock and Syverson, 2021[35]; Capello, Lenzi and Perucca, 2022[36]). Previous waves of digitalisation produced substantial investment in electronic health records, digital prescribing, and health information systems, but often failed to deliver proportional improvements in productivity, interoperability, or patient outcomes because organisational, governance, and incentive structures evolved more slowly than the technologies themselves. AI risks following the same trajectory: rapid deployment into systems that remain structurally unprepared to capture its full value.

Prior to the Madrid Conference, expert consultations identified systemic constraints and barriers to scaling AI in health:

• Stretched health workforce sceptical of AI use: Workforce scepticism reflects structural strain rather than technological resistance alone. The OECD projects a shortage of 8 million health workers by 2040, with nursing accounting for nearly 74% of the gap (OECD, forthcoming[6]). In this context, AI may be introduced into systems already burdened by staffing shortages and previous digital transformations that increased administrative workload and did not meet the needs of clinicians and patients on the ground. Resistance therefore reflects accumulated operational fatigue and concerns that new tools will add complexity, rather than opposition to innovation itself (Almyranti et al., 2024[4]). Concerns around liability compound this scepticism. Clinicians may be held liable both when they follow an AI recommendation that results in harm and when they decline a recommendation that would have prevented it; neither the quality of the tool nor the soundness of the clinician’s reasoning provides certain protection (Mello and Guha, 2024[37]). Where tools generate frequent alerts of low predictive value, clinicians bear both the workload imposed by those alerts and the risk associated with the rare correct alert that is overridden.

• Fragmented accountability without a clear sense of direction: Fragmented accountability creates a co‑ordination problem that few health systems have successfully resolved. AI implementation frequently requires action across ministries, regulators, payers, providers, technology vendors, and professional bodies, yet responsibility for achieving system-wide outcomes is rarely assigned to a single institution. What emerges is a governance vacuum in which many actors influence success, but none are accountable for it.

• Trust deficits in AI-driven systems. Trust deficits across patients, clinicians, and institutions represent a cross-cutting barrier to AI adoption in health, closely linked to gaps in accountability. At the patient level, concerns about data use, reliance on algorithms, and loss of human judgment reduce willingness to share data, potentially limiting the availability and diversity of data needed to develop, validate, and deploy AI systems effectively. At the institutional level, many AI tools are validated in controlled settings that do not reflect real-world clinical conditions, leaving unresolved questions around liability, performance, and governance under existing regulatory frameworks. Trust is self-reinforcing: well-governed deployments accelerate adoption, while failures such as poor quality outputs, errors, or data breaches can rapidly erode confidence and are difficult to rebuild. Trust therefore functions as institutional capital: it accumulates slowly but can be depleted quickly. Since AI systems depend on continuous data access, clinician engagement, and public legitimacy, trust is not a soft constraint but a core condition for system performance.

• Inconsistent regulatory and policy landscape not adaptable to change: Regulatory frameworks are increasingly struggling to match the speed and iterative nature of AI development (Singh, Zhou and Auclair, 2025[38]). While AI models can be updated within weeks, regulatory adaptation often occurs over years (OECD, 2026[5]). The result is a growing mismatch between technological and institutional cycles, creating uncertainty for developers, providers, and patients alike. In many jurisdictions, overlapping frameworks across digital health, data protection, medical devices, procurement, and liability further compound complexity (Aldhalimi, 2026[39]). Some of these frameworks have a homogenous treatment of AI solutions that do not acknowledge the diversity of risk profiles. The result is not necessarily stronger safety, but fragmented governance that increases compliance burden while slowing diffusion of beneficial innovations.

• Data fragmentation that prioritises in-house value. Early digital health investment produced facility-centric data models with limited interoperability, stretching scarce resources across incompatible systems (OECD, 2026[40]). As a result, health data remains fragmented across incompatible systems, preventing AI models from being trained, validated, and deployed across settings. In practice, systems can generate data but struggle to convert it into scalable learning: models developed in one context rarely generalise to others, and each deployment effectively restarts the development cycle. The result is a data infrastructure that supports local optimisation but constrains system-wide learning and limits the cumulative returns on AI investment.

• Technological dependence and limited strategic capabilities. Scaling AI in health also depends on access to trusted and resilient digital capabilities, including secure infrastructure, computational resources, foundational models, procurement mechanisms and technical expertise. Where these capabilities are concentrated outside the health system, countries may face new dependencies that affect sovereignty, resilience and sustainable scale. Health systems will therefore need to balance international collaboration with the development of trusted national or regional capabilities that can support effective AI adoption over time.

• Misalignment between innovation incentives and health system priorities. Much of the estimated USD 18 billion investment in health AI is driven by technical feasibility and commercial opportunity rather than by population health needs (World Economic Forum, 2026[28]). As a result, development is concentrated in areas that generate measurable short-term returns such as diagnostic imaging, administrative automation, and revenue cycle management, while domains with potentially greater health impact, including prevention, mental health, primary care co‑ordination, and multimorbidity management, remain comparatively underserved. In the absence of clearly articulated demand from health systems, developers often define the problems to be solved rather than seeking the needs of health systems. The result is not a shortage of innovation, but a mismatch between where innovation is concentrated and where disease burden and system need are greatest. This risks directing scarce investment toward solutions that scale commercially while delivering limited system-wide health gains.

• Misaligned incentives favouring standalone tools over shared foundations. Procurement processes prioritise short-term cost savings, while reimbursement mechanisms continue to reward activity rather than outcomes. As a result, efficiency gains generated by AI are rarely reinvested in the data infrastructure, workflow redesign, workforce capabilities, and governance arrangements needed to support scale. Health systems therefore tend to adopt standalone tools such as AI scribes or diagnostic algorithms at the facility or departmental level, without building the interoperable foundations that benefit integration and reuse. The result is a proliferation of pilots and local successes, but limited system-wide transformation.

Continuing to focus policy efforts solely on addressing each of these barriers independently may impose a higher burden on health systems and populations. Instead, these challenges suggest that health systems are at an inflection point, where the use of AI is being normalised and focus is shifting from promising pilots to creating the conditions that accelerate effective scale.

Collective action to design mutually reinforcing activities

The OECD Expert Group on AI in health identified the effective scale of AI in health as a key goal, and collective action as a condition to reach that goal. The AI in health policy checklist provides a living compass for identifying blind spots in the adoption of AI at national and international levels (OECD, 2026[5]) and sits within a growing landscape of frameworks produced by international organisations, governments, think tanks, not-for-profit organisations, and consulting firms (WHO, 2021[41]; WHO, 2026[42]; European Union, 2024[43]; ISO, 2023[44]). That fragmented landscape signals the need to align practices across organisations: the governance challenge is recognised, the tools are multiplying, and yet the gap between AI’s demonstrated potential and its system-wide impact continues to widen.

What is now needed is to move beyond guidance and provide a clear signal to countries to enable the scaling of health AI, anchored in collective action. The extensive literature on AI in health has established what works in controlled settings and identified systemic barriers to scale. What it has not yet produced is compatibility and shared learning on human capabilities and technical capacities across countries, institutions, and sectors to drive action toward common trust, data foundations, governance standards, and accountability frameworks that would allow AI to deliver value within and across borders.

Other sectors offer instructive precedents for governing complex, high-stakes systems at scale. The global banking system demonstrates how fragmented oversight and misaligned incentives can generate systemic risk in highly interconnected environments, while air traffic control shows how safety and efficiency can be maintained through interoperable standards, co‑ordinated governance, and real-time information sharing across jurisdictions. Both sectors illustrate that scalable innovation depends not only on technology, but on shared rules, aligned incentives, and collective mechanisms for managing risk.

Healthcare shares characteristics of both systems but operates under unique constraints. Like banking, it is vulnerable to fragmentation and systemic risk; like air traffic control, health relies on co‑ordination, interoperability, and shared situational awareness to function safely at scale. At the same time, health AI must balance three objectives simultaneously: accelerating innovation, protecting patient safety, and ensuring access to benefits for all. Realising AI’s potential will therefore depend on governance, infrastructure, and incentives that align these objectives rather than treating them as competing priorities.

In efforts to increase harmonisation across international frameworks and actively pursue co‑operation in designing activities that are mutually reinforcing, the OECD and the Spanish Ministry of Health co-hosted an expert-led AI in health Conference in Madrid on 28‑29 May (OECD, 2026[45]), bringing together experts from across countries and partner organisations. The conference was designed to define a shared agenda for action; one that draws on lessons from other sectors in navigating the tension between speed, safety, and scale, and translates the existing body of knowledge into the collective steps that only co‑ordinated action can deliver.

The next section summarises the Madrid Conference outcomes which led to the creation of the Expert-Led Madrid Action Plan (EL-MAP).