Ask an insurance executive whether AI catches fraud better than people do, and the honest answer is usually a question back: better at what?
That's not a dodge. AI and human investigators are good at fundamentally different things, and the difference matters — whether you're a policyholder whose claim just got flagged, an agent explaining a delay to a frustrated client, or a carrier deciding what to buy.
The scale of the problem
The most-cited estimate comes from the Coalition Against Insurance Fraud, whose 2022 study put total U.S. losses at roughly $308.6 billion a year across all lines: life insurance around $74.7 billion, property and casualty around $45 billion, workers' compensation around $34 billion. Four years on, that's still the operative number — the Coalition leads with it, and both the NAIC and the NICB were citing it in 2026. Comprehensive fraud estimates are rare enough that the 2022 study was itself the first update to a figure set in 1995. The Coalition also acknowledges the true cost can't be measured precisely, since undetected fraud is by definition uncounted.
For healthcare specifically, the National Health Care Anti-Fraud Association doesn't publish a dollar figure at all. It estimates losses in the tens of billions each year, using a conservative benchmark of roughly 3% of total health spending, while noting that some government and law-enforcement agencies put the share closer to 10%. Against the $5.3 trillion the U.S. spent on health care in 2024, that's a range from about $160 billion to more than $500 billion — a spread wide enough to be its own commentary on how much of this is estimation.
Divided across the population, the Coalition's total works out to roughly $900 per person per year, paid quietly through premiums. (The FBI's older estimate, covering non-health lines only, lands in the same neighborhood: $400 to $700 per household annually.) However you slice it, honest policyholders foot the bill.
One distinction is worth carrying through this entire article: suspicion is not proof. A claim can be unusual, poorly documented, or statistically strange — and completely legitimate. Hold onto that..
Not all fraud looks the same
The word "fraud" covers two very different things, and they call for very different detection strategies. Hard fraud is invented from nothing: a staged collision, an arson-for-profit fire, a fabricated injury, a death claim on someone who isn't dead. It's premeditated, often organized, and frequently involves rings — a clinic, an attorney, a body shop, and a rotating cast of "witnesses" all running the same playbook.
Soft fraud is opportunistic: an honest claim, inflated. The laptop that was three years old became "practically new." The pre-existing dent joins the accident. Most people who commit soft fraud don't think of themselves as criminals. They think they're squaring up with a company that's been taking their premiums for years. Soft fraud is far more common. Hard fraud is far more expensive per case. That asymmetry shapes everything: you need one system that can spot a $2,000 exaggeration buried in thousands of ordinary claims, and another that can see a $2 million ring hiding across files that look unrelated.
Why the old approach hit a wall
For decades, detection ran on two things: rules and instinct. Rules engines flag anything matching a preset pattern — a claim filed within 30 days of policy inception, a loss just under an investigation threshold, a total loss on a vehicle with no lienholder. Then an SIU examiner picks up the file and applies experience. Rules are static, and fraudsters read the room. Once people learn that claims over $10,000 get scrutiny, the claims come in at $9,800. A rule that worked last year is a roadmap this year. Human capacity is finite. SIU investigators carry dozens of active cases, each needing evidence gathering, interviews, surveillance coordination, report writing. Claim volume vastly outstrips available hours. Triage isn't a strategy choice — it's arithmetic.
What AI actually does
Machine learning is strong at exactly the thing humans can't do: look at everything, at once. Instead of applying fixed rules, models learn what a normal claim looks like across millions of examples, then score new claims against that baseline — including patterns nobody thought to write a rule for. Four things they do well:
Anomaly detection. Unsupervised models find claims that don't resemble anything else in the data, without anyone specifying in advance what "suspicious" means. This is how genuinely novel schemes surface.
Reading unstructured text. Most of the signal in a claim file isn't in the structured fields. It's in adjuster notes, medical narratives, and the claimant's own description. Language models can read thousands of these and flag inconsistencies a human would only catch by reading every file closely, which nobody has time to do.
Image forensics. Damage photos carry metadata: when they were taken, where, on what device. A model can check whether the timestamp matches the reported loss date, whether the same photo appeared in a prior claim, and whether the pixels show signs of editing.
Network analysis. This is where AI most clearly beats a human. A single staged accident looks like an accident. But when the same clinic, the same attorney, the same body shop, and the same three witnesses keep reappearing across claims filed by strangers in different months, a graph model sees the ring. An adjuster looking at one file cannot — not because they aren't sharp, but because they're looking at one file.
The timing advantage is real. In a study completed by CLARA Analytics in late 2024, machine-learning models analyzing 2,867 property and casualty claims filed between 2020 and 2024 identified claims with high SIU-referral potential as early as two weeks after first notice of loss — well ahead of typical investigative timelines. The models' predictions tracked closely with the referrals human professionals eventually made. They just got there sooner, which matters when costs compound with time. Adoption is near-universal: roughly 96% of insurers now use some form of anti-fraud technology.
A caution about the numbers you'll see elsewhere. Much of the eye-popping data circulating online — "10x investigator throughput," "90% cost reduction," "14 days cut to 2 hours" — comes from vendors selling the software. Treat it as you'd treat any sales claim. The independent evidence supports something more modest and more useful: AI is very good at surfacing suspicion, earlier and at a greater scale than people can manage alone.
The thing AI is worst at
False positives. A model can tell you a claim is statistically unusual. It cannot tell you why, and it cannot tell you whether "unusual" means "fraudulent." This isn't abstract for our readers. Consider who gets flagged as anomalous: someone with no U.S. credit history. Someone whose address houses three generations of one family. Someone who files a claim in imperfect English. Someone whose financial footprint looks nothing like a 40-year-old suburban homeowner's — because they're a 22-year-old international student. None of that is fraud. All of it is unusual relative to training data built mostly on people who don't look like that.
An algorithm trained on historical claims learns historical patterns, including whatever bias was baked into who got scrutinized before. If certain ZIP codes were over-investigated in the past, a model will cheerfully learn to over-investigate them in the future, and it will look perfectly objective doing it. This is the proxy problem: a system can produce discriminatory outcomes without ever touching a protected characteristic, simply by leaning on variables that correlate with one. Every false positive is a real person, waiting on money they're owed, treated as a suspect while they wait.
So: better, or not?
AI is better at surfacing suspicion. More claims, more signals, faster, without fatigue, and with a view of the connections between files that no individual can hold in their head.
Humans are better at establishing truth. Context, intent, plausibility. Is this injury medically consistent with this accident? Does this person's hesitation mean deception — or that they're terrified of losing their house? Is this a fraud pattern, or a legitimately unusual life?
A model that flags a claim has produced a hypothesis, not a verdict. Someone still has to test it. That's also, not coincidentally, what regulators expect. As we covered last time, the direction of U.S. insurance AI regulation is consistent across states: insurers must explain AI-influenced decisions, test models for unfair outcomes, and keep meaningful human review in high-stakes calls. An insurer generally cannot deny a claim on an algorithm's say-so.
The new problem: AI is helping the fraudsters too
The same technology has made fabricating evidence dramatically cheaper. Carriers are now seeing AI-generated damage photos, synthetic repair invoices, fabricated medical records, and doctored documents that would have taken real skill to produce a few years ago. Zurich and other insurers have publicly flagged the trend. The old verification instinct — does this document look right? — degrades as the forgeries improve.
The response is largely more AI: forensic models examining image metadata, detecting manipulation at the pixel level, cross-checking documents against known templates. It's an arms race, and an early one. Anyone claiming to have solved it is selling something.
If your claim gets flagged
An investigation is a question, not an accusation. What helps:
Provide accurate information, and stay consistent across every conversation.
Keep receipts and original, unedited photos. Never alter a document, even to make it clearer.
Respond promptly. Silence gets read as evasion.
Ask in writing what's needed, so there's a record.
If a claim is denied, request a written explanation and use the appeal process. Your state insurance department exists for this.
Where we land
AI catches fraud humans would miss — faster, at a scale no SIU could staff for, and it's genuinely superior at seeing the connections that turn scattered claims into a visible ring. Any insurer not using it is leaving money on the table, and that money ultimately comes out of honest policyholders' premiums.
But AI does not determine that someone committed fraud. It determines that something looks odd. Those are very different statements, and collapsing them is how you deny a legitimate claim to someone who did nothing wrong.
The goal was never to flag the most claims. It's to flag the right ones, and to be fast and fair to everyone else.
Sources: Coalition Against Insurance Fraud, The Impact of Insurance Fraud on the U.S. Economy (2022) and fraud technology surveys; NHCAA, "The Challenge of Health Care Fraud"; FBI insurance fraud estimates; CLARA Analytics P&C claims study (November 2024); NAIC Model Bulletin on the Use of AI Systems by Insurers (December 2023).
