Artificial intelligence should transform health systems, not automate their failures
Dr Ricardo Baptista-Leite
Chief Executive Officer, HealthAI – The Global Agency for Responsible AI in Health, Switzerland
Dr Ricardo Baptista Leite, HealthAI – The Global Agency for Responsible AI in Health, Switzerland — Health systems must be redesigned around prevention before advances in medical AI can reduce disease burden and narrow inequities in care.
Published on September 1st, 2026
The promise of AI in health
Artificial intelligence (AI) is already reshaping medicine. The Frontiers in Science lead article by Guni et al. describes the extraordinary expansion in its capabilities, ranging from multimodal diagnosis and molecular discovery to prediction, navigation, and increasingly autonomous clinical tasks (1). Yet the central policy question is no longer simply what AI can do. It is what health systems should ask AI to do.
Most health systems remain organized around a disease-reactive model: resources are mobilized after symptoms appear, services are paid for activity rather than health gained, and data are fragmented around encounters rather than longitudinal lives (2, 3, 4). If AI is layered onto this architecture without redesigning it, we may exacerbate inequities, biases, and inefficiencies within the system. Put simply, AI retrofitted into legacy workflows may ultimately make health systems very efficient at being inefficient. Faster documentation, image interpretation, and triage can create value, but optimization of a structurally misaligned system will not, by itself, reduce disease burden or extend healthy life.
The policy objective must therefore be broader than the adoption of safe AI products. It should be the responsible development, adoption, and diffusion of AI to create measurable improvements in health, equity, workforce experience, and system sustainability. This requires four shifts in governance.
From product performance to health-system value
Regulatory authorization appropriately asks whether a technology is sufficiently safe and performs as intended. It cannot answer the separate question of whether deploying that technology represents good health policy and effectively reduces the burden of disease. Health technology assessment, procurement, and reimbursement must therefore examine system impact: Does an AI application move care upstream toward prevention and earlier intervention? Does it improve outcomes across population groups rather than only average accuracy? Does it release clinical time for human care, reduce avoidable demand, or merely generate additional activity?
HealthAI’s analysis of implementation of the European Union’s Artificial Intelligence Act illustrates the problem. It identifies market access—not market authorization—as a binding constraint: regulatory approval does not lead to clinical adoption when reimbursement pathways, budgets, institutional responsibilities, and implementation support are absent (5). Europe’s regulatory ambition will produce limited health benefit unless payment and procurement reward validated outcomes and service redesign, not simply the acquisition of compliant technology. Regulators should remain focused on safety and performance; health systems, through their health technology assessment bodies, must also apply a formal “health-transformation test” when deciding what to fund and scale.
The same discipline should extend to public-sector portfolios. Governments should stop funding disconnected pilots that cannot be integrated, evaluated, or maintained. Funding should privilege solutions with an implementation pathway, interoperable data architecture, workforce plan, post-market monitoring, and a plausible route to equitable scale. A technically impressive model without these conditions is not an innovation strategy; it is an experiment without a health-system destination.
From one-off approval to continuous assurance
The prevailing regulatory model was built for comparatively stable products. AI models may drift, be updated, interact with changing populations and workflows, or depend on foundation models that evolve outside the direct control of the clinical deployer. As systems become more agentic, the fiction that safety can be established at a single pre-market moment becomes increasingly untenable.
HealthAI’s Global Landscape report finds persistent gaps in adaptive AI oversight, interagency coordination, and post-market surveillance across diverse jurisdictions (6). Governance must become lifecycle based. Health institutions need local performance monitoring, version traceability, clear escalation protocols, and the capacity to detect differential performance across populations. National authorities need interoperable incident-reporting systems, and regulators need mechanisms to exchange safety signals internationally. A global early-warning architecture for AI in health, led by national regulators, should perform a function analogous to pharmacovigilance: identifying rare, delayed, or context-specific harms that no single institution or country can detect alone.
Accountability must also follow this lifecycle. Clinicians should not become the default liability sink for failures originating in model design, inadequate validation, opaque updates, deficient procurement, or poor organizational implementation. Responsibility should be distributed explicitly among developers, AI providers, deployers, health institutions, and regulators according to their control over each stage.
From regulatory fragmentation to cooperative capacity
AI systems and their supply chains cross borders; regulatory capacity and health-system readiness remain profoundly uneven. The Global Landscape report shows convergence around shared principles and medical device regulatory frameworks, but also overlapping mandates, incompatible approaches, and major differences in enforcement capacity (6). The answer is neither a single global regulator nor identical national laws. It is regulatory interoperability: common minimum expectations, shared terminology, reliance on trusted assessments, coordinated post-market surveillance, and structured peer learning.
This is also an equity imperative. Countries with the greatest shortages of health workers and services may have the most to gain from AI, yet the implementation of appropriate digital health technologies may be difficult to accomplish, especially in low- and middle-income countries (7, 8). Governance capacity should therefore be financed as a global public good—through a hybrid model, including not only international development funding and financial institutions, but also philanthropic, private sector, and domestic funds—alongside digital infrastructure and workforce development. Otherwise, high-income markets will define the evidence, standards, and products, while lower-resource settings inherit technologies that were neither designed nor adequately assessed for their populations.
From automating care to redesigning health
The most important promise described by Guni et al. is the possibility of moving towards prediction, pre-disease identification, and prevention (1). But this transition will not be delivered by algorithms alone. It requires longitudinal and representative data, interoperable infrastructure, prevention-oriented payment models, redesigned clinical pathways, public participation, and a workforce able to act on earlier signals. Analysis by the Organisation for Economic Co-operation and Development similarly concludes that AI is not a panacea; policies, governance, and collective choices determine whether it produces equitable and sustainable impact (9).
Human agency must remain central. Meaningful oversight is not achieved by placing a nominal clinician at the end of an automated chain. Patients and professionals should help determine where and how AI is useful, when human judgment is indispensable, and which outcomes matter. Co-design, co-implementation, and co-adoption are critical to ensure that these technologies meet citizens’ health needs equitably. The aim should be to use machines for what they do well while returning time, attention, and empathy to human relationships and ensuring clear human accountability.
The gap between technical capability and clinical benefit is therefore a governance and leadership gap. Governance should not be treated as friction that slows innovation, but as the infrastructure that enables trustworthy innovation to scale—recognizing that innovation will move at the speed of trust. The measure of success will not be how many AI tools enter hospitals, but whether they help societies prevent illness, reduce inequity, improve care, and create healthier lives.
Copyright statement
Copyright: © 2026 [Baptista-Leite]. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in Frontiers Policy Labs is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Generative AI statement
During the preparation of this article, the author used ChatGPT (OpenAI) to assist with the synthesis of the author’s ideas, and refinement of language and presentation. The substantive arguments, interpretations, policy recommendations, and conclusions are those of the author. The author critically reviewed and edited all AI-assisted content and assumes full responsibility for the accuracy, integrity, and final content of the manuscript.
References
Guni A, Hu W, Morley J, Cohen IG, Barnaghi P, Ashrafian H. Large language medicine: defining a new paradigm in human health. Front Sci (2026) 4:1810095 doi: 10.3389/fsci.2026.1810095
World Economic Forum, Partnership for Health System Sustainability and Resilience. Acting early on non-communicable diseases: a framework for health system transformation. WEF (2026). Available at: https://www.weforum.org/publications/acting-early-on-non-communicable-diseases-a-framework-for-health-system-transformation/
Fleetcroft R, Steel N, Cookson R, Walker S,Hove A. Incentive payments are not related to expected health gain in the pay for performance scheme for UK primary care: cross-sectional analysis.BMC Health Serv Res (2012) 12:94 doi: 10.1186/1472-6963-12-94
Eisinger-Mathason TSK, Leshin J, Lahoti V, Fridsma DB, Kno AN. Data linkage multiplies research insights across diverse healthcare sectors. Commun Med (2025) 5:58 doi: 10.1038/s43856-025-00769-y
HealthAI – The Global Agency for Responsible AI in Health. Harnessing AI for health and economic competitiveness: translating the EU AI Act into action. HealthAI (2026). Available at: https://healthai.agency/knowledge-hub/knowledge-resources-on-ai-governance-in-health/#close
HealthAI – The Global Agency for Responsible AI in Health. AI governance in health: global landscape 2025 report (version 1).HealthAI (2025). Available at: https://healthai.agency/knowledge-hub/knowledge-resources-on-ai-governance-in-health/#close
Ciecierski-Holmes T, Singh R, Axt M, Brenner S, Barteit S. Artificial intelligence for strengthening healthcare systems in low- and middle-income countries: a systematic scoping review. npj Digit Med (2022) 5:162 doi: 10.1038/s41746-022-00700-y
World Health Organization. Global strategy on digital health 2020-2025. WHO (2021). Available at: https://www.who.int/publications/i/item/9789240020924
Organization for Economic Co-operation and Development. Scaling artificial intelligence in health. OECD (2026). Available at: https://www.oecd.org/en/publications/scaling-artificial-intelligence-in-health_a436e12d-en/full-report.html
