A number is moving around the workers' compensation industry right now, and it is a big one. Traditional rule-based fraud detection catches about 60 to 80 percent of what it is looking for. AI-powered systems are hitting 85 to 95 percent, and hybrid models are pushing past that. The commonly cited lift is up to 60 percent improvement over legacy methods (True Insurtech Solutions).
That is a compelling headline. It also skips over what actually matters, which is what happens after the flag.
What a 60 percent lift really represents
Deloitte's own framing is worth reading carefully. AI is detecting soft fraud at rates between 20 and 40 percent, and hard fraud at rates between 40 and 80 percent. The projected savings on fraudulent-claim costs are in the 20 to 40 percent range. That is real money, and it is real prevention. But it is also entirely dependent on what the human on the other end of the system does with the signal.
The Coalition Against Insurance Fraud has said this out loud: up to 16 percent of workers' comp claims have some element of fraud, adding roughly $9 billion in annual losses. Business Insurance covered this and quoted experts who all landed on the same conclusion, which is that AI is a signal, not a verdict (Business Insurance).
The failure mode most vendors do not talk about
Every fraud model produces false positives. That is not a flaw of the technology, it is a property of statistics. When a model flags a legitimate injured worker as suspicious and a human treats that flag as fact, the consequences fall on someone who is already hurt.
That is the piece the industry has to get right. A 60 percent lift in detection is only a win if the additional flags are triaged with the same discipline as the ones a claims professional would have caught manually. Rushed disposition of AI flags is how good tools become bad outcomes.
How we think about it inside Atlas workflows
Atlas uses pattern review inside our Utilization Review and Nurse Case Management workflows. Three principles drive how we deploy it:
- Every flag surfaces to a clinician before it touches a decision. The model prioritizes what humans look at. It does not decide.
- Transparent reasoning, not black-box scoring. Our clinicians see why a claim was flagged. That is table stakes for defensible URAC-aligned decision-making.
- Every AI-influenced decision is documented as such. An injured worker, an employer, and a carrier all have the right to know when a tool was in the loop.
The bar is not detection. The bar is integrity.
A 60 percent lift in fraud detection sounds like a technology story. It is really a workflow story. Any managed care partner running AI in the claims pipeline needs to be able to answer three questions on demand: Which flags did the model raise. Who reviewed them. What decision followed, and on what evidence.
If those three answers are not clean, the technology is not helping. It is just shifting risk. Atlas builds every automation with those three answers in mind.
AI-assisted claims oversight, human-led decisions.
Atlas uses machine-assisted pattern review inside our Utilization Review and Nurse Case Management workflows. Every flag surfaces to a licensed clinician before it ever touches a claim decision.
See our services