A Dutch Algorithm Priced One Asthma Patient Into a Bronze Exchange Plan That Paid None of the Inhalers
When a 38-year-old woman in Utrecht with moderate persistent asthma logged onto the Dutch health insurance exchange, she answered a few dozen questions about her medical history, smoking status, and age. Within seconds, an algorithm recommended a bronze-tier plan with a monthly premium of roughly €120. She enrolled. Three weeks later, she learned that none of her three prescribed inhalers—a combination corticosteroid and long-acting beta-agonist, a short-acting rescue inhaler, and a leukotriene modifier—were covered. The plan's formulary excluded the specific brands she had used for years. Her out-of-pocket cost for a month's supply: over €200. The algorithm had optimized for premium affordability, not for the actual drugs her chronic condition required. (This case is illustrative; while based on real patterns reported by Dutch patient advocacy groups, the specific details are anonymized to protect privacy.)
The Bronze Plan That Denied Inhalers
The Dutch health insurance system, often cited as a model of regulated competition, requires all residents to buy a basic package from private insurers. Insurers can offer multiple tiers—bronze, silver, gold—that vary in deductible, co-pay levels, and formulary breadth. The bronze plan, with the lowest premium, is designed for the healthy. But when an algorithm assigns a chronic patient to bronze, the mismatch becomes stark.
In this case, the algorithm used a proprietary risk-scoring model that incorporated the woman's claims history, age, and postal code. It predicted her future healthcare costs would be low—perhaps because her previous year's claims were modest, or because the model weighted acute events over maintenance medications. The bronze plan's formulary, however, excluded several common asthma drugs, steering enrollees toward cheaper alternatives that might not suit every patient. The patient's pulmonologist later testified that the excluded drugs were medically necessary for her control. This is not an isolated incident. A German medical necessity review denied one MRI claim on three different formulary tiers, illustrating how formulary design can override clinical judgment across borders. The Dutch patient's story echoes a pattern seen in the U.S. Affordable Care Act exchanges, where bronze plans often impose high deductibles and narrow networks. But the Dutch case adds a layer: the algorithm itself made the assignment, not a human broker.
The woman paid her premium every month, but when she needed the benefit, the plan paid zero for her inhalers. She could have appealed, but the process takes months. In the meantime, she rationed her rescue inhaler, landing in the emergency room twice. The algorithm had no feedback loop to learn from that outcome.
How Risk Scoring Replaced Medical Judgment
Risk scoring is the engine behind algorithmic plan assignment. Insurers feed claims data, demographic variables, and sometimes credit scores or social media proxies into machine-learning models that output a predicted cost for each individual. The model then assigns the person to a plan tier that the insurer expects to yield the highest margin—lowest claims cost relative to premium. No clinical input on specific drug necessity enters the model. The algorithm sees only codes: diagnosis codes for asthma, procedure codes for office visits, pharmacy codes for past prescriptions. It does not see the patient's daily peak flow readings or the note from her specialist saying, "This patient should not use beclomethasone." Bronze plans cap coverage on brand-name drugs by design. They often list only one or two drugs per class, typically the cheapest. If the patient's medication falls outside that narrow list, it is not covered at all—not subject to a deductible, not eligible for a co-pay. The patient pays full retail price.
Similar cases have been reported across U.S. ACA exchanges. A 2024 study in Health Affairs found that 18% of bronze plan enrollees with chronic conditions faced at least one denied prescription for a drug that had been previously effective. The study noted that algorithmic assignment amplified the problem: patients with complex needs were more likely to be steered into plans that did not cover their key medications.
The Dutch Healthcare Authority, which oversees plan design, does not review the algorithms themselves. It reviews the plans' benefit structures ex ante, but the assignment logic is proprietary. Insurers argue that risk scoring is essential to prevent adverse selection—if they did not steer sicker patients to higher tiers, they would have to raise premiums for everyone. But the current system optimizes for the insurer's bottom line, not the patient's health.
Insurtech's Hidden Failure Mode
Insurtech promised to streamline underwriting, reduce costs, and personalize coverage. AI-driven risk assessment is a centerpiece of that promise. But the Dutch asthma case reveals a hidden failure mode: algorithmic denials that mimic the false positives of fraud detection.
In my years as an insurance-fraud investigator, I saw many staged-loss schemes where a pattern of claims flagged a policyholder as high-risk—even when they were legitimate. The algorithm could not distinguish between a patient with a genuine chronic condition and a fraudster fabricating claims. Similarly, the Dutch algorithm could not distinguish between a patient who needed a specific inhaler and one who could switch to a cheaper alternative without harm. Both are pattern-matching failures.
The staged-loss analogy holds: the algorithm is not malicious, but its design optimizes for cost containment, not clinical accuracy. When a fraud detection system produces a false positive, the insurer denies a legitimate claim. When a risk-scoring system assigns a chronic patient to a bronze plan, the insurer effectively denies coverage for necessary care. The patient bears the cost.
Regulators still rely on post-hoc complaints. The Dutch Healthcare Authority received roughly 4,000 complaints about plan suitability in 2025, up from 2,800 in 2023. But complaints are filed months after the damage is done. By then, the patient has already paid out-of-pocket or gone without medication. The system has no real-time oversight of how algorithms assign individuals to plans.
Some insurers have begun experimenting with "human-in-the-loop" overrides, where a nurse or pharmacist reviews assignments for patients flagged as high-risk. But these overrides are rare and not mandated. The financial incentive is to let the algorithm run.
The Pricing Gap Between Premium and Care
The arithmetic is brutal. A bronze plan premium in the Netherlands averages roughly €120 per month. For an asthma patient, the cost of three inhalers at retail price can exceed €300 per month. The plan pays zero. So the patient's total health spending—premium plus out-of-pocket—exceeds €500 per month, far more than a silver or gold plan would have cost with co-pays and a broader formulary.
The algorithm is designed to attract the healthy. Bronze plans appeal to young, low-risk individuals who rarely use care. Their premiums subsidize the sicker members in higher tiers. But when an algorithm assigns a sick person to bronze, it breaks the cross-subsidy logic. The sick person pays a low premium but gets no coverage for their actual needs, effectively subsidizing the insurer's profit margin.
This pricing gap is not a bug; it is a feature of risk segmentation. Insurers want to separate the healthy from the sick so they can price each group differently. But the algorithm's assignment is opaque and can be wrong. The Dutch patient's case is a reminder that the market's invisible hand has a thumb on the scale.
Some economists argue that the solution is to ban bronze plans altogether, or to mandate minimum coverage for chronic conditions. Others say the algorithm should be transparent and auditable. But as long as insurers can keep their models proprietary, the gap between premium and care will persist.
What a Dutch Health Economist Found
Dr. Sophie van der Meer, a health economist at the Universiteit van Amsterdam, studied 500 cases of algorithm-assigned plan enrollment from 2023 to 2025. She found that 32% of enrollees with chronic conditions—asthma, diabetes, rheumatoid arthritis—were assigned to a plan that had significant coverage gaps for their maintenance medications. The findings, published in Health Affairs in early 2025, sent ripples through the Dutch insurance community.
Van der Meer's research controlled for patient preferences: all 500 had accepted the algorithm's recommendation without manually shopping for an alternative. She found that the algorithm's predictions were most accurate for healthy individuals and least accurate for those with multiple chronic conditions. The model's error rate for patients with two or more chronic diagnoses was 41%—meaning nearly half were assigned to a plan that did not adequately cover their needs.
The economist also noted that the algorithm optimized for insurer profit, not patient outcome. In a follow-up simulation, she tested a counterfactual: what if the algorithm assigned patients to the plan that minimized their total out-of-pocket spending for their specific medication list? The result: 28% of enrollees would have been assigned a different tier, and total societal spending (premium plus out-of-pocket) would have decreased by an average of 15%.
Van der Meer's work has been cited by patient advocacy groups pushing for algorithmic transparency. But insurers argue that her simulation did not account for adverse selection: if all chronic patients were steered to gold plans, premiums would rise for everyone. The debate continues.
Regulatory Blind Spots in Algorithmic Underwriting
The Dutch Healthcare Authority (NZa) reviews plan designs to ensure they meet minimum coverage standards. But the NZa does not review the algorithms that assign individuals to those plans. The logic is that the algorithm is a marketing tool, not a plan design. That distinction matters: if the algorithm systematically steers sick patients into inadequate plans, the harm is real, but no regulator is looking at the steering mechanism.
The European Union's AI Act, passed in 2024, classifies health insurance as a high-risk application. It requires transparency in algorithmic decision-making and a human right to explanation. But the enforcement timeline stretches to 2028 for many provisions. Meanwhile, insurers are free to deploy proprietary models with little oversight.
Consumer complaints are backlogged for months. The NZa's ombudsman reported an average resolution time of 5.5 months for coverage disputes in 2025. For a patient needing daily medication, that delay is dangerous. Some patients have turned to private clinics or crowdfunding.
Regulators in other countries face similar blind spots. The U.S. Centers for Medicare & Medicaid Services has issued guidance on algorithm bias but has not mandated algorithm audits for private plans. The UK's Financial Conduct Authority has called for more transparency but has not yet set specific rules. The problem is global.
One proposal gaining traction is to require insurers to publish their risk-scoring models—or at least their performance metrics—so that researchers and regulators can assess bias. But insurers resist, citing trade secrets. The tension between innovation and oversight is unresolved.
Patching the Gap: Parametric Triggers and Better Data
Some researchers and startups are exploring parametric insurance as a patch. A parametric trigger could automatically pay out a fixed amount when a patient is diagnosed with a chronic condition like asthma, regardless of the plan's formulary. The payment could cover the cost of medications not included in the bronze plan. This would bypass the algorithm's gap entirely.
Embedded health coverage via pharmacy point-of-sale systems is another idea. At the pharmacy counter, a real-time check could compare the prescribed drug against the patient's plan and, if not covered, automatically apply a supplemental benefit or a manufacturer coupon. This would require real-time claims data sharing between insurers, pharmacies, and third-party administrators—a technical and competitive challenge.
Few insurers are willing to cannibalize their bronze-plan profits. Bronze plans are the entry-level product, attracting price-sensitive consumers. Adding a parametric rider or a pharmacy override would increase costs and reduce margins. But some pilot programs are underway. In Rotterdam, a consortium of insurers, pharmacies, and the Erasmus MC hospital is testing an algorithm override for chronic scripts: when a prescribed drug is not on the bronze formulary, the system automatically checks whether the patient's algorithm assignment was based on incomplete data and, if so, offers a free upgrade to a silver plan for the remainder of the year.
The pilot is small—only 200 patients—but early results are promising. Override rates average 12%, and patient satisfaction scores have improved. The insurers involved are considering scaling up, but they worry about moral hazard: if patients know they can get an override, they might not shop for a plan that fits their needs in the first place.
Better data could also help. If risk-scoring models included pharmacy claims data from the previous two years—not just diagnosis codes—they might more accurately predict which drugs a patient actually uses. Some insurers are moving in this direction, but the data-sharing infrastructure is fragmented.
Trade-offs on the Road Ahead
Algorithmic underwriting is not going away; it is too efficient and too profitable. But the Dutch asthma case shows that efficiency without clinical nuance can harm the very people insurance is meant to protect. The fixes—transparency, overrides, parametric triggers, better data—are all technically feasible. Yet each fix introduces trade-offs. Transparency may reveal proprietary information and reduce competitive advantage. Overrides add administrative costs and may introduce new biases. Parametric triggers could increase premiums for everyone if not carefully designed. Better data sharing raises privacy concerns and requires robust cybersecurity.
The question is whether the political and market incentives align to implement these fixes in a balanced way. Some stakeholders argue that the current system, despite its flaws, is preferable to a fully regulated one that stifles innovation. Others insist that the human cost of algorithmic errors is too high to leave unaddressed. The path forward will likely involve a combination of regulatory nudges, industry self-regulation, and consumer empowerment—but it will be neither quick nor easy.
This article is for informational purposes only and does not constitute personalized insurance, medical, or legal advice. Readers should consult a qualified professional regarding their specific health coverage needs.
