FRAUD PREVENTION · PREDICTIVE AI
A European health insurer — 4,8 million insured members across seven countries — is preparing to deploy a predictive AI system for healthcare reimbursements. Annual fraud is estimated at €82 million, increasing by 8 % per year. The model has already been technically validated, with projected precision of 78 %.
What remains to be decided is not the model's performance: it is the trigger thresholds, degree of automation, information provided to the insured member, appeal procedure and governance of the system. In other words, how a score becomes a decision.
THE PROBLEM
A predictive system learns regularities present in historical data. Some do describe fraudulent schemes. Others describe a chronic condition, a complex care pathway, a change in circumstances, an age, a territory, or a way of filling in forms.
The model does not distinguish these two families of regularities. It treats them as signal. And when a threshold turns that signal into a control, the issue ceases to be statistical: it becomes a decision made about a person.
The useful question is therefore not “ does the model detect more fraud ? ”, but: whom does it identify as suspicious, and what happens next to that person?
WHAT IS THE SAME FOR EVERYONE
WHAT DIFFERS FOR EACH GROUP
THE FALSE POSITIVE
A legitimate person incorrectly classified as high risk does not experience a statistical error. They experience a control, a delay, a request for documentation, and the feeling of being suspected.
This sequence is an analytical mechanism, not a result. It is a reminder that a model does not merely produce a score: it triggers human consequences, and those consequences are what the simulation examines.
BIAS AND CONSEQUENCE
A statistical bias and a human consequence are not measured at the same point. Between them lie a threshold, an automation rule and a procedure. Ethical analysis must cover the entire chain, not the model alone.
ONE DECISION, MULTIPLE OBJECTIVES
FRAUD CAPTURE
€82 million in estimated annual fraud, increasing by 8 % per year, driven by increasingly sophisticated schemes.
FALSE POSITIVES
Legitimate insured members incorrectly subjected to controls, whose cost is visible neither in the model's precision nor in the savings achieved.
EQUITY ACROSS POPULATIONS
The overall error rate says nothing about how errors are distributed. That distribution is the ethical issue.
AUDITABILITY AND COMPLIANCE
A system that cannot be explained, traced and audited cannot be deployed, regardless of its performance.
WHAT IS TESTED
The protocol does not test “ an algorithm ”. It tests eleven ways of combining a prediction, a threshold, a human decision, information and an appeal. These six dimensions vary from one configuration to another.
5 200 synthetic insured members were interviewed individually, distributed proportionally across the 17 types, and exposed to the 11 configurations combined with different information and appeal arrangements, with dynamic follow-ups at tipping points: perception of the control, feeling of unfair suspicion, satisfaction with the appeal.
SIMULATED POPULATION
The reconstructed population covers 2,2 million European health-insurance members, calibrated on public data — INSEE, DREES, open insurance-sector data — and on the insurer's proprietary data: member types, healthcare-use history, documented fraud history, demographic structure. No personally identifiable data entered the system.
It is structured into 17 types of relationship with insurance and controls, combining age, family structure, chronic or acute health status, sensitivity to controls, previous experience of administrative interactions and level of trust in institutions.
RARE CHRONIC CONDITIONS
atypical care pathway due to medical necessity, high exposure to being flagged, limited room for explanation
TRANSITIONAL LIFE SITUATIONS
recent professional or family change, significant transactional variations read as anomalies
SENIOR INSURED MEMBERS
high and varied healthcare use, multiple providers, strong sensitivity to the feeling of being suspected
SIMPLE CARE PATHWAYS
median and regular healthcare use, low probability of being flagged, little experience of controls
INSURED MEMBERS WITH HIGH INSTITUTIONAL TRUST
control interpreted as a normal procedure as long as it is explained and time-bounded
PREVIOUS EXPERIENCE OF AN ERROR
past administrative dispute, immediate interpretation of the control as repeated injustice
ADMINISTRATIVE VULNERABILITY
difficulty gathering supporting documents, high cost of pursuing an appeal, risk of abandonment
INSURED MEMBERS NOT CONTROLLED BUT INFORMED
never flagged, aware that the system exists, yet still forming a judgment of the insurer
These configurations reveal part of the population's heterogeneity. The simulation operates on synthetic individuals, not a handful of persona types.
REACTIONS
ERRORS ARE NOT RANDOMLY DISTRIBUTED
False positives are concentrated in three types: rare chronic conditions with atypical pathways, people in transitional situations, and senior insured members with high healthcare use. These types represent 8 % of the insured population and 62 % of the false positives generated by the model. The overall error rate says nothing about this concentration.
TRANSPARENCY COSTS LESS THAN OPACITY
An unannounced control produces 34 % satisfaction when resolved favorably and 12 % when the claim is rejected. An explained control — nature of the system, reason for the flag, right of appeal — produces 72 % and 51 %. Opacity is read as illegitimate suspicion; explanation, as respect owed to the insured member.
CONTESTABILITY MATTERS MORE THAN PRECISION
At identical precision, the same model produces 24 % average satisfaction without a human appeal procedure and 68 % with a formal appeal available within fifteen days: a 44 points difference obtained without changing the model. The algorithm cannot be the final decision-making authority.
DIFFERENTIATION IS NOT A COMPROMISE
Differentiated thresholds by type — more permissive for the three sensitive types, stricter elsewhere — reduce false positives by 62 % without degrading fraud capture, with 34 % lower human-processing costs and 78 % lower projected reputational risk. Ethical optimization and operational optimization point in the same direction here.
Imagine All The People does not ask a model whether predictive AI is acceptable. Eleven decision architectures are tested against a heterogeneous synthetic population to observe who is controlled, who manages to justify their case, who abandons the process, and what each person then concludes about their insurer.
CONDITION FOR DEPLOYMENT
WHAT THE SYSTEM HAD TO DEMONSTRATE
Full traceability of the configurations tested, validation by the insurer's ethics committee before deployment, involvement of legal and compliance teams, annual external audit by an independent firm, and public reporting on bias and false-positive indicators.
WHY THIS CHANGES THE DESIGN
Explainability is not a convenience feature here. It conditions authorization to deploy. A higher-performing but non-auditable configuration was not an available option.
RESULT
Comparative reading based on simulated reactions across five architectures representative of the eleven configurations tested. The false-positive column is read in the opposite direction from the others: a high level signals a large volume of unjustified controls. A cautious configuration may detect less; an aggressive configuration controls more legitimate insured members.
THE MOST ROBUST
Differentiated thresholds by type transparent information human appeal within 15 days prior ethical governance continuous model reassessment
THE MOST FRAGILE
Single threshold automated decision no information provided to the insured member
The ethics committee selected the leading configuration, reinforced with a permanent committee including representatives of user associations, a published charter and an annual external audit. At 18 months after full deployment, annualized fraud savings reach €62 million versus €54 million projected, the false-positive rate among sensitive types is 4,2 %, insured-member satisfaction after a control is 71 %, and human appeal reverses the algorithmic decision in 34 % of cases where it is used. No significant media or advocacy challenge emerged over the period.
TAKEAWAY
The model was good. That is what made the decision difficult. Its overall precision said nothing about how its errors were distributed: 8 % of the population accounted for 62 % of false positives, and that population was precisely the one whose care pathways are atypical by necessity.
The decisive adjustment therefore concerned not the algorithm but the architecture around it: differentiated thresholds, reasoned explanation, human appeal, governance established before deployment. With model precision unchanged, these parameters shifted insured-member satisfaction by 44 points.
POSSIBLE FUTURES
A
AUTOMATISER
The predictive score directly triggers the control, with a single threshold for the whole population
B
ASSISTER
The score becomes one signal among others, validated by a human case manager
C
GOUVERNER
Differentiated thresholds, transparent information, formal appeal, external audit, quarterly reassessment
WHAT THE SIMULATION DOES NOT DO
This case does not conclude that these systems should be deployed, nor that they should be rejected. Predictive detection delivers a real operational gain against costly and growing fraud. The model itself is not the issue.
What is simulated are the consequences of how its prediction is transformed into a decision: who is affected, with what errors, what asymmetries, and what reputational and regulatory consequences. The simulation makes these architectures comparable. It does not rank them on behalf of the institution.
METHOD
Operational and reputational trajectories were projected over 24 months for each configuration. This case adds a capability to the library: testing not only the reaction to a decision, but how an automated system itself produces decisions about populations.
This case is based on a simulation conducted for a European health insurer. The company is not named, no personally identifiable data entered the system, and detailed results remain the client's property. The comparisons published here are qualitative and constitute neither an algorithmic audit nor an assessment of regulatory compliance.
Your next decision
Describe your need. We can point you to the right level of support.
State your decision. See the future it produces.
