A new report from the OECD paints a picture that anyone working in health technology should study carefully. While 100% of OECD member countries now use AI in healthcare administration, only 10% have scaled AI to national level for clinical applications like medical imaging. The gap between experimentation and deployment is enormous — and the reasons behind it are not primarily technical.
They are structural. Fragmented data foundations, regulatory uncertainty, workforce capacity gaps, and governance blind spots are holding back what could be the most consequential shift in healthcare delivery since the adoption of electronic health records. The OECD’s response — a policy checklist organised into four pillars covering enablers, guardrails, engagement, and trustworthy deployment — is the first serious attempt at creating a cross-border framework for responsible AI scaling in health.
For healthcare technology companies, medical device manufacturers, and digital health startups, this report is more than a policy document. It is a roadmap for where procurement budgets, regulatory requirements, and institutional priorities are heading over the next three to five years.
The Scale Problem: Why 100% Adoption Doesn’t Mean 100% Impact
The OECD finding that every member country uses AI in administration but almost none have scaled it clinically reveals the real bottleneck. Administrative AI — scheduling, billing, claims processing, resource allocation — operates on structured data within well-defined workflows. Clinical AI — diagnostic imaging, pathology analysis, treatment recommendation, drug interaction prediction — operates on messy, unstructured, highly sensitive patient data within workflows where errors have direct human consequences.
The infrastructure requirements are fundamentally different. Administrative AI can run on standard cloud compute. Clinical AI at national scale requires GPU-accelerated inference, real-time processing of imaging data, federated learning architectures that keep patient data within institutional boundaries, and validation pipelines that meet regulatory standards in every jurisdiction where the system operates. The compute costs alone are significant — and growing.
The Data Foundation Challenge
The OECD checklist puts data foundations first for a reason. AI models are only as good as the data they’re trained on, and healthcare data is notoriously fragmented. Patient records sit in different formats across different systems in different institutions, often within the same city. Imaging data from one hospital may use different DICOM standards than the hospital across the street. Lab results, clinical notes, genomic data, and wearable sensor data all live in separate silos with different access controls and different levels of quality.
The emerging solution — country-led health data authorities that ensure data is findable, accessible, interoperable, and reusable (the FAIR principles) — is the right idea but will take years to implement. In the meantime, companies building AI for healthcare have to work with what exists: incomplete datasets, inconsistent formatting, and access processes that vary by institution, region, and country.
The Compute Economics of Healthcare AI
One aspect the OECD report touches on but doesn’t fully explore is the compute infrastructure required to scale AI in health. Training a diagnostic imaging model on millions of scans requires substantial GPU resources. Running that model in production across a national health system — processing thousands of scans per day with sub-second latency — requires even more. And when you add the emerging category of large language model applications in healthcare — clinical note summarisation, patient communication, literature analysis, clinical trial matching — the API costs start to compound rapidly.
Health systems and digital health companies are increasingly using AI APIs from providers like Anthropic, OpenAI, and the cloud-hosted AI services from Azure and Google Cloud. Many entered these platforms through startup grants, research partnerships, or promotional credit programmes — and now find themselves with credits allocated to one provider while their actual usage has shifted to another. If your organisation has unused Anthropic capacity from a research grant or pilot programme that didn’t scale, you can sell Anthropic credits through brokers who match sellers with buyers looking for discounted AI API access. It’s a practical way to recover value from credits that would otherwise expire — capital that can be redirected toward the AI workloads you’re actually running.
The Workforce Question
The OECD checklist identifies workforce capacity as a critical enabler — and the data supports it. Only 29% of member countries have established a national approach to improving AI use in the health workforce. The gap isn’t just about training clinicians to use AI tools. It’s about creating entirely new roles: clinical AI validators who can assess model outputs against medical evidence, health data engineers who can build compliant data pipelines, and AI ethics officers embedded within health institutions rather than technology companies.
The countries that move fastest on workforce development will have a structural advantage in AI adoption. Korea is already mandating AI education within health professional curricula. The UK’s Digital and Data Professional Capability Framework is proactively mapping the skills needed across both clinical and technical roles. These aren’t future plans — they’re being implemented now.
What This Means for Health Tech Companies
The OECD checklist is a signal of where health system procurement is heading. Companies building AI for healthcare should align their product development, compliance documentation, and commercial strategy to the four pillars: data foundations, scalability assurance, workforce integration, and trustworthy deployment. The companies that can demonstrate alignment with these priorities will have a significant advantage in public health system procurement over the next three to five years.
Frequently Asked Questions: AI in Healthcare
What percentage of OECD countries use AI in healthcare?
100% of OECD member countries now use AI in healthcare administration. However, only 10% have scaled AI to national level for clinical applications such as medical imaging — highlighting a massive gap between administrative adoption and clinical deployment.
Why hasn’t clinical AI scaled in healthcare?
The primary barriers are structural rather than technical: fragmented data foundations, regulatory uncertainty across jurisdictions, governance gaps, and insufficient workforce capacity. Clinical AI also operates on sensitive, unstructured patient data where errors have direct human consequences — requiring a higher bar for validation than administrative applications.
What is the OECD AI in Health Policy Checklist?
It is a framework developed by the OECD in partnership with the Global Digital Health Partnership and Coalition for Health AI. It is organised into four pillars — enablers, guardrails, engagement, and trustworthy deployment — with nine policy categories and 43 questions designed to help decision-makers identify priorities and blind spots in their AI health strategy.
What are the FAIR data principles?
FAIR stands for Findable, Accessible, Interoperable, and Reusable. In healthcare AI, FAIR principles ensure that patient data can be discovered, accessed through standardised protocols, used across different systems, and reused for both primary care and analytical purposes — all while maintaining compliance with data protection regulations.
How many countries have a national AI health strategy?
According to the OECD report, only 18% of member countries have established a strategy or action plan specifically at the intersection of AI and health. Several more are currently developing such strategies, but the majority of countries lack a dedicated national framework.
What is a regulatory sandbox for AI in health?
A regulatory sandbox is a controlled environment where AI developers can test their solutions under relaxed regulatory requirements while being monitored by authorities. Only 18% of OECD countries have established a national approach to regulatory sandboxes with a focus on AI in health — a mechanism that can significantly accelerate innovation while managing risk.
What is a model card in healthcare AI?
A model card is a standardised document that accompanies an AI model and certifies its compliance, transparency, and accountability. Developed by organisations like the Coalition for Health AI, model cards describe what a model was trained on, how it performs across different populations, its known limitations, and its intended use cases — enabling implementers to assess fitness for their specific clinical context.
How is AI currently used in healthcare administration?
Common administrative AI applications include automated scheduling and appointment management, claims processing and billing optimisation, resource allocation, supply chain management, patient flow prediction, and clinical documentation assistance. These applications operate on structured data and well-defined workflows, making them easier to deploy than clinical AI.
What compute infrastructure does healthcare AI require?
Clinical AI at scale requires GPU-accelerated inference for imaging and diagnostics, real-time processing capability, federated learning architectures that keep patient data within institutional boundaries, and validation pipelines that meet regulatory standards. Language model applications add API inference costs on top of existing infrastructure. The compute requirements are substantial and growing.
What is federated learning in healthcare?
Federated learning is a machine learning approach where models are trained across multiple institutions without moving patient data outside each institution’s boundaries. Instead, the model travels to the data — each institution trains on its local data and only shares model updates (not patient records) with a central server. This preserves patient privacy while enabling AI models to learn from diverse, multi-institutional datasets.
Which countries are leading in AI health workforce development?
Korea has mandated AI education within health professional curricula. The United Kingdom has developed the Digital and Data Professional Capability Framework, which proactively maps skills needed across clinical and technical roles. Overall, 29% of OECD countries have established a national approach to improving AI use in the health workforce.
What are the main risks of AI in healthcare?
The OECD identifies several key risks: skewed or biased training data that produces inequitable outputs, privacy and security risks from handling sensitive patient information, insufficient transparency in how AI models reach their conclusions, potential job displacement among healthcare workers, and de-personalisation of the patient-provider relationship.
How are health systems paying for AI infrastructure?
Through a mix of direct cloud provider contracts, startup and research grants, enterprise agreements, and promotional credit programmes from AI providers. Many health systems and digital health companies received credits during pilot phases that are now expiring as priorities shift — creating a growing pool of unused AI and cloud capacity.
Can health organisations sell unused AI credits?
Yes. Organisations with unused cloud or AI API credits — whether from startup grants, research partnerships, or pilot programmes that didn’t scale — can recover value by selling them through brokers who match sellers with buyers. This is particularly relevant as health AI strategies evolve and organisations shift between providers.
What should health tech companies focus on to align with the OECD framework?
Align product development with the four pillars: demonstrate robust data foundations and FAIR compliance, build scalability evidence through model cards and real-world validation, invest in workforce integration through training and usability, and embed trustworthy AI principles including transparency, bias monitoring, and human oversight. Companies that can document alignment with these priorities will have a structural advantage in public health procurement.