Scaling Artificial Intelligence in Health

1. Responsible scale of AI in health

Interest in artificial intelligence (AI),1 is exponentially growing in healthcare. While the main developments to date have been predominantly made in medical imaging and automating administrative tasks, advancements in other fields will likely follow and become an integral part of healthcare (OECD, forthcoming workforce publication, 2026), (Secinaro et al., 2021[1]). Today, AI is used in health systems across all OECD countries (100% of those interviewed. However national-level scale‑up remains limited; only 10% of medical imaging applications being used on a national scale (OECD, forthcoming workforce publication, 2026).

Beyond imaging, AI is enhancing analytical capabilities, diagnostic accuracy, patient monitoring, health system administration, drug discovery, and workflow efficiency. Other areas, such as predictive medicine, personalised medicine and health literacy can also meaningfully benefit from advances in AI, resulting in more human-centred care.

At the same time, the adoption of AI adds risks due to potential workforce displacement, concerns over data protection and security, the application of AI trained on skewed data, and unequal access to the benefits from AI solutions.

Many of these risks arise from the challenge of scalability of AI solutions across institutions, regions, countries and populations. Consider the time it took to reach 40% consumer adoption of previous technologies across sectors: 64 years for the telephone, 45 years for electricity, 23 years for computers, 16 years for mobile phones, and 13 years for the internet (DeGutsa, 2012[2]). The question remains, “How long will it take AI to scale, when considering the societal acceptance and ethical debate for AI in healthcare?”. In some cases, the barriers to AI adoption are justified, given the concerns about the current state of safety of these systems in the health sector. In other cases, barriers to scaling arise from the lack of robust policy; fragmented data and digital infrastructure; minimal co‑ordination across the AI in health ecosystem; governance frameworks that fail to incentivise safe and scalable solutions; and previous efforts that eroded trust among both patients and healthcare professionals.

The difficulty to scale AI solutions in health undermines its potential economic and health benefits. The inability to scale leads to duplicative investments and poor-quality solutions based on data that are not representative of target populations. Without action, the benefits of AI will not be distributed evenly, and the reach of innovation will be limited. This disparity is already evident, with wealthier institutions able to invest in the necessary technical infrastructure, financial resources and skilled personnel to implement and sustain AI solutions whereas less affluent institutions lack the people, data, or infrastructure to design, develop, implement, sustain, and evolve these AI solutions. A fragmented approach leads to AI solutions that are more expensive and generate sub-optimal results with limited reach.

A collective approach to AI in health – where governments partner with stakeholders to act on their role as stewards of the health system – is necessary to foster safety, privacy, and innovation. This approach can enable governments to proactively shape the integration of AI into health systems while ensuring alignment with public interest objectives.

To that end, in January 2024, OECD Health Ministers, in the Declaration on Building Better Policies for More Resilient Health Systems (https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0500) outlined several key objectives related to the use of AI in the health sector:

“assist in our efforts to harness the transformative potential of new technologies for health such as genomics and genetics, and artificial intelligence, while ensuring that their application delivers better health outcomes by tracking and assessing the health and health systems implications including on health spending of these developments”,

“assess and make policy recommendations on the implications of using artificial intelligence in health systems”,

“develop a sector-specific framework for AI in health, that is aligned with existing multi-sectorial frameworks, to encourage responsible use of artificial intelligence, exploiting cross-sectoral synergies to promote fairness, transparency and accountability, while ensuring consistency across policy domains”.

Based on the Declaration, the OECD established a time‑bound AI in Health Expert Group within the OECD.AI Network of Experts in May 2024 (OECD, 2024[3]). This Expert Group involves individuals from 27 countries, representatives from leading organisations (e.g. World Health Organization, Coalition for Health AI, HealthAI), and members of civil society. Aligned with the 2024 Health Ministerial Declaration, and building on the OECD Recommendation for Health Data Governance (OECD, 2016[4]) and OECD AI Principles (OECD, 2024[5]), the Expert Group was focussed on addressing the challenge of scalability for AI in Health. To that end, the experts focussed their efforts on developing an AI in Health Policy Checklist for the responsible scale of AI in health, considering both within jurisdictions and across borders. In total, the group has met six times to develop the checklist.

This report is structured around three core components: the rationale for a sector-specific approach; the “AI in Health Checklist”; and its application to assess readiness for the responsible scale of AI in health:

1. Rationale for the responsible scale of AI in health (Why): The case for leadership and policy clarity at the intersection of AI and health, outlining why a formalised approach is beneficial to guide safe and responsible scale (remainder of Chapter 1).

2. AI in Health Policy Checklist (What): A tailored AI in Health Policy Checklist, developed through expert consultations and review of existing evidence and structured around the pillars of Enablers, Guardrails, Engagement, and Trustworthiness. (Chapter 2).

3. Current state of implementation of AI in health across OECD Members (How): OECD countries are advancing the implementation of AI in health systems across legislation, governance, technical infrastructure and capacity. Leveraging the AI in Health Policy Checklist, an analysis of the current state of implementation is structured around core components of checklist (Chapter 3).

The report closes by summarising the report and identifying possible future actions for policy work at the intersection of health and AI (Chapter 4).

Rationale for formalising an approach at the intersection of AI and health

The transformative potential of AI in health is clear, but so are the risks. Without a coherent comprehensive AI in Health Policy Checklist, countries face fragmented adoption, uneven guardrails, and unclear return on investment. This may lead to challenges of trust or unsustainable return on investment. This is exacerbated in countries with a federated governance structure – such as Germany, Australia or Canada – where the accountability for health is at the state or provincial level can cause further fragmentation. A checklist could help regional health authorities align their activities for collective benefit while enabling local delivery of health services.

To further reinforce the case for an AI in Health Policy Checklist, key concerns related to privacy, safety, validation and value have been identified. These interconnected areas exhibit challenges that are significant considerations in the health sector due to concerns about the potential for human harm and the opportunity for human benefit. Without a focus on policies that address these concerns, there is risk that decisions are delayed or inconsistent. The OECD AI in Health Policy Checklist aims to help policymakers identify key challenge areas and provide insights into emerging practices for actionable steps. The following sections provide context on these concerns and the benefit of a collaborative approach.

Health data access, use, privacy, and security

The OECD Recommendation on Health Data Governance (OECD, 2016[4]) provides guidance in several areas related to health data access, use, privacy, and security. The Recommendation provides considerations for protecting privacy and security of individual health data, while also supporting the use of data for purposes of direct patient care, health system management, public health, and innovation (the latter three of which are considered examples of analytic data use).

This will be especially important for amplifying the benefits and addressing the risks of emerging technologies and capabilities such as quantum computing (OECD, 2025[6]) and genomics (OECD, 2021[7]).

Consistency in this practice fosters trust of the public and providers while enabling an efficient foundation on which AI solutions are built and being prepared to adapt to future opportunities and risks. A checklist will help avoid blind spots in establishing that consistency.

Health safety, quality and equity

If health technologies fail, they can have long-lasting and profound consequences on trust among health providers and the public. Likewise, failure to leverage the advanced capabilities of AI in the provision of health can also cause harms from inaction. As opportunities to improve healthcare quality emerge – due to the use of technology including AI – the standards of healthcare quality and safety will evolve.

If solutions are not designed to scale, reaching the quality-of-care standards will become more difficult for low-resource or hard-to-reach settings to achieve. This includes supporting health in remote and rural areas, supporting those with poor health and data literacy, and assuring indigenous data sovereignty (where applicable).

Enabling responsible scale of solutions will help the benefits of AI reach every person. A checklist will help identify blind spots for barriers to scale the benefits of AI while assuring safety.

Validating innovation for use in health

Emerging studies indicate that 75% of AI in health solutions evaluated through randomised control trials demonstrate a positive impact (Han et al., 2024[8]). For emerging innovators, particularly small and medium enterprises (SMEs), there is a risk that the fragmented patchwork of policies and related processes slow down innovation and competitiveness.

While variation in policies and processes is expected across jurisdictions and countries, unnecessary variation can add significant costs (e.g. financial, time) for researchers and innovators without commensurate value of increased protection for the jurisdiction or the public. The unintended cost of variation includes lack of clarity for how approvals will operate. For emerging capabilities, such as adaptive and agentic AI2 systems3, the health system has been slow to clarify their approach across medical device approvals and health technology assessments among other areas.

Enabling a consistent approach will support capacity and capability building toward a robust innovation ecosystem across delivery, approvals, and scale. A checklist will help to identify blind spots in establishing that ecosystem that accelerates innovation for all.

Value dynamics of healthcare

The evaluation of AI initiatives in health mirror those in the existing landscape of health system valuation, which considers a wide range of indicators – generally aligning with the quintuple aim for healthcare improvement (Nundy, Cooper and Mate, 2022[9]). Benefits from AI in health could translate to better health outcomes (e.g. using AI in medical imaging to improve accuracy diagnosis and treatment), healthcare experience (e.g. optimising and reducing waiting time), provider experience (e.g. reduced administration leading to more time to patient care), improved equitability (e.g. identifying populations that would benefit form a targeted health intervention), and reduced cost (e.g. reducing redundant testing, optimisation of sparse resources, improvements to productivity). For example, evidence from the American Medical Association (AMA) shows that Ambient Voice Technology reduced documentation time by 85.8%, saving an average of 5.27 minutes per case compared to manual typing (Karavassilis et al., 2025[10]).

Importantly, these benefits may be achieved over both a long-time horizon and broad population. Within the private sector and pharmaceutical research and development, AI also offers promise to improve the discovery, development and supply chain delivery of new therapies (Sanofi, 2023[11]).

The benefits of AI in health can be well understood, but analysis is needed to evaluate the social and economic impact of AI solutions. A checklist will help avoid blind spots that prevent understanding and evaluating those benefits across populations and geographies.

References

[2] DeGutsa, M. (2012), Are Smart Phones Spreading Faster than Any Technology in Human History? | MIT Technology Review, https://www.technologyreview.com/2012/05/09/186160/are-smart-phones-spreading-faster-than-any-technology-in-human-history/, https://www.technologyreview.com/2012/05/09/186160/are-smart-phones-spreading-faster-than-any-technology-in-human-history/ (accessed on 15 October 2025).

[8] Han, R. et al. (2024), “Randomised controlled trials evaluating artificial intelligence in clinical practice: a scoping review”, The Lancet Digital Health, Vol. 6/5, pp. e367-e373, https://doi.org/10.1016/S2589-7500(24)00047-5.

[10] Karavassilis, M. et al. (2025), “Ambient Voice Technology in Same Day Emergency Care: Enhancing Documentation Efficiency and Patient Flow”, Acute Medicine, Vol. 24/1, pp. 49-57, https://doi.org/10.52964/AMJA.1006.

[9] Nundy, S., L. Cooper and K. Mate (2022), “The Quintuple Aim for Health Care Improvement”, JAMA, Vol. 327/6, p. 521, https://doi.org/10.1001/jama.2021.25181.

[13] OECD (2026), “The agentic AI landscape and its conceptual foundations”, OECD Artificial Intelligence Papers, No. 56, OECD Publishing, Paris, https://doi.org/10.1787/396cf758-en.

[6] OECD (2025), “A quantum technologies policy primer”, OECD Digital Economy Papers, No. 371, OECD Publishing, Paris, https://doi.org/10.1787/fd1153c3-en.

[12] OECD (2025), Recommendation of the Council on Artificial Intelligence, https://legalinstruments.oecd.org/en/instruments/oecd-legal-0449.

[5] OECD (2024), AI principles, https://www.oecd.org/en/topics/ai-principles.html.

[3] OECD (2024), OECD AI, https://oecd.ai/en/.

[7] OECD (2021), “Building and sustaining collaborative platforms in genomics and biobanks for health innovation”, OECD Science, Technology and Industry Policy Papers, No. 102, OECD Publishing, Paris, https://doi.org/10.1787/11d960b7-en.

[4] OECD (2016), OECD Recommendation on Health Data Governance, https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0433.

[11] Sanofi (2023), Artificial Intelligence in R&D, https://www.sanofi.com/en/our-science/digital-artificial-intelligence/artificial-intelligence-rd (accessed on 14 October 2025).

[1] Secinaro, S. et al. (2021), “The role of artificial intelligence in healthcare: a structured literature review.”, BMC medical informatics and decision making, Vol. 21/1, p. 125, https://doi.org/10.1186/s12911-021-01488-9.

Notes

← 1. Defined as “a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment” (OECD, 2025[12]).

← 2. Agentic AI are systems of co‑ordinated agents that can break down tasks, work together over extended periods, and use external tools to achieve complex goals. These systems are designed to operate in more open-ended, less predictable environments and often require less human supervision than individual AI agents (OECD, 2026[13]).

← 3. Artificial intelligence system is a machine‑based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment (OECD, 2025[12]).