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

4. Achieving collective action through the Expert-Led Madrid Action Plan

Why co‑operative action matters now

From work prior to and at the Madrid Conference, there was significant consensus among experts that the Action Plan needed urgency to address the thematic challenges with concrete actions. Those concrete actions should facilitate the achievement of a common vision and be guided by common principles. Experts in attendance at the Conference expressed consensus on the vision, principles, and action plan.

The urgency among experts was described as follows:

Al is already transforming health. It helps to detect diseases earlier, improve treatments, ease workloads, streamline operations, protect public health, and stimulate innovation. It also brings new risks, challenges, errors, legal and privacy concerns, and unclear responsibilities. Acting together will support transformation that unlocks human benefits while earning trust.

A key theme for the Action Plan was the need to establish a clear signal for policymakers, industry, and civil society on the actions needed to enable the effective scale of AI in health. To date, the urgency to act has led to the development of multiple uncoordinated frameworks, leaving industry challenged to interpret them effectively and policymakers tasked with reviewing multiple approaches to determine which are best suited to their context. This slows the scale of innovation locally and impairs the ability to scale innovation across borders.

With that in mind, experts determined that it was important to identify a clear and coherent signal for action for industry and policymakers. This signal would translate into a collection of leading practices for the effective scaling of AI while considering the trade‑offs associated with different approaches. The signal would also build on what already exists and works and only provide additional guidance where there are gaps in knowledge. The signal produced would not be prescriptive for either policymakers or industry; however, it could inform country-led actions that also support international collaboration and simplify the ability to scale across borders.

Having established urgency and an approach to the strategy, experts agreed on a common direction for the desired signal. This would act as a guiding vision for the effective scale of trustworthy AI in health.

Expert agreement on a shared vision for AI-power in health

Effective and safe scaling of AI in health will transform health systems, resulting in:

• People‑centred health: People will be empowered to use their data and AI safely to achieve better health outcomes, supported by integrated healthcare teams using the most up-to-date knowledge.

• Health for everyone: Health systems will effectively scale trustworthy AI to advance people‑centred access to the full range of quality health services, ensuring all people receive the care they need, when and where they need it, without financial hardship.

• Dynamic action: Health professionals, researchers, health system management, patients and innovators will work together to match human needs with opportunity through trusted use of information and AI to protect and promote public health, improve the effectiveness of care and prevention pathways, and accelerate the translation of scientific discovery into action.

This vision will be achieved through collective action across organisations, governments, civil society, and industry working co‑operatively with mutual trust and respect for sovereignty.

This vision reflects that action will be designed to support better care for individuals and communities, while enabling analytical applications of AI for health system operations, public health management, and research and innovation.

Experts also agreed that action should follow agreed guiding principles. The OECD AI Principles, adopted in 2019 and updated in 2024 (2026[59]), include five value‑based principles:

• Inclusive growth, sustainable development, and well-being

• Human rights and democratic values, including fairness and privacy

• Transparency and explainability

• Robustness, security, and safety

• Accountability

In addition, the OECD AI Principles included recommendations for policymakers:

• Investing in AI research and development

• Fostering an integrated AI-enabling ecosystem

• Shaping an enabling interoperable governance and policy environment for AI

• Building human capacity and preparing for labour market transformation

• International co‑operation for trustworthy AI

Subsequent to the publication of the OECD AI principles, several organisations developed their own principles introducing sector-specific nuances. Experts at the Conference reflected on these principles.

Actions will be Guided by Common Principles in Health

In 2024, HealthAI – the Global Agency for Responsible AI in Health synthesised the G20 (OECD AI, 2019[60]), WHO (WHO, 2021[61]), OECD AI (OECD, 2026[59]) and World Bank (World Bank, 2025[62]) principles (among others) to identify a common set of nine principles. These are functionally equivalent to the OECD principles, applied to the design and deployment of health AI (HealthAI - the Global Agency for Responsible AI in Health, 2025[63]). These are synthesised here:

• Promote human and societal well-being

• Respect fundamental rights and human dignity

• Preserve human agency and enhance lifecycle oversight

• Establish transparent and understandable governance and processes

• Foster clear accountability and proportionate responsibility mechanisms

• Optimise and improve health outcomes while minimising harm

• Ensure fair and secure access for robust and safe outcomes

• Consider long-term human and financial impacts

• Promote participatory governance across diverse stakeholders

Experts, building on the OECD AI principles (2026[59]) for International co‑operation, Sustainable development and an AI-enabling ecosystem, agreed that additional principles were important to include, given recent experience with AI in health and the importance to develop pathways for safe and effective scale.

• Respect national sovereignty and value international collaboration

• Leverage what works and scale fiscally sustainable solutions

• Invest in data foundations that are accurate and reusable, for AI that is both safe and scalable

Expert-Led Madrid Action Plan (EL-MAP)

The above are foundations for the Action Plan, to support its collaborative development and ongoing evolution. Below are the actions in EL-MAP, which address the challenges identified through the course of the Conference. Each action is described with several components:

• Rationale: Description of challenge to be addressed by the action and why the action will help.

• Possible outcome: Possible deliverable based on the rationale.

• Sample existing initiatives: Examples of relevant initiatives shared by participants.

The actions below are organised by three key themes identified by experts (summarised in Section 3) to guide collective action:

 Establish and sustain trust by understanding where AI could be used, how health systems will be re‑oriented around AI, and what competencies and capacities are required among the public sector, private sector, and civil society.

• Strengthen foundations by simplifying the path from AI experimentation to system-wide scale. Pilots can be delivered in isolation; however, achieving the benefits of AI in health at scale depends on integrating AI solutions into the digital health ecosystem. The digital health ecosystem is described by foundations of policy, data and digital infrastructure, understanding of AI life cycle management, adaptability for future change, and practices for oversight of integration of AI in health.

• Develop economic and financial models by estimating the economics of AI in health, drawing from broader cross-border and cross-industry experience, as well as describe leading practices to incentivise the effective adoption, scale, and use of AI in health solutions.

Underlying these actions is an additional key supporting action for measurement and co‑ordination of the readiness, use, and impact from the adoption of AI in health.

Figure 1. Identified key priority themes for collective action

Actions to establish and sustain trust

 

ACTION 1: Co-design use cases for AI-powered, person-centric learning health systems (Co-design of use cases)
Rationale: To date, AI in health has largely focussed on technology-driven innovation. That is, the innovation is developed first and then health systems see how to integrate the innovation into their systems. This action would shift the focus toward needs-based innovation, where AI use cases are articulated and aligned with linguistic, demographic and disease‑burden needs, as well as priorities for improving population and patient outcomes, provider experience, or other core measures of health system effectiveness.
Possible outcome: Document “typical” use cases where AI may be employed to deliver better care, research and innovation in health. Includes for example use cases for patient care (including direct patient use of AI) and use cases for personal decision support, clinical decision support, administration, health system management, population health protection, and innovation. Considers roles across the use case including accountability, consent, and communication. Once use cases are defined, action could be taken to focus on the core use cases that generate the greatest benefit to individuals and populations with the lowest risk, and appropriate readiness to develop, adopt, and scale. Could include the development of “catalogues” or “maps” that simplify discoverability of AI initiatives for leverage and re‑use.
Sample existing initiative(s): ‒ Although not health-specific, the United Kingdom’s AI Growth Lab provides a model for testing AI use cases through collaboration between users, regulators and developers before large‑scale deployment (Gov UK, 2025[64]). ‒ The first large‑scale mapping of AI initiatives across publicly funded healthcare in Sweden, Norway, Denmark, Finland, and Iceland. The framework categorises over 800 AI initiatives into eight use cases and three stages: concept, pilot, and routine operation (AI Sweden, 2026[65]). ‒ WHO/Europe mapped the use of AI in healthcare across 50 WHO European Region Member States, including common AI applications and use cases (WHO Regional Office for Europe, 2025[66]). ‒ SmartCHANGE is a Horizon Europe project developing trustworthy and explainable AI tools to predict long-term cardiovascular and metabolic risks in children and adolescents. It illustrates a person-centred AI use case that combines participatory design, privacy-preserving methods, explainability and real-world validation (European Commission, 2023[67]). ‒ France conducted a comprehensive mapping exercise of AI use across public administrations, reviewing existing AI initiatives, governance arrangements, use cases, barriers and productivity impacts through surveys, interviews and benchmarking with foreign administrations and the private sector. The exercise included a dedicated assessment of the health and social sector, examining how regional health agencies, social security institutions and healthcare providers are using AI to support administrative, operational and service‑delivery functions (Cunéo et al., 2026[68]) ‒ Brazil’s Ministry of Health has launched a national mapping exercise of AI initiatives across the health sector. By classifying initiatives according to health domain, technical approach, readiness level, risk, data use, oversight and financing, it provides a replicable method to establish a baseline of AI readiness and use, helping inform where AI can be aligned with health system needs and scaled effectively.
ACTION 2: Co-develop competencies and capacity to work with AI (Capacity building)
Rationale: Users of AI may be overwhelmed by the need to become more literate in many areas. The objective should be to increase literacy to a level that is sufficient for the specific AI use case and adapted to the professional role and user needs. It follows that AI developers should design their applications to meet people at their level of literacy.
Possible outcome: Document baseline “literacy” requirements for AI, data, digital, health for each of (1) the public and patients, (2) providers, (3) policymakers, (4) payers, (5) regulators, and (6) institutional leaders. This “literacy” baseline would be tuned to their role and emerging new roles to enable the effective scale and use of AI in health for prioritised use cases (see Action 1). Consider methods for delivering “literacy” programmes, ongoing reviews and updates of baseline to ensure optimal outcomes and minimal harm. Could work with industry to design solutions to reduce literacy burden on users, such as encouraging usability testing with intended users prior to deployment
Sample existing initiative(s): ‒ The University of Mons’ AI in Belgium developed a Digital Medicine programme that embeds AI training within medical education to strengthen health professionals’ ability to evaluate and responsibly apply AI tools in clinical practice (Pizzolla et al., 2023[69]). ‒ Data Saves Lives (DSL) is a multi-stakeholder initiative and led by the European Patients’ Forum (EPF). Its aim is to raise patient and public awareness of the importance of health data, improve understanding of how health data is used, and foster effective use and good practices across Europe to ensure that AI systems are trained on high-quality health data (Data Saves Lives, 2025[70]). ‒ The European Patients’ Forum AI Survey Report 2026 provides evidence on patients’ AI literacy needs, trust requirements and communication preferences, helping inform patient-focussed education and capacity-building efforts for the effective use of AI in healthcare (The European Patients’ Forum, 2026[71]). ‒ The IEEE Standards Association’s Healthcare Professionals (HCP) AI Factsheet initiative aims to improve transparency and trust in AI in health AI by providing healthcare professionals with clear, standardised and clinically meaningful information that enables them to assess the intended use, capabilities, limitations and risks of AI tools (IEEE Standards Association, 2026[72]). ‒ Korea is strengthening AI literacy and workforce capacity across healthcare, government, industry and the wider public. Initiatives include practical AI training for public officials, interdisciplinary medical-AI education for specialised professionals, and a planned National AI Health Assistant to help patients better understand prescriptions, clinical records and medical terminology (HealthcareITNews, 2026[73]). ‒ Complementing the Slovenian AI Factory (SLAIF), the Slovenian Competence Center for Artificial Intelligence (KCUI) supports safe and effective AI adoption by providing a national entry point for AI knowledge, training, regulatory and technical advice, and practical use cases (KCUI, 2026[74]). ‒ Belgium’s public health institute, Sciensano, has established an internal AI Board to strengthen organisational capacity for AI adoption. The Board promotes AI literacy through training and supports the development of AI skills across the organisation (Sciensano, 2026[75]).
ACTION 3: Co-facilitate multi-party co‑operation and partnerships (Collaboration and partnerships)
Rationale: Scaling AI in health is a significant task requiring many actors to collaborate toward a common goal. Establishing blueprints for structures that facilitate co‑operation will help remove roadblocks to scale and prevent wasteful (and potentially harmful) duplicative activities.
Possible outcome: Documents baseline leading practices for co‑operation and partnership across and between public sector, private sector, and civil society to advance evidence‑based innovation. Such partnership could share responsibility, risks, and value created while improving patient outcomes. Includes practices for non-competitive environments and/or collective impact to encourage co‑operation among legal entities and optimise common outcomes while minimising harm. Also includes leading practices for participatory engagement models with the public and health workforce across the full AI life cycle.
Sample existing initiative(s): ‒ The European Charter of Patients’ Rights provides a shared framework for patients, healthcare providers and policymakers around core principles of patient autonomy, transparency and participation, helping build the trust and common understanding needed for patient-centred innovation, including AI-enabled healthcare (European Commission, 2002[76]). ‒ The proposed EU Charter of Digital Patients’ Rights illustrates how a shared rights-based framework can provide common principles for patients, healthcare providers, policymakers and technology developers, supporting trust and effective adoption of AI-enabled healthcare (van Kolfschooten, 2024[77]). ‒ Canada’s Declaration of Personal Health Data Rights provides a shared framework for patients, healthcare providers, governments, researchers and technology developers to align around principles for access, control, transparency and effective use of health data, supporting trust-based collaboration in data-driven health innovation (Patients' Annual Summit, 2026[49]). ‒ Under the AI Basic Healthcare Strategy, Korea is advancing collaborative pilot projects, including public healthcare AI platforms and an integrated emergency AI platform, to support joint implementation across ministries, public institutions and healthcare providers (Korea.net, 2026[78]).

Actions to strengthen foundations

 

ACTION 4: Identify leading practices to oversee and lead the integration of AI in health systems (Leadership)
Rationale: Most AI in health activities have been bottom-up, starting with pilots and running into difficulty with scaling. Less than 30% of OECD countries have implemented a form of formalised AI in health leadership. This action would identify leading practices (including from other sectors) where leadership has been established for the effective scale of AI solutions. This could include oversight of the evaluation of AI in health initiatives that consider safety, impact, and scalability to help focus efforts on scaling what works and adds human and economic value. It could also consider practices to enable proportionate assurance, oversight, and governance requirements appropriate for the risk profile of the AI solution.
Possible outcome: Documents leading practices for decision structures to support the effective scale of AI while doing no harm. Decision structures would clarify accountability for decision making and issue resolution including identification and resolution of common barriers to the effective scale of AI in health.
Sample existing initiative(s): ‒ The Oviedo Convention provides a legally binding human rights framework for the application of biology and medicine, establishing principles to safeguard human dignity, autonomy, informed consent, privacy, equity, and the protection of vulnerable individuals in healthcare. While not AI-specific, its principles provide an important foundation for the effective integration of AI in health by guiding approaches to human oversight, accountability, transparency and the protection of patients’ rights in AI-enabled care (Council of Europe, 1999[79]). ‒ Duke’s AI governance model demonstrates how health systems can build the institutional capacity needed to scale AI safely by embedding structured evaluation, accountability and monitoring throughout the AI lifecycle. This approach can support leaders in identifying valuable use cases, managing risks and building trust as AI becomes integrated into clinical care (Duke University, 2025[80]). ‒ The OECD’s Scaling Artificial Intelligence in Health report provides a strategic framework for moving from experimentation to effective scale, identifying key enabling conditions at the system level, including governance, data access, workforce capacity, evaluation and trust (OECD, 2026[5]). ‒ Korea’s National AI Strategy Committee supports cross-government co‑ordination on AI policy, while the AI Basic Healthcare Subcommittee brings together relevant ministries and public institutions to guide healthcare AI scale‑up (Republic of Korea, 2026[81]). ‒ Brazil has consolidated national leadership for digital health and AI within the Secretariat of Information and Digital Health (SEIDIGI) of the Ministry of Health, which holds a formal mandate over health data governance, interoperability, telehealth, data integration and the Ministry’s AI policy (Ministério da Saúde Brazil, 2025[82]). ‒ Although not specific to health, the Australian National AI Centre (NAIC) illustrates how a national leadership mechanism can support AI adoption through best-practice guidance and direct engagement with businesses, not-for-profits, social enterprises and First Nations organisations (Australian Government, 2026[83]).
ACTION 5: Develop baseline policies that support the effective scale of AI in health (Policy)
Rationale: Innovators report challenges scaling AI in health across health facilities, regions, and countries due to inconsistent and poorly understood policies. This action will help the scale of AI in health solutions by providing greater clarity for innovators (and policymakers) for key policies. Further, this would help policies to be more compatible when scaling solutions within and across borders. This would also improve adoption by front line workforces with greater clarity for their individual accountability.
Possible outcome: Documents baseline guidance for the effective and safe scale of health AI across key policy areas for prioritised use cases (see Action 1). This could include areas such as (a) common language for policymaking, (b) risk classification of AI solutions, (c) medical device regulation, (d) health technology assessments, (e) testing and independent evaluation of AI solutions, (f) procurement, (g) model cards, and (h) legal frameworks – including the treatment of liability – among others. Guidance should recognise that not all AI is the same and be tuned based on context, as well as should be built on what already exists and is working.
‒ Sample existing initiative(s): ‒ CHAI’s Applied Model Card illustrates how baseline governance requirements can support effective AI scale by ensuring that health AI applications are accompanied by structured information on their purpose, risks, performance and appropriate use. Such common documentation standards can facilitate procurement, oversight and comparison of AI solutions while strengthening trust among providers, patients and regulators (CHAI, 2026[84]). ‒ Ukraine’s AI and Blockchain Sandbox, as part of the first large‑scale application of the HUDERIA methodology, illustrates how regulatory sandboxes can support the effective scaling of AI by enabling innovators and public authorities to assess risks, test governance approaches and develop practical tools before wider deployment (Council of Europe, 2026[85]). ‒ Although formal guidelines for conducting prospective, non-interventional AI evaluations (often referred to as silent trials) in healthcare settings are still lacking, silent trials provide a valuable approach for testing AI models in intended clinical environments without influencing patient care, generating evidence on local performance, safety, workflow integration and potential risks before deployment (Tikhomirov et al., 2026[86]). ‒ Health Data Governance Principles (HDGP) provide a globally recognised framework designed to support a rights-based approach to the management and use of health data. It provides operational guardrails that balance the necessary protections of individual data rights with the effective secondary use of data for innovation and AI deployment (Transform Health, 2026[87]). ‒ Informed by the Health Data Governance Principles, the Health Data Governance Legislative Blueprint provides a non-prescriptive resource to support the development of national and regional legislative and regulatory frameworks. It is complemented by a legislative and regulatory assessment tool (Transform Health, 2026[88]). ‒ Korea’s AI Basic Act establishes legal requirements for high-impact AI and reinforces principles on safety, trustworthiness, security and risk management. This is complemented by proposed legislation on digital healthcare and health information use, as well as HIRA guidance on how AI-enabled medical technologies should be assessed for coverage and reimbursement (Information Technology & Innovation Foundation, 2025[89]). ‒ The Clinton Health Access Initiative is advancing collaborative work across five core deliverables to facilitate co‑operation with a focus on LMICs. These are: (1) a data commons marketplace with federated learning, (2) a patient simulator for benchmarking and post-deployment monitoring, (3) paired procurement playbooks for governments and innovators, (4) a clinical decision-support “skills” repository, and (5) standardised product cards (analogous to model cards) for AI products (Clinton Health Access Initiative, 2026[90]).
ACTION 6: Develop baseline policies, processes, and approvals for the full AI in health lifecycle (Life Cycle)
Rationale: While there are documented AI in health lifecycles, these have not consistently been integrated into business and development practices. There is opportunity to describe the range of expectations at each stage of development of the AI in health lifecycle. This would be designed to optimise adoption while protecting safety. A life cycle could be developed that would be adaptable by country that defines responsibilities, considers risk levels and approval criteria, incorporates change management, and establishes minimum monitoring requirements throughout.
Possible outcome: Documents baseline expectations across the AI in health lifecycle, building on leading practices. Expectations would support trustworthiness for users and payers and could include efficient re‑use of common resources (e.g. digital public infrastructure) to realise economies of scale from common investments. Lifecycle components could include (a) design and development of AI algorithms and tools, (b) independent evaluation and deployment for effectiveness and scale, (c) post-deployment monitoring, and (d) update of AI algorithms from continual learning among others.
Sample existing initiative(s): ‒ CERSI-AI (The Centre of Excellence for Regulatory Science and Innovation in AI and Digital Health) has been developing a standard model of the AI lifecycle from design to retirement. CERSI-AI is working with innovators, regulators, patients and the NHS to optimise the regulation of AI and Digital Healthcare to accelerate innovation that will improve people’s lives (CERSI-AI, 2026[91]). ‒ Korea’s Integrated Review and Assessment System for Innovative Medical Devices provides a concrete example of lifecycle management for AI-enabled medical technologies, linking early clinical access with real-world evidence generation and subsequent evaluation. Technologies recognised as safe and potentially valuable may enter clinical practice through temporary listing arrangements, with accumulated evidence later assessed through the New Health Technology Assessment before proceeding to standard National Health Insurance listing (HIRA, 2026[92]). ‒ The European Testing and Experimentation Facilities initiative supports the AI lifecycle by providing testing, validation, regulatory guidance and experimentation services for health AI and robotics in realistic healthcare settings, helping innovators move from proof of concept to safe clinical deployment at scale (TEF-Health, 2026[93]).
ACTION 7: Strengthen pre‑competitive data and digital infrastructure with protection to support scale (Infrastructure)
Rationale: Data and digital infrastructure are the backbone of 21st century health systems and a pre‑competitive foundation for scaling AI in health. Artificial Intelligence requires timely access to quality data and sufficient digital infrastructure for training, validation and deployment. Similarly, the operation of AI in production requires quality data inputs, secure hardware to operate AI algorithms, and digital connections to ensure that the right information is available to the right people, at the right place and time. Strengthening these shared foundations can support all innovators – from small enterprises to large corporations.
Possible outcome: Documents leading practices for developing and governing pre‑competitive and digital infrastructure that enables the effective use of health data and digital infrastructure while assuring proportionate protections are in place. This would cover the range of accountability related to health data governance, (semantic and technical) interoperability, data quality, secure data access, cyber security, infrastructure management and value‑sharing with a focus on common enabling foundations that support innovation across the ecosystem.
Sample existing initiative(s): ‒ The OECD Recommendation on Health Data Governance provides a framework for building trusted health data ecosystems by promoting privacy protection, secure data access, interoperability, transparency and effective governance arrangements. By supporting countries in establishing the institutional foundations for effective health data use, it helps create the enabling environment required for future data-driven innovation while maintaining public trust and protecting individual rights (OECD, 2016[54]). ‒ The European Health Data Space (EHDS) illustrates how shared digital infrastructure can enable large‑scale data use by establishing common rules, standards and mechanisms for electronic health data access and exchange across Member States. By reducing fragmentation and strengthening citizens’ control over their data, the EHDS creates foundations for future data-driven applications, including AI, while ensuring appropriate safeguards (European Commission, 2025[56]). ‒ Canada’s Connected Care for Canadians initiative illustrates how national action can address fragmentation in a federated health system by improving interoperability and enabling more seamless exchange of health information across provinces and territories. By strengthening the connectivity of health data ecosystems, the initiative supports more co‑ordinated care and creates conditions for effective secondary use of health data for research, innovation and future AI applications (Government of Canada, 2026[55]). In this context, the pan-Canadian health data platform, VITAL, is building the infrastructure needed to connect clinical data from hospitals across multiple provinces, enabling AI-driven discovery and advancing precision medicine (Government of Canada, 2026[94]). ‒ Brazil’s National Health Data Network (RNDS) provides a national interoperability platform to connect health information systems across the country. Decree No. 12560/2025 further strengthens this foundation by formalising governance, interoperability and data-sharing arrangements, creating enabling conditions for more integrated care and the effective scaling of AI in health (Presidency of the Republic of Brazil, 2025[95]). ‒ The United Kingdom’s Cyber Security and Resilience Bill demonstrates how governments can strengthen the reliability of digital infrastructure by establishing baseline cybersecurity, risk management and incident reporting requirements. Such approaches are essential to ensuring that increasingly connected health information systems remain secure and resilient as digital health services expand (Department for Science, 2024[96]). ‒ Spain’s Sectoral Data Spaces Programme supports the development of trusted data-sharing infrastructures and governance mechanisms across strategic sectors, including health. By funding data-space demonstrators and associated use cases, the programme seeks to strengthen interoperability, data access, and secure data reuse, creating the digital foundations required for AI-enabled learning health systems (Ministerio para la Transformación Digital y de la Función Pública, 2026[97]). ‒ Korea is strengthening the data and computing infrastructure needed to scale AI in health, including through the planned Public Healthcare AI Highway and a National Health Data Hub integrating insurance claims and clinical data (Seoul National University Hospital, 2026[98]). ‒ Slovenia’s AI Factory (SLAIF) provides shared national AI capacity, including high-performance computing, data access, tools, training and support services for researchers, public institutions and industry. Its health vertical illustrates how common digital and compute infrastructure can support AI development and scale across the health ecosystem (Slovenian AI Factory, 2026[99]).
ACTION 8: Enable effective adaptability to emerging capabilities (Adaptability)
Rationale: Health systems would benefit from considering artificial intelligence as an emerging capability, rather than a singular one‑off innovation. This calls for putting into place policies and practices that can review and respond to those innovations at an appropriate pace to realise benefits while minimising predictable harms.
Possible outcome: Documents approaches for systemic review of emerging capabilities to support scale, for example, regulatory sandboxes and horizon scanning to assess the likelihood, impact and timing of emerging technologies (including quantum computing and genomics), and their people, policy, process, technical and cybersecurity implications. The aim is to optimise speed and value while minimising harm to individuals, communities and populations.
Sample existing initiative(s): ‒ Applied to healthcare, the OECD AI Principles provide a foundation for adaptive governance by enabling health systems to respond to rapidly evolving AI capabilities while maintaining safeguards for safety, accountability and human rights. Their risk-based and technology-neutral approach allows governance approaches to evolve alongside emerging clinical applications and new forms of AI-enabled care (OECD, 2026[59]).

Actions to develop economic and financial models

 

ACTION 9: Develop model for economics returns from AI in health (Economics)
Rationale: An empirical grounding for investments in AI in health is necessary to ground decision making in economics rather than hype. Such a model would need to consider socio‑economic factors for availability and adoption of AI tools in addition to modelling for the timeframes under which benefit is expected to be realised. Modelling could be integrated with other economic considerations for changing population demographics.
Possible outcome: An economic model to estimate the potential return on investment for AI in health, with a focus on prioritised use cases (e.g. administration, prevention, innovation) and consideration of time horizon for returns. For example, considering short-term gains in provider well-being, patient experience, and reduced human effort (documentation time), medium-term gains from better direct health outcomes, and long-term gains from improvements in population health. This modelling should be complemented by real-world evaluations of costs and benefits to assess whether projected returns are actually realised once AI is deployed in practice, including by leveraging Health Technology Assessment approaches to assess clinical effectiveness, cost-effectiveness, workflow impact, and broader system value. Models should differentiate between parties bearing costs and capturing benefits, as these frequently differ. Further, models should consider total cost of ownership, operational sustainability, maintenance, monitoring, cybersecurity, training, and shared investment mechanisms.
Sample existing initiative(s): ‒ Rather than evaluating AI investments only at the level of individual tools, health systems need approaches that estimate the broader returns generated by common enabling assets (e.g. interoperable health data infrastructure, secure data-sharing environments and evaluation platforms). These models can help justify collective investment and inform value‑sharing mechanisms across providers, researchers, innovators and public authorities (OECD, 2024[100]). ‒ RIHN 2.0 reimbursement in France ties conditional public funding for AI-enabled diagnostics to a mandatory, HAS-approved protocol demonstrating clinical utility, workflow impact, and economic value, before allowing progression to permanent reimbursement (ODELLE Technology, 2026[101]; Ministere de la santé et d’accès aux soins, 2024[102]).
ACTION 10: Develop considerations for financial management related to AI in health (Financing)
Rationale: The bottom-up approach to AI innovation in health has led to fragmented pilots with limited ability to scale and integrate into routine care. As health systems move from experimentation to sustained use of AI, financing models and incentives need to evolve to support the effective integration of AI-enabled solutions into healthcare delivery. This includes incentives for health providers and facilities to adopt and appropriately use validated AI applications, mechanisms that encourage patients and users to engage with AI-enabled services, and incentives for innovators to develop solutions aligned with health system needs. These changes form part of a broader transformation in how future health systems finance, reward and sustain new models of care.
Possible outcome: Documents leading practices for financing AI-enabled health systems. Considers (a) shifts in remuneration to incentivise the appropriate use of authorised AI applications, (b) incentives for the co‑operation across health facilities, researchers, and innovators to use and re‑use common foundational health assets, (c) expenses to sustain ongoing safety of AI in health solutions, and (d) the role of finance in risk- and value‑sharing.
Sample existing initiative(s): ‒ As part of the revision of the national health insurance reimbursement system, Japan introduced financial incentives for hospitals to adopt and use generative AI for clinical documentation by incorporating its use into staffing calculations linked to reimbursement (Ministry of Health, Labour and Welfare Japan, 2026[103]). ‒ Korea has established National Health Insurance evaluation and reimbursement principles that recognise the potential additional value of AI-enabled medical technologies. HIRA determines whether these technologies are covered under existing benefits or require New Health Technology Assessment, while practical experience with additional reimbursement for AI-enabled medical technologies is still emerging (HIRA, 2026[92]).

EL-MAP support through measurement and co‑ordination

 

ACTION: Measuring and co‑ordinating EL-MAP across experts and organisations
Rationale: Leading practices for measuring and developing key indicators around health AI have yet to be established. Developing them would support quantification of health AI’s usage and benefits. In turn, this would support evidence needed for effective scale‑up, as well as enable routine monitoring and evaluation of AI in health implementation.
Possible outcome: Supports EL-MAP through effective measurement, evaluation and monitoring of AI use cases, assessing their readiness, usage, and impact at individual and population levels. This could support evaluation of human and economic benefits across sectors and countries. Measurements could include indicators for preparedness, safety, experience, health outcomes, well-being, economic impact, and sustainability.
Existing initiative(s): ‒ OECD (forthcoming[6]).

Moving toward the future

Participants at the Madrid Conference were mutually committed to advancing the effective scale of AI in health. Their contributions before, during, and after the Conference have led to the Expert-Led Madrid Action Plan (EL-MAP).

Over the coming months, experts and international organisations will collaborate on how and when to proceed with the action plan. This will identify regional networks or organisations that are already providing leadership against specific actions as well as where formal mandates need to be agreed and funded. EL-MAP could represent a change in how disparate organisations and experts collaborate respectfully and collectively as well as foster learning from each other – in particular where there are gaps in levels of maturity.

The OECD could support this initiative with its expertise in economic analysis, financial modelling, health data governance, and measurement and co‑ordination among other areas. Further, the OECD could contribute to measuring the usage of AI in health including across prioritised use cases and monitor their impact related to achieving human and economic outcomes across sectors and countries.

While AI and related technologies in health are evolving at an unprecedented pace, taking the time to establish this strong collective action will create the clear signal so that investments are less duplicative and resources can be better used and re‑used to deliver future health needs. The risk of continuing the current path is that health systems will develop local solutions serving local populations but are unable to leverage foreign expertise to support broader expansion and benefit wider communities.

Experts at the Conference left with a clear sense of urgency to act together to realise the promise of AI for the benefit of all. The Expert-Led Madrid Action Plan is an important first step to realising this vision.