By Chris Ciappa
Founder & Chief Coherence Architect
Samirac Partners
When AI Accuses the Wrong Doctor
Now bear with me a moment and let me tell y’all a healthcare story. I may just be a farm boy, but I know this much: the minute you let AI start flagging people or businesses for potential fraud without controlling drift—especially when payments and money flow are involved—you’ve just opened a real can of worms, and by that I mean some very serious liability.
Imagine a national health insurer deploys a new AI fraud detection system.
Within weeks it flags something alarming.
A cluster of physical therapy clinics suddenly begins billing a rehabilitation code three times more often than the national average.
The model flags the providers.
Risk scores spike.
Payments are frozen.
Investigations begin.
From a machine-learning perspective, the system worked perfectly.
The anomaly was real.
The statistical pattern was undeniable.
There was just one problem.
The clinics weren’t committing fraud.
What the AI Saw
The model was designed to detect unusual billing behavior.
And that’s exactly what it did.
A group of providers suddenly billing the same code at abnormal frequency is exactly the type of signal fraud systems are built to catch.
From the model’s perspective, this looked suspicious.
From the payer’s perspective, the signal triggered investigations.
But the model was missing something critical.
What Actually Happened
Six months earlier, a new orthopedic surgery center opened in the region.
It specialized in complex knee reconstruction procedures.
Those surgeries required significantly longer rehabilitation periods.
Patients needed more physical therapy.
Therapists used the rehabilitation billing code more frequently.
Billing patterns changed.
The AI detected the change.
But the anomaly was not fraud.
It was medicine.
Where the System Failed
The model itself did not fail.
The anomaly detection worked exactly as designed.
The failure occurred one layer above the model.
Without stable identity anchoring and clearly defined identity invariants built into the system architecture, the model’s output was treated as a fraud signal rather than what it actually was — a statistical observation.
Once the signal was interpreted as fraud risk, the system escalated automatically.
Payments were delayed.
Investigations were launched.
Providers filed complaints.
What began as statistical analysis became an operational decision.
The Dangerous Assumption in Modern AI Systems
Many AI systems today operate under a simple assumption:
If a pattern is unusual, it must be suspicious.
But in complex systems like healthcare, unusual behavior often reflects:
New clinical practices
Regional treatment differences
Provider network changes
Policy shifts
Population health changes
Detection systems can identify patterns.
But patterns do not explain themselves.
And when systems treat statistical anomalies as operational signals without architectural context, they begin making confident mistakes.
Detection Is Not Decision Architecture
Fraud detection systems are often designed around analytics layers:
anomaly detection
network clustering
risk scoring
predictive modeling
These tools are powerful.
But they operate at the pattern detection layer.
Fraud enforcement decisions exist at a completely different layer — the decision architecture layer.
Without structure between those layers, statistical suspicion becomes operational authority.
The Five Layers That Prevent Decision Drift
Stable decision systems across domains tend to converge around similar structural layers.
In fraud detection, those layers include:
Identity Anchors
Who exactly is being evaluated.
Reference Frames
What contextual rules interpret the data.
Coherence Boundaries
Which inputs are admissible evidence.
Drift Detection
How behavioral changes are interpreted.
External Validation
Who confirms the signal before enforcement.
When these layers are missing, AI detection systems become extremely good at identifying anomalies — and dangerously bad at interpreting them.
The Architecture Problem
This is precisely the class of problem addressed by structural architectures such as the Drift Stack™.
Rather than treating model outputs as decisions, the Drift Stack separates system layers:
Detection generates signals.
Architecture establishes context.
Validation confirms meaning.
Only then does action occur.
In other words:
AI detects.
Architecture decides.
The Future of Fraud Detection
Artificial intelligence will dramatically improve fraud detection.
But smarter models alone will not solve the problem.
As detection capabilities increase, governance architecture becomes more important, not less.
The organizations that succeed will not simply deploy better models.
They will design systems where intelligence operates inside stable identity anchoring, structural context, and governed execution boundaries.
Because the most dangerous failure in fraud detection is not missing fraud.
It is confidently accusing the wrong person.
And that is the deeper architectural problem.
Detection systems can identify anomalies internally. But when the system itself drifts, only an external correction layer can determine whether the signal represents fraud — or simply a change in reality.
By Chris Ciappa
Founder & Chief Coherence Architect
Samirac Partners
The Drift Stack™
—