Claims Leakage: What It Is, What It Costs, and the Fraud Gap That Drives It

Claim leakage forms everywhere money changes hands in a claim: a deductible nobody applied, an invoice paid twice, depreciation skipped on a roof estimate, a subrogation opportunity missed at intake, a reserve that stair-stepped upward until the settlement anchored to it. Fraud is one cause of leakage among many - but at most carriers it is the largest controllable one, which is why this page covers the whole concept and then goes deep on the fraud gap.

Within that fraud slice, claims fraud leakage is the dollars a carrier pays on fraudulent claims that should have been investigated and denied or reduced. It is the portion of insurance fraud the detection-and-investigation stack failed to catch - a direct hit to loss ratio and combined ratio.

Leakage is not just about hard fraud (fully fabricated claims). The largest contributor in most carriers is soft fraud - inflated estimates, exaggerated injuries, padded amounts on legitimate underlying events that an investigation would have reduced but never received. Soft fraud is harder to detect, easier to commit, and more frequent than hard fraud by an order of magnitude.

For the operator-level reduction playbook, see insurance claims leakage: how uninvestigated claims drain profitability. For where in the claims lifecycle leakage actually forms - from FNOL through reserving to settlement - see the claims process end-to-end.

The industry splits leakage along two axes. Hard vs soft: hard leakage is a payment that plainly breached the policy or process - an uncovered loss paid, a deductible never applied, the same invoice paid twice. Soft leakage is judgment drift - a generous estimate, a reserve that crept upward, a settlement rushed to clear a desk. Overpayment vs underpayment: overpaying wastes indemnity dollars; underpaying trades indemnity for regulatory exposure and litigation. Every leak type maps onto both axes and onto a specific stage of the claim lifecycle.

No single tool or team addresses all ten rows, and most leakage programs quietly pick two or three. The rest of this page quantifies the biggest block - the fraud gap - because it is the one a carrier can close fastest without replacing its core claims platform.

Leakage compounds from five distinct sources. Most leakage reduction programs target only one or two; the leverage is in addressing all five.

1. Under-investigation: detection flags claims, but investigator capacity processes only 25-30% of them. The remaining 70-75% pay out without investigation.
2. Detection miss: fraud the detection layer never flagged. Soft fraud and emerging patterns escape rules-based and statistical detection.
3. Investigation depth: investigations that close prematurely due to time pressure miss fraud that a thorough investigation would have caught, and a denial resting on a thin file review fails the same duty to investigate as one resting on an AI output nobody can explain (see AI fraud investigation and court admissibility).
4. Recovery gaps: confirmed fraud that gets denied at the claim level but never gets pursued for recovery (subrogation, restitution, criminal referral). Recovery leaks outside fraud too: roughly one in four US property-liability insurers records no salvage or subrogation recovery in a given year (Bisco and Fier, Journal of Insurance Regulation, 2023), as covered in why missed subrogation is lost at intake, not at closure.
5. Provider and network leakage: organized fraud that requires cross-claim analysis to detect; a single claim looks legitimate, the pattern across many claims does not.

On the medical/provider side specifically, see medical record fraud in insurance claims.

A typical mid-size P&C carrier processes 2,000 flagged claims per month with 10 SIU investigators. The math:

• Monthly flagged volume: 2,000 claims
• Manual investigation capacity: ~100 cases/month (10 investigators × ~10 cases each)
• Monthly capacity gap: ~1,900 uninvestigated flagged claims
• Annual uninvestigated: ~22,800 claims
• Fraud density among uninvestigated: typically 15-25% (these claims were flagged for a reason)
• Average claim value: ~$8,000
• Average inflation per fraudulent claim: 20-40% of value
• Annual leakage estimate: 22,800 × 20% × $8,000 × 30% ≈ $11M per year

Most carriers do not model leakage explicitly. Once modelled, it reframes the business case for investigation capacity investments - the cost of not investigating dwarfs the cost of investigation infrastructure. To measure it rather than model it, the claims leakage audit checklist works stage by stage, sizes the sample, and explains why NAIC sampling guidance bars extrapolating a sampled leakage rate beyond the files actually drawn.

For a CFO-ready memo that walks through this leakage math and frames the investment case for leadership, see the CFO ROI memo: the business case for AI claims investigation. Once a program is live, see measuring the ROI of fraud investigation AI for the metrics that prove leakage recovery after deployment. When a vendor leads with efficiency, ask for the outcome column: across the five autonomy levels of Shift's ARISE framework the efficiency figure multiplies eightfold while the indemnity impact moves three percentage points, and 43% of US insurers do not capture the Coalition's Impact Ratio at all - see the two columns of the ARISE ladder.

Four mechanics decide whether the audit produces a number anyone can act on. Stratify the sample by line, severity band and handling office, or high-volume cells drown out the expensive ones. Use reviewers independent of the team that handled the files - self-audits reliably find less. Code causes, not just dollars, because a leakage figure without a cause distribution cannot be fixed. And resist extrapolation: NAIC market conduct sampling guidance treats a sampled error rate as evidence about the files drawn, not a book-wide multiplier. The working version of all four, with sample sizing by book, is in the claims leakage audit checklist.

Which benchmark should you use while the audit runs? Published leakage ranges disagree by an order of magnitude, largely because almost nobody names a source. Here is where each circulating number actually comes from:

The honest reading: model with EY's 7-14% of total claims spend, treat The Lab Consulting's 20-30% claim as a ceiling argument rather than a planning number, and replace both with your own closed-file audit result within two quarters. A carrier quoting an industry percentage in a board deck without an audit behind it is reporting someone else's book.

The dominant source of leakage is the 75% of flagged claims that pay out without investigation. This is not a process failure - it is the inevitable consequence of detection systems generating flags faster than manual investigation can process them. For the full operational analysis, see why 75% of flagged insurance claims are never investigated.

Three strategies historically used to manage the gap, and why each has limits:

• Triage: investigate the highest-value cases, close the rest. Works at the top of the queue but leaves substantial leakage in the mid- and tail.
• Outsourcing: contract additional investigation capacity. Adds cost without changing the structural ratio between flag volume and investigation throughput, because panel firms bill for investigator time drawn from a national pool of 39,500 private detectives and investigators (see outsourced investigation services vs autonomous AI).
• Detection precision: tune rules to flag fewer claims. Reduces queue volume at the cost of recall - the carrier misses more genuine fraud.

Five interventions, deployed sequentially, address each leakage source:

1. Documented triage criteria - reduces investigator-to-investigator variance, lifts confirmed fraud rate by 20-40%.
2. Documentation automation - reclaims 1-2 cases per investigator per month from report writing.
3. Database query automation - reclaims 2-3 cases per investigator per month from manual NICB, ISO ClaimSearch, LexisNexis pulls.
4. Autonomous AI investigation - the largest single lever; investigator throughput rises from ~10 to 50-100+ cases per month, closing coverage from 25% to 85-100% without adding headcount.
5. Detection precision tuning - reduces false positive volume so the investigated portion of the queue has higher fraud density.

For the procurement framework when evaluating investigation platforms, see evaluating AI fraud investigation vendors. To sharpen the first intervention - triage - so the right claims reach investigators before they leak, see the claims triage automation guide and, for the decision rules that route each flagged claim to the right depth of review, the claims triage decision framework.

Three platform categories address claims fraud leakage from different angles. Choosing the right one - or combining them - depends on which leakage source dominates at your carrier.

For the hidden costs of bundled claims-suite AI modules, see hidden integration costs of adding AI modules to legacy claims management suites. For a side-by-side evaluation of the major platforms and their fraud capabilities, see claims management systems comparison and top claims management systems for P&C carriers 2026. Inside the system of record the SIU workflow is real, but it records the investigation rather than running it: Guidewire ClaimCenter ships referral routing, accept-or-reject dispositions, supervisor escalation and an outcome record whose fields SIU staff enter after the investigation is finished (see Guidewire ClaimCenter fraud). Across the whole claim lifecycle, vendor density runs inverse to where leakage concentrates - seven vendors serve intake against three at subrogation - as mapped stage by stage in the best AI claims processing software in 2026.

Most leakage writing stops at overpayment. The symmetric risk costs more per dollar. Underpaying or wrongly denying claims is leakage too - it leaks through regulatory penalties, bad-faith verdicts, department-of-insurance complaint volume, and the quiet cost of non-renewal by customers who were paid short. The NAIC Unfair Claims Settlement Practices Act, Model #900 makes failing to conduct a reasonable investigation an unfair claims practice in its own right, so a leakage program that closes files faster by investigating less manufactures the exposure it was meant to remove.

That is the constraint on automating leakage away. Under the NAIC model bulletin on insurer AI use, the carrier remains accountable for every adverse decision its systems touch. Leakage reduction that survives a market conduct exam is investigation-grade: each reduced or denied claim backed by a documented, human-reviewable evidence file, not a score (see human-in-the-loop AI investigation). Done that way, deeper investigation shrinks both halves of leakage at once: it catches the inflated claim and documents why the legitimate one deserved full payment.

Leakage is not evenly distributed, and the dominant leak type changes by line - which is why a single corporate leakage percentage hides more than it reveals:

Run the closed-file audit by line before by cause: the cause codes that dominate an auto book (estimate inflation, injury buildup) barely appear in a property book (depreciation, pricing data), so a blended corporate number points every fix at the wrong queue.

Leakage is a carrier-level symptom of an industry-level number. The Coalition Against Insurance Fraud estimates that insurance fraud steals at least $308.6 billion a year from American consumers - the first refreshed estimate in 27 years, up from the long-cited $80 billion. The property and casualty share alone is roughly $45 billion, with workers compensation adding another $34 billion, per the American Academy of Actuaries September 2024 issue brief. When your own uninvestigated queue is leaking $11M a year, it is a rounding error inside a $308B system - which is exactly why the problem persists.

The mechanism that generates most of that number is soft fraud, not hard fraud. The American Academy of Actuaries notes that the NAIC identifies two categories - hard fraud (deliberate fabrication) and soft fraud (exaggerating an otherwise legitimate claim) - and that soft fraud "is more common." Soft fraud is precisely the leakage that a thorough investigation reduces but a rushed one waves through, which is why coverage - not just detection - is the lever. For the full data set on fraud by line of business, see insurance fraud statistics 2026: the $308 billion problem.

How much of a carrier's spend actually leaks? EY puts indemnity and leakage at approximately 7% to 14% of carriers' total claims spend - a range that dwarfs most SIU budgets and reframes investigation capacity as a loss-ratio decision, not a cost center. The largest single blocks sit in the two lines where soft fraud is easiest to commit: see workers compensation fraud investigation and auto insurance fraud investigation for the line-level playbooks. Severity itself is now a leakage driver on the settlement side: plaintiff firms assemble AI-built bodily-injury demand packages in hours while carriers still examine them by hand over weeks, an asymmetry covered in AI demand packages and claim severity.