Chapter 3

The AI-Native Overhaul

While Chapter 2 established the current baseline of the Dutch healthcare system - characterized by a 62% effective capacity and a 38% violation rate of legal wait times - this chapter examines a proposed “target state” known as the SOLL model. In this version of the simulation, we keep the same 45 citizens, three general practitioners, and single hospital, but we adjust five specific parameters to measure the systemic impact of an AI-native architecture.

The Mathematics of Capacity

The primary shift in the SOLL model is the dramatic reduction of administrative tasks. By changing only five variables, the simulation shows a significant increase in the time doctors can spend with patients:

0.62
IST Effective Capacity
0.83
SOLL Effective Capacity
+34%
Clinical Availability Gain

This represents a 34% increase in clinical availability without the need to hire additional staff. In practical terms, a practice designed for 2,400 patients would gain roughly 500 extra “slots” for patient care. If applied across all 12,000 practices in the Netherlands, this would create approximately 6 million additional patient-equivalent openings nationally.

Five Key Interventions

These changes do not require new medical discoveries or new buildings; instead, they focus on restructuring how information moves through the system.

InterventionChangeTechnical Mechanism
Ambient AI ScribesAdmin 30% → <5%Systems record consultations and automatically generate medical records.
AI-Driven TriageWait 12wk → <4wkAI matching routes patients to specialists based on urgency and availability.
Digital TwinsReactive → ProactiveRisk scores trigger a doctor’s visit before physical symptoms appear.
A2A SchedulingIsolated → CoordinatedSoftware agents share real-time schedules to book referrals in hours.
Data SharingInterop 11% → 66%Shared standards allow specialists to see a patient’s full history instantly.

The 5% administrative burden remaining in the model accounts for essential human tasks, such as a doctor signing off on a complex referral or reviewing an AI-generated summary.

Illustrative Example: The Digital Twin

A “Digital Twin” in this simulation is not a creative AI; it is a risk calculator that reads a patient’s medical history. For example, if a citizen named Hendrik has a history of heart disease and COPD, his Digital Twin might calculate a 32% probability of heart failure. Because the system flags this early, Hendrik visits his GP while his condition is still “mild.” Treating a mild condition is generally cheaper and more effective than waiting for a patient to arrive at the hospital in a “critical” state.

Proactive vs. Reactive The key difference is timing. In IST, patients seek care when symptoms become unbearable. In SOLL, the system detects risk trajectories and intervenes before the patient reaches a critical state. This shifts the cost curve from expensive emergency interventions to routine primary care.

The Agentic Mesh: How the System Communicates

The SOLL simulation uses a three-layer structure to allow different parts of the healthcare system to “talk” to one another. We call this the Agentic Mesh.

Agent Cards

Every doctor and hospital department publishes a machine-readable “Agent Card.” This card describes what they do, how long their waiting list is, and their current capacity. Instead of a GP sending a fax and waiting weeks for a reply, the GP’s digital agent checks the specialist’s Agent Card and negotiates a referral almost instantly.

FHIR Memory Store

Every action in the simulation - a diagnosis, a prescription, or a wait time - is stored as a standardized “resource.” This ensures that every agent has access to the same history, preventing the data fragmentation seen in the current system.

The Cognitive Loop

Every agent follows a three-step cycle: they look at their Memory (past visits), they Reflect on their current state (e.g., “I have more time because my admin work is done”), and they Plan their next action (e.g., “I will refer this patient to the specialist with the shortest queue”).

Results: IST vs. SOLL

When we compare the two systems across 200 simulation runs, the outcomes shift noticeably:

MetricIST (Baseline)SOLL (Target)Change
Specialist Wait9.4 weeks<4 weeks−58%
Treeknorm Violations38%<10%−74%
System-Failure Deaths4.2 / run<1 / run−76%
Effective Capacity62%83%+34%
Worker Shortage Offset301K projected~110K offsetTechnology-driven efficiency

Research Limitations and Bias

As mentioned previously, I have worked in Big Tech for the last six years. This simulation is a personal project used to explore these ideas, and the results are based on theoretical parameters.

There are several reasons these results might be different in the real world:

Specific limitations of the SOLL model
  • Small Sample Size: With only 45 citizens, a single complex case can change the results significantly.
  • Adoption Barriers: The model does not account for the time it takes for doctors to trust new tools or for laws to change.
  • Human Behavior: If a doctor saves 25% of their time, they might choose to spend more time with each patient rather than seeing more patients. The model assumes they see more patients to fix the backlog, but reality is often more complex.
  • Infrastructure Cost: The model does not price the implementation cost of building an Agentic Mesh across 12,000 practices.
  • Parameters derive from published sources (CBS, RIVM, NZa, IZA, LHV) but the SOLL targets are projections, not observations.
Sources & References
  1. IZA - Integraal Zorgakkoord 2023–2026 (transformation targets)
  2. ABF Research - Arbeidsmarktprognose zorg en welzijn 2023–2033
  3. NZa - Tariff schedules and capacity norms 2025
  4. Park et al. - Generative Agents: Interactive Simulacra (Stanford, 2023)
  5. HL7 FHIR R4 - International healthcare data standard
  6. Wegiz - Wet elektronische gegevensuitwisseling in de zorg (2023)

The Purpose of the Model

The goal of this simulation is not to provide a perfect prediction of the future. Instead, it provides a set of numbers that we can argue about. You might think that a 5% administrative load is too low, or that the “AI multiplier” is too optimistic. That is a useful disagreement because it focuses on specific parameters rather than vague opinions. By using this simulation, we can test different ideas and see which ones actually have the potential to fix the “care gridlock” in Dutch healthcare.

Variables, Not Vibes The value of the SOLL model is not in its predictions but in its testability. Every parameter can be changed and every result can be re-run. If you disagree with the assumptions, adjust them and see what happens. The code is open source.
🎮 Try the Simulation

Data Sources & References

  1. CBS - 172 duizend mensen overleden in 2024 (mortality statistics)
  2. CBS - Bijna een derde werktijd zorg gaat op aan administratie (~31% admin burden)
  3. CBS / AZW - Ziekteverzuim in zorg en welzijn 2024 (7.3–8.9% sick leave)
  4. CBS / VZinfo - Levensverwachting per leeftijd en geslacht (M: 80.5, V: 83.3 jaar)
  5. RIVM / VZinfo - Chronische aandoeningen en multimorbiditeit (96% of 75+ has ≥1 chronic condition)
  6. ABF Research / VWS - Prognosemodel Zorg en Welzijn (301,000 worker shortage by 2035)
  7. NZa - Maximumtarieven huisartsenzorg 2025 (€12.43 regulier consult, BR/REG-25136)
  8. NZa / Staat van VenZ - Overschrijding Treeknorm wachttijden (specialist wait time violations)
  9. LHV - Peiling Administratieve lasten huisartsen 2023 (GP admin burden survey)
  10. IZA / Rijksoverheid - Integraal Zorgakkoord 2023–2026 (€2.8B transformation budget)
  11. Chen et al. - Risk of cardiovascular comorbidity in COPD: systematic review. Lancet Respir Med 2015; OR 2.46 for CVD.
  12. ESC - 2021 ESC Guidelines on CVD Prevention in Clinical Practice. Eur Heart J 2021;42(34):3227–3337. Diabetes RR≈2.0, Hypertension RR≈1.8.
  13. Park et al. - Generative Agents: Interactive Simulacra of Human Behavior. Stanford/Google, 2023.