Escaping Pilot Purgatory: The OECD Blueprint for Scaling AI in Health

Maya: Did you know that across all OECD member countries, 100% of them use artificial intelligence in administration, but when we look at national-level scale-up for medical imaging applications, that number drops to just 10%?

Sam: That massive gap between local pilots and enterprise scaling is exactly why we are unpacking a document published by the Organisation for Economic Co-operation and Development, titled Scaling Artificial Intelligence in Health.

Maya: And right off the bat, we need to let you know that our voices are AI-generated, and this is a summary of a public document. Our goal is to figure out how health systems can move these incredible tools from isolated experiments into everyday care.

Sam: I love that framing. We hear so much about the promise of these algorithms, but health system executives and medtech leaders are currently stuck trying to navigate massive infrastructure and reimbursement bottlenecks. It is often referred to as pilot purgatory.

Maya: To put this scaling challenge into perspective, the document provides some fascinating historical context on how long it takes for a technology to reach 40% consumer adoption.

Sam: Yes, it is a great baseline. It took 64 years for the telephone to hit that mark. Then, it took 45 years for electricity, 23 years for computers, 16 years for mobile phones, and 13 years for the internet.

Maya: Which immediately raises the question: how long will it take for algorithms to scale in healthcare? It is fundamentally different from a mobile phone. You are dealing with highly sensitive data, patient safety, and a culture centered on doing no harm.

Sam: But the OECD makes a really compelling counterpoint. While caution is absolutely necessary, there is also a significant risk in inaction. If we do not scale these tools, we are actively denying patients better care.

Maya: Exactly. The report notes that 75% of artificial intelligence in health solutions evaluated through randomised control trials actually demonstrate a positive impact. We know they work. We just need to figure out how to deploy them safely across borders and hospital networks.

Sam: To do that, the OECD proposes a policy checklist built around 4 thematic pillars. Enablers, guardrails, engagement, and trustworthiness. Let us start with the enablers pillar, because the foundational element of any algorithm is data.

Maya: And the healthcare sector has a massive data problem. According to the document, health data represents the largest segment of data from any industry. It is estimated at 30% of all data generated.

Sam: But despite creating 30% of the world's data, the health sector is using less than 5% of it for decision making purposes. The rest is just sitting in fragmented silos.

Maya: That fragmentation is the enemy of artificial intelligence. If data is locked away in disconnected, proprietary systems, developers are forced to duplicate their financial investments and computational efforts, and the resulting models end up trained on highly skewed or unrepresentative populations. The OECD says we need to fundamentally restructure our approach by ensuring all health data aligns with the FAIR principles.

Sam: Findable, accessible, interoperable, and reusable. The interoperability part seems to be gaining some real traction. The report points out that 27 countries have now adopted a national health interoperability strategy.

Maya: Interoperability does not mean everyone uses the exact same software vendor or database architecture. It means the technical and semantic standards are compatible across the entire health ecosystem. For example, the report highlights Finland, where healthcare providers are mandated by the Act on Client Data, specifically law 1050/2021, to use Kanta Services for standardised electronic patient records. This ensures that critical health and social care data flows seamlessly across both public and private entities.

Sam: Denmark is doing something similar with Sundhed.dk and the Danish Health Data Authority, upholding standardised reference architectures. And looking at the bigger picture, the European Health Data Space is setting standards for cross-border data exchange across member states.

Maya: But even if the data is interoperable, you still need a way to legally and securely access it for research and innovation. This is where Health Data Access Bodies come into play. Currently, 20 countries have established or designated a central agency of this kind.

Sam: How exactly does a Health Data Access Body help a medtech company that wants to build a new diagnostic tool?

Maya: They act as a highly trusted, centralised gatekeeper and an enabler for the entire ecosystem. They process complex data access applications, issue the necessary legal permits, and provide highly secure processing environments where researchers and innovators can access high-quality datasets that are properly pseudonymised or anonymised. It gives developers a legally compliant pathway without forcing them to negotiate fragmented data sharing agreements with many individual hospitals.

Sam: France has had a national Health Data Hub doing this since 2019. Finland uses an authority called Findata, established under the 2019 Act on the Secondary Use of Health and Social Data. It requires all key data controllers to produce standardised metadata.

Maya: And the financial and reputational stakes for getting privacy wrong are incredibly high, which explains why so many hospital administrators are inherently risk-averse. The report mentions that under the GDPR in the European Union, penalties for unauthorised access, processing, or the loss of personal data can reach up to 4% of global turnover.

Sam: And it is not just Europe. In Australia, amendments to the Privacy Act in 2022 drastically increased penalties for a repeated or serious offence. With fines like that, it is no wonder hospital administrators avoid sharing data.

Maya: Which is why having clear, government-backed national frameworks, like Health Data Access Bodies, takes that immense legal pressure off individual hospitals. Another vital enabler is simply having the underlying technical infrastructure and computing power to train and run these highly complex, data-heavy models. The report notes that 25 countries have now defined a national cloud infrastructure for health to handle high-performance computing securely.

Sam: So we have the data, the legal pathways, and the cloud computing. Moving to the guardrails pillar, how do health systems actually buy and evaluate these tools? The OECD found that only 11% of countries are adapting national procurement guidelines to account for these solutions.

Maya: That is a massive bottleneck. If national procurement guidelines are outdated, individual hospitals are forced to figure out cybersecurity, interoperability, and assurance requirements from scratch every time they want to buy a new tool. Modern procurement sends a consistent signal to the market.

Sam: But when they get it right, the scale is incredible. The United Kingdom just launched a massive tender to integrate these technologies into the NHS, focusing on diagnostics and operational efficiency.

Maya: Once a tool is procured, you have to prove its value so it can be reimbursed. The traditional process is called a Health Technology Assessment. But only 24% of countries are streamlining their national approach to health technology assessments to include these new adaptive models.

Sam: Traditional medical devices, like a standard MRI machine or a pacemaker, stay exactly the same after they are manufactured. But an algorithm learns, adapts, and changes its outputs based on new data it encounters in the real world. So the health technology assessment has to completely evolve to look at ongoing performance drift, long-term safety, and algorithmic transparency. How are countries adapting their frameworks to handle that?

Maya: France has a really innovative approach called the PECAN scheme. It provides a 1-year derogatory reimbursement for mature digital health tools. This allows developers to validate their clinical benefits while the solution is already being used and reimbursed.

Sam: Belgium is pushing for something similar. Their National Institute for Health and Disability Insurance introduced a fast-track evaluation process aiming for reimbursement decisions within 270 days.

Maya: And Germany has the DiGA Fast-Track, which enables rapid reimbursement for certified digital health applications that demonstrate a positive health impact. These accelerated pathways are crucial for developers trying to survive pilot purgatory and scale globally.

Sam: But before a tool even gets to reimbursement, it has to be deemed safe by regulators. 23 OECD countries have integrated specific risk classifications into their medical device regulations. But the report highlights a major problem: there is no global alignment on how risk is defined.

Maya: The classic example they give involves ambient voice scribes. These are tools that listen to a doctor-patient conversation and automatically generate the clinical notes.

Sam: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Maya: And now, back to the document.

Sam: It seems straightforward, but the United States and the European Union look at the risk completely differently, right?

Maya: Exactly. In the United States, the FDA evaluates risk based on the output. Because a human doctor is in the loop to review and sign off on the final notes, the scribe tool is considered low risk.

Sam: But the European Union evaluates both the input and the output. Because the tool uses highly sensitive personal health information as its input to generate the text, it was originally classified as high risk there.

Maya: If you are a software developer trying to build a global product, that regulatory fragmentation is a nightmare. It adds massive costs, duplicates efforts, and delays patient access to beneficial tools.

Sam: A major way countries are trying to help innovators navigate these complex rules and gather evidence is through regulatory sandboxes. The report says 18% of countries have established a national approach for a regulatory sandbox with a focus on health.

Maya: A sandbox is a controlled environment where tools can be tested in real-world settings under heavy regulatory supervision. For instance, the United Kingdom recently launched its AI Airlock, which is their initial regulatory sandbox specifically for health. Denmark has also launched a sandbox to significantly reduce the time from development to operation.

Sam: It is a great way to gather real-world evidence safely. But what about overarching legislation? Are countries writing new laws specifically for these medical algorithms?

Maya: Very few. Only 3% of countries are developing national legislation specific to artificial intelligence in health. However, 22 countries are in the process of adopting national legislation to compliment the broader European Union AI Act.

Sam: Looking at high-level strategy, 27 countries have a general national strategy for artificial intelligence that includes some reference to healthcare. But only 18% have a strategy or action plan specifically dedicated to the intersection of health and these algorithms.

Maya: The countries that do have specific health strategies take interesting approaches. Finland has built a comprehensive vision for 2035. France focuses heavily on citizen rights, building trust, and ensuring transparency.

Sam: Spain highlights reliability and universality. Ireland focuses on using the tools to improve operational efficiency and the experience of both patients and healthcare staff. And the United States emphasises democratising the technologies to promote wider access.

Maya: Regardless of the specific focus, all these strategies are ultimately aiming for what the industry calls the quintuple aim for healthcare improvement. That means better health outcomes, improved patient experience, improved provider experience, greater equity, and reduced costs.

Sam: Achieving that requires strong governance. Yet only 18% of countries have established a national oversight body to coordinate all of this. What does an oversight body actually do in practice?

Maya: An empowered national oversight body measures post-market effectiveness to ensure tools remain safe after deployment. They rigorously track any algorithmic performance drift, manage formal incident reporting, and ensure the tools are continually evaluated against public interest objectives. In New Zealand, Health NZ established a National Artificial Intelligence and Algorithm Expert Advisory Group. By requiring all projects to be registered with them, they oversee deployments and focus heavily on mitigating operational risk and algorithmic skewness.

Sam: Luxembourg recently designated their Agency for Medicines and Health Products as their national oversight body. And the United Kingdom has the NHS AI Safety Lab to ensure safe tools are brought into the system.

Maya: We have talked a lot about the technology, the data, and the regulators. But the engagement pillar of the OECD checklist focuses heavily on the health workforce. If the frontline staff do not trust the tools, they will never scale.

Sam: The report says 29% of countries are establishing a national approach to improve workforce awareness and competency. And the impact on the workforce can be life-changing. We mentioned ambient voice scribes earlier. The report cites evidence from the American Medical Association showing that ambient voice technology reduced documentation time by 85.8%.

Maya: That is an astonishing reduction in administrative burden. In practical terms, it saves an average of 5.27 minutes per case compared to manual typing. For a busy clinician, that gives them valuable time back to actually focus on patient care and reduces burnout significantly.

Sam: But integrating these tools requires training. It is not just handing a doctor a tablet. Providers need to understand the parameters of the tools and know how to effectively challenge the outputs when the algorithm contradicts their clinical judgment.

Maya: Several countries are investing heavily to make sure their workforce is prepared for this shift. France has taken a very aggressive approach by mandating training across all health professional programmes starting in the 2025-2026 academic year. They backed this mandate with substantial funding to effectively train 500 000 professionals over a 5-year period.

Sam: That is a massive national commitment. Estonia is taking a different route with their AI Leap 2025 program. They are expanding foundational literacy by targeting 20 000 high school students and training 3 000 teachers. They are preparing the next generation before they even enter medical school.

Maya: Germany offers certified training through a national platform called KI-Campus and is advancing plans for a national Masters in Digital Medicine. Building this capacity is essential because it is not just doctors using these tools. We need technologists, data stewards, privacy officers, and administrative staff to support the infrastructure.

Sam: The engagement pillar also extends to the public. 15 OECD countries have developed national public engagement initiatives to increase public understanding, and 16 countries have established formal public assemblies to help design these systems.

Maya: Public assemblies are a fantastic mechanism. France established a Citizen Assembly for Digital Health where diverse members of the public discuss issues and make recommendations to the government. Finland uses their AuroraAI program to systematically involve the public in co-designing healthcare services.

Sam: Why is it so important to have public assemblies for highly complex technical tools?

Maya: Because trust is not generated by technical performance alone. Engaging the public ensures that the tools reflect diverse perspectives and respect cultural differences. It transforms patients from passive data sources into empowered participants in their care.

Sam: That leads perfectly into the final pillar of the OECD checklist: trustworthiness and bioethics. The document stresses that humans must always remain at the core of any solution brought into the health ecosystem.

Maya: The framework relies on 4 foundational bioethical principles. Patient beneficence, patient non-maleficence, patient autonomy, and justice. The OECD argues that whenever developers and buyers evaluate tools for scale, they must explicitly state how their solution aligns with those specific bioethical principles.

Sam: To make that transparency operational, the report strongly advocates for the use of model cards. We are starting to see this with CE marking in Europe and the Coalition for Health AI, or CHAI, in the United States.

Maya: A model card acts like a highly detailed, standardised nutrition label for a clinical algorithm. It comprehensively summarises the intended use cases, the specific performance metrics, the known limitations, and the representativeness of the data it was trained on. And because these algorithms adapt over time, the report insists the model card has to be treated as a living document that gets updated regularly to reflect model changes, retraining, and real-world performance.

Sam: It is incredible how all these pieces connect. You cannot have a reliable model card without fair data, and you cannot get fair data without Health Data Access Bodies and interoperability.

Maya: Exactly. To escape pilot purgatory, health system leaders and medtech executives need to stop looking at these tools as isolated software purchases. It is an enterprise-wide transformation. You need the robust data foundations, the updated procurement and reimbursement pathways, and massive investments in workforce training to ensure you scale fast while doing no harm.

Sam: To recap, escaping pilot purgatory means actively building the data foundations, modernising the procurement guidelines, and training the workforce. If you want to dive deeper into the checklist, the source document is linked in the show notes.

Maya: And as always, a quick reminder that this podcast is not medical advice. Thanks for joining us for this deep dive into the OECD blueprint.