Landmark OECD report supports need for AI infrastructure and oversight
- John Kalafut
- Apr 15
- 7 min read
Updated: 2 days ago
NEWSLETTER - APRIL 2026
AI GOVERNANCE IN HEALTHCARE | HEALTH AI STRATEGY & OBSERVABILITY | 5 MIN READ
John Kalafut, PhD | Laure Tessier-Delivuk PMP|Charlotte Kalafut |Aaron Hillman CDH-E
A landmark OECD report[1] confirms what clinicians and health technologists already suspect. We have the tools, but not the infrastructure, oversight, or the will to deploy them responsibly at the system level. Here is what that means for your organization.
100% OECD nations use AI in health admin | 10% decision making AI used at national scale | 18% have national AI oversight for AI in health | 29% have workforce readiness policies |
The pilot paradox: AI is everywhere in health, but rarely at scale
In 2026, the OECD published its most comprehensive assessment yet of artificial intelligence in healthcare. The headline is both reassuring and quietly alarming. Every single member country uses AI in health administration. Yet when it comes to patient decision making AI, one of the most mature and clinically validated domain of health AI, only 10% of applications are being used on a national scale (OECD, 2026).
The Real Blockers to AI Adoption in Healthcare
Much of the discourse around health AI focuses on model quality: whether an algorithm can match or outperform a radiologist, a pathologist, or a triage nurse. The evidence on that front is increasingly positive though the OECD report situates the challenge elsewhere entirely.
72% of OECD countries have enacted policies supporting better use of health data. That sounds solid but look beneath that headline and you find fragmentation: data that is not findable, not interoperable across systems, and not representative of the full patient population. FAIR principles (Findable, Accessible, Interoperable, Reusable) remain an important policy direction, but the OECD makes clear that data foundations, interoperability, representativeness, and access governance remain uneven in practice (OECD, 2026, p.16).
More troubling still is what happens after an AI solution is approved and deployed. Only 18% of OECD member states report established national oversight for the use of AI in health. That does not mean individual organizations have no governance, but it does mean healthcare organizations function in variable and often unclear regulatory and compliance environments, leading to lost opportunity because of a need to ‘figure it out’ on their own. (OECD, 2026, p.49).
Why medical imaging is a canary in the coal mine
Medical imaging AI has a decade-long head start over other clinical domains. Deep learning models for radiology were among the first to achieve regulatory clearance. If any domain should have cracked the deployment challenge, it is imaging.
The OECD notes that only a small percentage of member countries report national‑scale deployment of medical imaging AI. For a U.S. audience, “national scale” is not directly comparable, given the absence of a centralized health system. Still, the underlying lesson holds. The main constraint is not regulatory approval, which largely exists, but downstream execution: integration with PACS and RIS infrastructure, clinician adoption within real workflows, continuous post‑deployment monitoring, and clarity around accountability when tools underperform (OECD, 2026, p. 8).
From the OECD Report (2026) Emerging leading practices include the development of model cards, developed in collaboration with bodies such as the Coalition for Health AI, which certifies compliance, transparency, and accountability of AI solutions when applied in real-world settings. These represent a practical bridge between technical specifications, transparency expectation and operational governance (OECD, 2026, p.7). |
The absence of oversight is compounded by a wider gap in the enabling conditions for AI adoption. The report finds that only 20% of OECD members have active policy actions enabling the use of AI in health systems. As the report notes, without plans for adoption and scale, the value of AI will be limited (OECD, 2026, p.58).
The Four pillars of AI infrastructure and the one most organizations still overlook
The OECD’s AI in Health Policy checklist is organized around four pillars, with capacity and capability called out as a critical area of action: (1) establishing enablers, (2) implementing guardrails, (3) engaging meaningfully, and (4) deploying trustworthy AI. Understanding where your organization sits against each is a practical starting point for any AI governance review (OECD, 2026, p.15).
Establishing Enablers Data foundations, FAIR compliance, workforce capacity, and assurance frameworks. The prerequisites for responsible deployment. | Implementing Guardrails Oversight mechanisms, performance monitoring, incident reporting, and policy frameworks to govern AI throughout its operational lifecycle. | Engaging Meaningfully Meaningful involvement of clinicians, patients, and communities as co-designers of trustworthy systems. | Deploying Trustworthy AI Operationalizing transparency, explainability, safety, and accountability, aligned with the OECD AI Principles and emerging regulatory frameworks. |
Most organizations have made at least partial progress on the first pillar. Data strategies are in motion. Models have been procured or built, and in many cases validated in controlled settings. The gap emerges after that point. Achieving the remaining pillars requires mechanisms to observe models once they are live, to understand how they are performing in day‑to‑day clinical use, and to intervene when performance drifts. But performance monitoring cannot stop at drift detection. Organizations also need to answer a harder question: “is this AI system actually delivering the value we expected?” That means tracking outcomes across operational, clinical, and financial dimensions, quantified where possible and assessed qualitatively where not. In practice, this means continuous monitoring, clinician‑visible performance dashboards, auditability that holds up to regulatory scrutiny, and defined escalation paths when outputs fall outside expected norms. Without these capabilities, AI stays a pilot. Without effective oversight, it is also likely that AI in health solutions will continue to be developed in silos, be challenged to scale, and fail to achieve their potential (OECD, 2026, p.58).
The workforce problem is a hidden risk
There is an interesting, secondary finding in the OECD report regarding workforce development and readiness. Only 29% of member countries have enacted meaningful workforce readiness policies, covering training, curriculum integration, and role adaptation, to support AI deployment in clinical settings. This creates a particular kind of fragility (OECD, 2026, p.49) because you can deploy a technically excellent AI solution, with all the appropriate regulatory approvals, into a clinical environment where radiologists, radiographers, and referring clinicians have received little structured guidance on how to interpret, interrogate, or appropriately override its outputs. The report identifies capacity and capability development among the health workforce as a priority area where only 29% of countries show active, coordinated progress, with 47% only partly engaged (OECD, 2026, Table 4.1).
The OECD emphasizes that a skilled and knowledgeable health workforce is essential for uptake and sustained use of AI, and it points to proactive planning and targeted upskilling as emerging leading practices, that remain the exception rather than the norm (OECD, 2026, p.21).
Even where awareness is growing, another barrier remains underappreciated: many health systems still lack mature pathways to evaluate, procure, and scale AI consistently.
Procurement and assessment still lag
The OECD analysis also highlights the impact immature procurement processes have on the broader adoption of AI into healthcare. Only 11% of OECD countries report having procurement guidance adapted to AI in health, 24% report health technology assessment approaches that include AI, and 18% report regulatory sandboxes specific to AI in health.
What this means in practice for AI in medical imaging
For organizations that have deployed or are evaluating the use of AI technologies into their operations the OECD findings point to a set of operational imperatives that administrators should be addressing:
1) Post-deployment monitoring - Is there a mechanism, technical and organizational, to continuously observe model performance in your specific population and clinical context?
Does that monitoring go beyond the technical dimensions of the AI application or system, are you able to answer “how is this AI tool delivering value?” The report points to oversight, measurement, and monitoring as essential and highlights emerging leading practices such as indicator for clinical effectiveness, economic impact, and post-market surveillance (OECD, 2026, p.22)
2) Model cards and transparency artifacts are moving from best practice to expected standard. The Coalition for Health AI framework, referenced positively in the report, provides a practical template. Having model documentation that follows your AI solution from approval through deployment and ongoing use is increasingly a condition of responsible operation (OECD, 2026, p.19).
3) Data quality is not a one-time concern. The FAIR principles apply to operational data, not just training data because the consistency and composition of the data given to any AI/ML system will affect its reliability. The report notes that if data are of poor quality, then the resulting AI solutions will be of poor quality, and measuring and reporting data quality will help collaborators work together with confidence while building trust with stakeholders (OECD, 2026, p.17).
Key Takeaway Inaction carries its own risk. The OECD suggests that without responsible scale, health systems may miss meaningful opportunities to improve outcomes, efficiency, and access. |
The Asher Informatics perspective
At Asher Informatics, we work at the intersection of AI governance and observability in healthcare environments. The OECD's findings resonate with what we see in practice: organizations that have successfully navigated a few pilots successfully often find themselves without the operational infrastructure to know, with confidence, whether their deployed AI is performing as intended across their full patient population.
The most important shift in framing we can offer is this: AI observability and assessment is not a technical add-on. It is part of the clinical governance layer required to make AI deployment responsible rather than speculative. Continuous monitoring of model outputs, demographic performance analysis, clinician interaction patterns, and systematic audit trails are all needed to realize the potential of AI technologies in the practice of healthcare.
The OECD's four-pillar framework gives healthcare executives a structured vocabulary for a conversation that too often defaults to technical detail. Bring it into your next board-level AI discussion. Ask where your organization sits on each pillar. The answer will tell you more about your AI readiness than any vendor demo.
Key actions for healthcare executives
|
References
1OECD (2026). Scaling Artificial Intelligence in Health. OECD Publishing, Paris. https://doi.org/10.1787/a436e12d-en
About Asher Observer
Asher Observer is the intelligence publication of Asher Informatics, delivering analysis, opinion, and insight at the intersection of clinical AI assessment, health system governance, and observability.
Asher Informatics provides health AI strategy and observability solutions to
simplify Health AI Enterprise Management with Oversight solutions to
Responsibly Govern, Strategically Assess, and Independently Monitor.
[1] OECD (2026), Scaling Artificial Intelligence in Health, OECD Publishing, Paris, https://doi.org/10.1787/a436e12d-en
Comments