Why I Built This
I work at Microsoft. You should assume I'm biased toward AI.
That's exactly why I built a system that could prove me wrong. In several important ways, it did.
Cammelot is a simulated Dutch town. Forty-five citizens with names, ages, pre-existing conditions, and GP assignments. Three general practitioners. One hospital. The citizens walk around a 16-bit pixel map that looks like it belongs on a 1994 Super Nintendo. Their diseases are drawn from 2025 RIVM prevalence data. Their healthcare costs come from NZa tariffs. Their waiting times follow the Treeknorm - the Dutch legal norm for how long you should wait for specialist care.
I built it because I was tired of the vibes.
"AI will solve the zorginfarct." How? Which interventions? In what order? With what side effects?
"AI is overhyped." Compared to what? The system that's already producing a 301,000-worker shortage by 2035?
Both sides argue with anecdotes and intuition. I wanted a machine that produces numbers. Not predictions. Outputs you can measure. Run it, measure it, break it, run it again. Two hundred times. With statistics.
The architecture: how it actually works
Every citizen in Cammelot is an autonomous agent running a cognitive loop inspired by Park et al.'s "Generative Agents" - the Stanford paper that simulated 25 AI agents in a town called Smallville. Their agents organized a Valentine's Day party. Mine enter waiting lists.
Each agent, every simulation cycle, does three things:
1. Memory. Every action - GP visits, referrals, queue entries, diagnoses - is logged as a FHIR R4 resource. The international healthcare data standard. Not a toy format; the actual standard that hospitals and GP systems use. Each agent's "life history" is a chronological FHIR record: Conditions, Encounters, Observations, MedicationRequests.
2. Reflection. Agents assess their own state. A patient might "realize" their breathing is getting worse. A GP might notice their burnout climbing. These reflections drive behavior - when to seek care, when to refer, when to escalate.
3. Planning. Based on memory and reflection, agents act. Patients visit GPs. GPs triage and refer. Specialists process queues. The system runs forward one tick. Repeat 3,000 times.
Disease progression is driven by Markov chains - probabilistic state machines calibrated to RIVM data. A patient with COPD transitions between states: healthy → mild → moderate → severe → terminal. Each tick, the chain rolls. Comorbidities multiply the odds: diabetes worsens heart disease (1.4×), hypertension compounds both (1.3×). When a patient reaches terminal, their sprite turns grey. They become a ghost.
Providers communicate through A2A (Agent-to-Agent) protocol. Every GP and specialist publishes an Agent Card - a machine-readable description of their skills, current wait time, and queue depth. Referrals flow as structured messages. Wait times are real-time, not estimated. Think of it as an evolution of the Personal Health Train concept that the Netherlands pioneered through Health-RI - but with stateful agents that discover each other, instead of stateless containers dispatched to known nodes.
Two worlds: IST and SOLL
The simulation runs in two modes. Each is a complete set of parameters - not a toggle, but a different reality.
IST ("Ist-Zustand" - the current state) — full parameters
This is today's Dutch healthcare system, parameterized from CBS, NZa, LHV, and IZA data:
- 30% of GP time lost to administrative paperwork (NZa/LHV)
- 8% workforce capacity lost to sick leave (CBS)
- 12-week Treeknorm ceiling for specialist care (NZa)
- Basic clinical triage - severity-based, but without AI enhancement
- No proactive monitoring - problems are detected when patients show up
- No AI assistance - efficiency multiplier is 1.0
- Standard disease progression - Markov chains tick identically for treated and untreated patients
The effective GP capacity in IST: C_total × (1 - 0.30 - 0.08) × 1.0 = 62%. Your GP is operating at sixty-two percent before they see their first patient.
SOLL ("Soll-Zustand" - the target state) — full parameters
This is what an AI-native primary care system could look like. Every parameter has a source, and I'll be honest about which are evidence-based and which are aspirational:
- 5% admin load - ambient AI scribes handle documentation (IZA 2030 target: 20%. I'm more aggressive.)
- 5% sick leave - reduced through better workforce support (aspirational)
- 4-week Treeknorm - AI-driven scheduling + capacity optimization (aspirational - the hardest one)
- Severity-based triage - sickest patients jump the queue, not first-come
- Digital Twin monitoring - continuous risk assessment for every citizen. Alert threshold: 25% risk (vs. 60% in IST). This means problems are flagged much earlier.
- 1.34× AI efficiency multiplier - GPs see 34% more patients effectively (IZA/Vilans research estimate)
- 0.3× HP drain - patients waiting in SOLL deteriorate 70% slower (because of active monitoring + ketenzorg)
- Treatment-modified disease progression - ketenzorg and hospital care actually slow Markov transitions: 20% deceleration on worsening, 5% chance of improving one state
- Fairness guardrail - if any age group's average wait exceeds 120% of the population mean, optimization biases auto-disable
The effective GP capacity in SOLL: C_total × (1 - 0.05 - 0.05) × 1.34 = ~121%. The same GP, nearly double the output.
Wait - that's a lot of changes at once
Yes, and that's deliberate.
I'm not testing whether "one AI tool helps." I'm testing what happens when you redesign the system end-to-end. SOLL isn't "IST plus a chatbot." It's a fundamentally different architecture.
Spoiler: some hold. Some don't. And one mechanism turned out to matter more than all the others combined.
The methodology: why you should take this semi-seriously
Statistical protocol — 200 simulation runs
- 100 independent runs per mode (IST and SOLL) - 200 total simulations
- 3,000 cycles per run - roughly 10 simulated years
- Welch's t-test for significance (unequal variances)
- Cohen's d for effect size (how big is the difference, not just "is it nonzero")
- 95% confidence intervals on every metric
- Every null result published alongside every positive one
N = 45 agents is small. I know. The stochastic variance is high - especially for elderly subgroups, where 3 agents per run means a single death swings mortality by 33 percentage points. I'll flag this limitation every time it matters. And I'll show you the raw data, not just the averages.
Why you should be skeptical of me
Three reasons to discount this work
1. I work at Microsoft. My employer builds and sells AI products. If this simulation showed "AI fixes everything," you should discount it heavily. It doesn't - several findings are embarrassing for the AI-optimist position - but the conflict of interest is real and I won't pretend otherwise.
2. I built the simulation. I chose the parameters, the architecture, and the disease models. A different builder would make different choices. The code is open - you can change anything and re-run.
3. N = 45. This is a proof of concept, not a clinical trial. The findings are directional, not definitive. I use proper statistics to separate signal from noise, but small populations have fundamental power limitations that no amount of t-testing can overcome.
What I can promise: I report every finding honestly. When the data contradicts my hypothesis, I say so. When a result is not significant, I don't dress it up. When I got something wrong in an earlier analysis, I show the correction.
What's coming
Over the next five posts, I'll walk through what happened - in the order I discovered it:
- The first death. Hendrik Veenstra, 70, dead at cycle 71. The aggregate: ~3% fewer deaths in SOLL. Not statistically significant. AI didn't save the patients.
- The admin tax. GP burnout drops 85%. Admin waste drops 83%. Deaths barely move. Fixing the workforce doesn't fix the patients - not by itself.
- The queue. Basic triage still kills the elderly at 5× the rate of the young. Same rules, radically different graves. AI-enhanced triage doesn't fix it.
- The bias. I optimized for throughput. The system pushed back against the elderly. The guardrail fired in 97% of runs - and it actually worked. Bias is significantly lower.
- The treatment gap. 143 proactive alerts per run. Ketenzorg tripled. Deaths still didn't drop. The model taught me: detection without treatment that changes disease trajectories is just documentation of decline.
The code is on GitHub. The simulation runs in a browser. The research runner produces 200 trials of statistical output in minutes.
I'm looking for healthcare professionals, researchers, and skeptics. Especially skeptics.
What would you test in Cammelot?