Dutch healthcare faces significant challenges. An aging population leads to rising demand for care, while the structural staff shortage increases the workload on doctors, nurses, and support staff. One of the biggest sources of frustration and lost time in daily practice is the administrative burden. Here, Generative AI and in particular Large Language Models (LLMs) offer new perspectives. This article provides an in-depth, factual analysis of the ways in which this technology is currently being deployed in healthcare, the applicable legal frameworks, and the crucial ethical boundaries.
Concrete Deployment of AI Language Models in Current Practice
The integration of AI in healthcare has transformed in recent years from a theoretical promise to concrete, operational applications. Currently, these applications focus primarily on language-intensive processes.
Administrative burden relief and reporting
The most mature application of LLMs in the healthcare sector is so-called ambient clinical intelligence, or the automated generation of medical reports. During a consultation (in-person or via video call), a speech-to-text model records the conversation between the healthcare provider and the patient. Subsequently, an LLM automatically structures this transcript into the standard medical format, such as the SOAP method (Subjective, Objective, Assessment, Plan).
The model is trained to ignore irrelevant conversation fragments (such as greetings or small talk) and correctly categorize medically relevant information. Although hard, nationwide data on time savings per discipline are still being gathered, small-scale pilots in Dutch hospitals and general practices consistently indicate that this technology has the potential to significantly reduce the time spent on reporting. This allows the healthcare provider to focus more attention on the patient during the consultation, rather than on the screen.
Triage support and patient communication
Another emerging application is at the 'front door' of healthcare: triage and digital patient communication. Patients are increasingly using portals to request e-consultations. LLMs are deployed to draft concept responses to common, low-complexity questions. In addition, language models can assist with the initial symptom intake. By processing natural language, an AI system can summarize the patient's complaints and present this, linked to an urgency classification, to the triage officer. It is essential here that the model does not diagnose the patient independently, but prepares the human professional.
Why Administrative Applications Grow Faster Than Clinical Ones
When analyzing the adoption rate of AI in healthcare, there is a clear difference between administrative deployment on the one hand and direct clinical or diagnostic deployment on the other. This difference is entirely attributable to the risk profile and the nature of the technology.
Language models are by nature generative and probabilistic; they predict the next word based on patterns in their training data. When summarizing a conversation that has just taken place, the context is already defined. If the model misses a minor nuance in a conversation report, the doctor can easily correct this during review. The margin of error here relates to efficiency, not patient safety.
Clinical applications, where an AI model independently proposes a treatment plan or diagnoses a rare disease based on patient records, carry a much higher risk. A phenomenon such as 'hallucination' (generating factually incorrect but highly convincing-sounding information) is unacceptable in a clinical setting. Therefore, we see that the healthcare sector currently prefers the administrative route, pending more specialized and robust architectures, such as reasoning models that are explicitly trained to verify logical steps before generating an answer.
Strict Requirements: Medical Devices and the AI Regulation
The inzet of AI in the healthcare sector is bound by one of the strictest regulatory frameworks in the world. Software is not simply admitted into the medical domain. The two main legal pillars in this regard are the Medical Device Regulation (MDR) and the new European AI Regulation.
The Medical Device Regulation (MDR)
Under the European MDR (Regulation 2017/745), software can be classified as a medical device. The key question is the software's intended purpose. If an LLM is used purely to transcribe and linguistically structure text, it generally falls outside the scope of the MDR, or is classified as Class I (lowest risk).
However, as soon as the LLM is given functionalities related to triage, diagnostics, disease prediction, or recommending specific medication based on patient characteristics, it almost certainly qualifies as a Class IIa or higher medical device. This means manufacturers must provide extensive clinical evidence, set up strict risk management processes, and pass a conformity assessment by a Notified Body before obtaining a CE marking.
The European AI Regulation (AI Act)
On top of the MDR, the EU AI Act is entering into force. This horizontal legislation classifies AI systems based on risk. A crucial intersection is that AI systems falling under the MDR as a medical device and requiring a third-party assessment are automatically classified as 'high risk' under the AI Regulation.
This introduces additional requirements on top of medical regulations, specifically targeting the nature of artificial intelligence. This includes strict obligations regarding data governance (to prevent bias in training data), extensive technical documentation on the model's architecture, and robust mechanisms for human oversight throughout the system's entire lifecycle.
Oversight in practice: In the Netherlands, oversight of these complex, overlapping regulations lies with multiple authorities. The Health and Youth Care Inspectorate (IGJ) is the primary regulator for medical devices and safe healthcare applications. For a deeper understanding of how this is institutionally organized, you can read more about oversight in the Netherlands.
Privacy, Data Sharing, and Professional Secrecy
An article about AI in healthcare is not complete without addressing the fundamental aspects of privacy and data protection. Under the General Data Protection Regulation (GDPR), health data qualifies as 'special category personal data'. The processing of this data is prohibited in principle, unless there is an explicit legal exception or explicit consent from the patient, combined with medical professional secrecy (laid down in, among others, the WGBO).
The use of public, consumer-oriented LLM services is out of the question in healthcare practice. When a doctor enters patient information, even anonymized, into a publicly accessible chatbot whose data may be used to further train the model, this constitutes a serious data breach and a violation of professional secrecy.
Healthcare institutions wishing to deploy AI must therefore ensure closed ecosystems. This can be achieved by signing strict processor agreements (explicitly agreeing that patient data will not be used for model training and will be deleted immediately after processing) or by deploying locally hosted models on the hospital's own servers. For more general information on securing data in this context, we refer to our guide on privacy and data security within business environments.
Supporting versus Deciding: The Role of the Healthcare Professional
Despite the impressive capabilities of language models, the fundamental ethical and legal rule in healthcare remains: the human professional is always ultimately responsible. AI systems in healthcare are designed according to the Human-in-the-Loop (HitL) principle. They support decision-making but do not make the decision themselves.
A real danger here is so-called 'automation bias'. This occurs when a healthcare provider over-relies on the output of an automated system and fails to critically verify it. Because LLMs produce text with a high degree of linguistic confidence, a doctor (especially under high workload) may be inclined to approve a generated report or triage advice too quickly.
To counteract this, significant investments are being made in designing user interfaces that build in friction. For example, a system can force the healthcare provider to actively accept the AI-proposed urgency class with an extra click, or it can explicitly highlight which sentences were generated directly by the model and which were added manually. Training staff is essential; they must learn not only how to operate the technology, but especially how to recognize its limitations.
Conclusion
The deployment of AI language models in Dutch healthcare is gaining momentum. While clinical decision support is still hindered by justified concerns about safety, reliability, and strict regulations such as the MDR, administrative applications are already proving their value today. Automated reporting and patient communication offer a tangible solution to the towering administrative burden, provided this is done within the strict frameworks of the GDPR and the AI Regulation. The successful healthcare organization of the future will not be the one that has care delivered exclusively by AI, but the one that best integrates AI as a tireless assistant to the human professional.