AI in healthcare: where to start, and what can wait

Why start with administration?

Administration is where the time goes, and administrative AI is a smaller step than clinical AI.

There are also not enough people. The World Health Organization estimates a projected shortfall of 11 million health workers by 2030, mostly in low- and lower-middle income countries. WHO adds that countries at all levels of development struggle to educate, deploy and retain their workforce. AI can give part of that time back. According to the European Commission, AI can streamline administrative tasks like patient scheduling, billing and electronic health records management, which can free healthcare professionals to focus more on patient care. WHO’s guidance on large multi-modal models lists “clerical and administrative tasks” as one of five broad applications, for example documenting and summarizing patient visits in the electronic health record.

There is a second reason. Clinical AI, which helps decide what is wrong with a patient, carries more risk and falls under stricter rules. Administrative AI is usually a smaller step, both for the organisation and for the regulator.

We need many more people at the bedside, while healthcare runs an expensive back office with planning and a lot of contact with patients and other organisations. The money AI saves there can go to care by and for people: a no-brainer.

Why the Dutch choice matters

The Netherlands put administration first and diagnostic AI later, with a national programme and money attached.

The Netherlands has written the order down. In a letter to parliament of 6 October 2025, the Dutch health minister launched a national programme for AI in healthcare. The first focus is on reducing the administrative burden. The programme starts with two kinds of tools:

• voice-driven reporting
  the care professional speaks, and software turns it into a note in the patient record;
• capacity planning
  rosters and planning based on demand for care and on who is available.

Diagnostic AI follows at a later stage. According to the letter, the care sector itself considered these first two areas feasible in the short term. The numbers behind that choice are Dutch, but the pattern is familiar to many health systems. A letter to parliament of 2 June 2026 states that Dutch health workers spent around 30% of their working time on administration in late 2025, based on figures from Statistics Netherlands. The parties to the Dutch care agreement have set a shared target of at most 20% by 2030. The same letter projects a shortage of almost 301,000 workers in Dutch health and social care by 2035.

The Dutch plan for the programme was completed in September 2026. It has five lines: AI readiness (including self-evaluation tools), preconditions such as a validation framework and legal and ethical frameworks, AI functions that are broadly usable across care, communication and education, and a national support structure. Support for care organisations starts in early 2027, for example with an AI readiness scan. There is money attached, but not only for AI. The Dutch government has set aside €200 million for 2027, €300 million for 2028 and €300 million for 2029 for, among other things, scaling up labour-saving AI, digital and hybrid care and medical technology. The Dutch care agreement adds an ambition: by 2028, AI should be in everyday use in 50% of the care processes it is meant for. That covers the tools the agreement gave priority to and that have proven their value. That is an agreement between parties, not a law and not a measured result.

What is already in use?

AI is already at work from hospitals to home care, but its integration into clinical practice is progressing slowly.

The Dutch letter of October 2025 gives examples from practice. At Amsterdam UMC, AI helps to assess biopsies after a kidney transplant faster and to assess MRI scans in neuroradiology in a more standardised way. Many care professionals, for example in elderly care, already use voice-driven reporting. At the epilepsy centre Kempenhaeghe, AI automatically structures medical data in patient records. In nursing and home care, organisations are experimenting with AI rostering that combines demand for care, staff availability and routes.

In the United States, the FDA keeps a list of AI-enabled medical devices that are authorised for marketing. The FDA itself notes that the list is not a comprehensive resource. It also says it will explore ways to tag devices that use large language models, so that healthcare providers and patients can recognise when that functionality is present. The European Commission is more cautious about the pace. It writes that AI and machine learning medical devices still have a relatively limited market presence and that the integration of AI into clinical practice is progressing slowly.

Which rules apply to AI?

In the EU, five sets of rules count: AI literacy, high-risk AI, medical devices, liability and health data.

AI literacy (EU). Since 2 February 2025, the AI literacy obligations of the EU AI Act apply, according to the European Commission. Organisations that use AI must work on the AI literacy of their people. That includes hospitals. The obligation remains, but since the simplification of the AI Act the law no longer requires a specific level.

High-risk AI (EU). The Commission names AI-based software intended for medical purposes as an example of a high-risk AI system. Such systems must meet requirements including risk mitigation, high-quality data sets, clear user information and human oversight. For high-risk AI built into products that already have their own EU safety rules, the Commission gives a transition period until 2 August 2028, with AI in robot-assisted surgery as its example. That date is the result of the AI Omnibus, the simplification of the AI Act that entered into force on 27 July 2026. Once such a system is in use, the organisation that deploys it must ensure human oversight and monitoring.

Medical devices (EU). Software used for diagnosis, prevention, monitoring, prognosis or treatment is a medical device under the EU Medical Device Regulation (EU 2017/745). The Dutch health care institute Zorginstituut Nederland adds a useful distinction: voice-driven reporting software is usually not a medical device, but that changes as soon as the system starts to interpret medical data and make diagnoses on that basis. That is a Dutch reading of the EU rules, so check it with your own regulator.

Liability (EU). Under the new Product Liability Directive, software, including AI systems, is a product with no-fault liability, according to the European Commission. A developer of an AI system is treated as a manufacturer.

Health data (EU). The European Health Data Space Regulation entered into force on 26 March 2025. According to the Commission, key parts apply from March 2029, including rules for the secondary use of most data categories, such as data from electronic health records. From March 2031, medical images, lab results and hospital discharge reports should follow for primary use. According to the Commission, the health data space plays a crucial role in giving AI in medicine access to diverse and high-quality health data.

What does WHO advise?

WHO’s six principles start with keeping humans in control of medical decisions.

In 2021, WHO published six principles for AI in health as the basis for regulation and governance:

1. protecting human autonomy, so that humans remain in control of health-care systems and medical decisions;
2. promoting human well-being and safety and the public interest;
3. ensuring transparency, explainability and intelligibility;
4. fostering responsibility and accountability;
5. ensuring inclusiveness and equity;
6. promoting AI that is responsive and sustainable.

In 2024, WHO added guidance on generative AI with over 40 recommendations for governments, technology companies and health care providers. It warns of documented risks of false, inaccurate, biased or incomplete statements, and of “automation bias”: professionals and patients overlooking errors they would otherwise have spotted, or handing difficult choices to the system.

Where does care get stuck?

Data, trust, financing and integration into daily work are the main obstacles.

The European Commission lists several challenges: access to high-quality health data, building trust and acceptance, sustainable financing, particularly in public hospitals, and integrating AI into clinical workflows “not merely as an added tool”. A Commission study groups the obstacles into four areas: technology and data, legal and regulatory barriers, organisational and business obstacles, and social and cultural factors.

The Dutch round tables with the care sector found much the same: no shared vision and strategy, too little knowledge and AI literacy, technical preconditions such as reliable data and systems that can talk to each other, and too little trust in the quality and safety of the AI systems available today. The Dutch minister also sees risks in dependence on non-European technology companies.

How do you prepare?

Start where the time goes, measure before you start and keep the professional accountable.

1. Start where the time goes.
   Look at where your people lose the most time to documentation and planning. That is where the Dutch programme and its funding point too. Then go through it process by process, such as the note after a visit, the shift handover or the roster, and ask: what can AI do, and what stays with the care professional?
2. Measure before you start.
   How much time goes to administration now? Without a baseline you cannot show later what AI has delivered.
3. Check whether a tool is a medical device.
   A tool that only writes notes is usually not. A tool that interprets data to support a diagnosis usually is. In the US, check the FDA list; in the EU, check whether it carries a medical CE marking.
4. Keep the professional accountable.
   According to the Dutch letter, AI is explicitly there to support, and final responsibility lies with the care professional who uses it. WHO’s first principle points the same way.
5. Plan for oversight and errors.
   Decide who checks the output, how often, and what happens when a note or a plan is wrong. Automation bias is a known risk.
6. Train your people.
   AI literacy has been an obligation in the EU since February 2025, and a lack of knowledge is one of the obstacles that both the Commission and the Dutch care sector name.
7. Get your data in order.
   Clinical AI later depends on data you organise now. In the EU, the health data space sets the timetable.

The future of healthcare

A sober order of steps, while Microsoft’s MAI-DxO still has to prove itself with real patients.

The Dutch choice is a sober one: first win back the time that now goes into paperwork, and only then take the larger step towards AI that supports diagnosis. We wrote earlier about how generative AI is changing the role of doctors, and about the wider picture in The insane future of healthcare. What that larger step could look like, Microsoft showed in October 2026: in its own benchmark, its MAI-DxO system reached 83% diagnostic accuracy on complex NEJM cases, against 36% for doctors with reference material. A study with real patients is still to come. Whether the Dutch target of at most 20% administration time by 2030 will be met, cannot be said yet.

Voice-driven reporting or smarter planning only works if people actually want to use it. AI is one of several new technologies coming to healthcare, and with every new step the question is the same: what does it do for the people at the bedside? That starts with leaders who are curious themselves, give their teams room to try things and stay critical at the same time. Enthusiasm for change and a critical eye belong together here. That side of change, the people and the organisation, is also the subject of the keynote on the organisation of the future.

Frequently asked questions

Where should a hospital start with AI?

Administrative work with a clear outcome, such as voice-driven documentation and staff planning. The Dutch national programme starts there, and the European Commission names scheduling, billing and electronic health records as tasks AI can streamline.

Is an AI scribe a medical device?

Usually not, as long as it only turns speech into a note. According to the Dutch health care institute, that changes as soon as the system interprets medical data and makes diagnoses on that basis. Then the EU Medical Device Regulation applies.

Which EU rules apply to AI in healthcare?

The AI Act (AI literacy since February 2025; requirements for high-risk systems such as AI-based software for medical purposes), the Medical Device Regulation, the Product Liability Directive and, for health data, the European Health Data Space Regulation.

Who is responsible for an AI mistake?

The care professional who uses the tool remains responsible for the decision, according to the Dutch health ministry. Under the EU Product Liability Directive, the developer of a defective AI system can be liable as a manufacturer.

Is there a list of approved AI devices?

In the United States, the FDA publishes a list of AI-enabled medical devices authorised for marketing. The FDA notes that the list is not comprehensive.