---
title: "Inflated insurance claim investigation: the loss is real, the number is not"
description: "The biggest category of insurance fraud is a real claim with a wrong number on it. A fraud score cannot say which line is unsupported, and cutting a line without evidence is a partial denial."
date: "2026-11-15"
lastModified: "2026-11-16"
author: "Pankaj Dhariwal"
tags: ["Guides"]
canonical: "https://gethesperai.com/blog/inflated-claim-investigation/"
---

# Inflated insurance claim investigation: the loss is real, the number is not

> **TL;DR** An inflated insurance claim is a real loss submitted at the wrong amount, and FRISS attributes 44% of fraud to exaggeration of legitimate claims. A fraud score returns a probability for the whole file. A partial denial needs a finding on each line with the source behind it, which is a different instrument entirely.
>
> - 44% of fraud is exaggeration of a real claim, per FRISS
> - 89% of carriers run red flags; 57% cite false positives
> - Washington: a database alone is not a reasonable investigation

- **44%** - of fraud FRISS attributes to exaggeration of a legitimate claim (FRISS, The State of Insurance Fraud 2026, Chapter 1, accessed 15 November 2026. Methodology undisclosed: the page states no sample size, population definition, geography or fielding dates. The same page puts fake injury claims at 29%, organized crime at 13% and rate evasion at 9%)
- **15%** - of BI and PIP dollars paid went to claims with the appearance of buildup (Insurance Research Council, Fraud and Buildup in Auto Injury Insurance Claims, published February 2015 on 2012 data. A separate IRC figure puts the excess added by fraud and buildup together at between 13% and 17% of total payments. The two are different quantities and do not combine)
- **89%** - of carriers run automated red flags and business rules, the most widely used anti-fraud technology (Coalition Against Insurance Fraud, 2024 State of Insurance Fraud Technology Study, sixth edition, produced with Shift Technology. 35 insurer-member responses, fielded 3 October to 14 November 2024)
- **57%** - name high rates of false negatives and false positives among their biggest deployment challenges (Coalition Against Insurance Fraud, 2024 State of Insurance Fraud Technology Study. Limited IT resources at 86% and data integration and poor data quality at 60% rank above it on the same barrier list)

The largest category of insurance fraud is not a fake claim. It is a real one with a wrong number on it. That makes an inflated insurance claim investigation a different job from a fraud investigation, and it needs a different instrument.

The dispute is over a dollar figure on a line, not the existence of the claim. A fraud score returns a verdict on the whole claim, so the file never reaches an SIU and lands on an adjuster. The biggest category of fraud is the one nobody is equipped to resolve: too real to deny, too small to refer, too slow to verify by hand.

## The category the fraud stack was not built for

**Which kind of insurance fraud is most common?**

An inflated insurance claim is a real loss submitted at the wrong amount, and in FRISS's 2026 fraud report it is the largest single category of fraud. FRISS attributes 44% of fraud to opportunistic exaggeration of legitimate claims, against 29% for fake injury claims and 13% for organized crime.

FRISS's own label for the 44% category is "Opportunistic fraud (exaggeration)", printed on [chapter one of The State of Insurance Fraud 2026](https://www.friss.com/downloads/the-state-of-insurance-fraud-2026-chapter-1), methodology undisclosed: the page states no sample size, no population definition, no geography and no fielding dates. Treat the split as a vendor's read of its own respondents.

The definitional boundary - what separates [soft fraud and hard fraud](/glossary/soft-fraud-vs-hard-fraud/) - is settled elsewhere. What matters operationally is that the two sides of it need different instruments.

A different method reached the same ordering. The Insurance Research Council, in a study published in February 2015 using 2012 data, defined buildup as the inflation of otherwise legitimate claims and found that [the most common type of abuse was claim buildup](https://insurance-research.org/news/insurance-research-council-finds-fraud-and-buildup-add-77-billion-excess-payments-auto-injury).

## Why a fraud score cannot answer the question

**Why does a fraud score not resolve an inflated claim?**

A fraud score answers whether a claim is suspicious. An inflated claim needs a different answer: which line is unsupported, and by how much. The loss happened, the policy responds, and the claimant is owed something, so there is no clean yes or no for a binary instrument to return.

The instrument emits a binary. In the Coalition Against Insurance Fraud's [2024 State of Insurance Fraud Technology Study](https://insurancefraud.org/wp-content/uploads/2024-State-of-Technology-Study.pdf), produced with Shift Technology from 35 insurer-member responses, automated red flags and business rules remain the most widely utilized at 89%. A rule fires or it does not. A referral is accepted or it is not. Nothing in that chain has the shape of the answer an inflated claim needs.

**Figure 1 (The fraud-flag path) - Three binary gates, and the overstated line is still unexamined.** A claim whose loss is real and whose amount is wrong passes through the detection stack without ever being asked which line is unsupported.

The claim as filed: Real loss, policy responds; One or more lines overstated

1. **Rules and scoring** - Automated red flags and business rules: 89% of carriers
2. **Referral threshold** - No single padded line clears it; the aggregate never scores
3. **SIU workup** - Manual and scarce: about 25% of flagged claims get a full one

Escalation from stage 3: Neither gate returns a finding on the line that is overstated

**Where it actually lands - Back on the adjuster's desk**: Paid as submitted, Or cut, basis undocumented

**Takeaway:** A binary instrument cannot return a per-line finding, so the line nobody tested gets paid or gets cut on a hunch.

Sources: Coalition Against Insurance Fraud, 2024 State of Insurance Fraud Technology Study, 35 insurer-member responses fielded 3 Oct - 14 Nov 2024, for the 89% on automated red flags and business rules. The share of flagged claims that get a full workup is a Hesper internal benchmark. NICB, revised March 2015: indicators are not actual evidence and only indicate the need for further investigation.

It misfires in both directions, and carriers say so: in the [same Coalition study](https://insurancefraud.org/wp-content/uploads/2024-State-of-Technology-Study.pdf), 57% named high rates of false negatives and false positives among their biggest deployment challenges. The bureau that receives the referrals is blunter still.

> Indicators of possible fraud are not actual evidence, they only 'indicate' the need for further investigation. ... Some claims, although questionable, may be paid due to a lack of conclusive evidence of fraud.
>
> - National Insurance Crime Bureau, Indicators of Contractor / Public Adjuster Property Repair Fraud, revised March 2015

An indicator tells an adjuster to look. It does not say which line is wrong or by how much, and without that a questionable claim gets paid. Prevention, detection and investigation are three separate layers, and this category falls between the last two.

## What an inflated insurance claim investigation tests, line by line

**How is claim inflation actually executed?**

Claim inflation is executed at the line-item level, not at the claim level. In property it is prior damage folded into a new loss, whole-dwelling estimates written for isolated damage, and deductibles recovered by raising a price. In auto injury it is treatment volume.

The NICB publishes the concrete signals. Read them as line-item questions, not a fraud checklist. Its [indicator list for contractor and public adjuster property repair fraud](https://www.nicb.org/sites/files/2020-02/Indicators-of-Contractor-Public-Adjuster-Property-Repair-Fraud-updated-2015.pdf), revised in March 2015, names:

- Invoices which contain inflated or fictitious statements
- An estimate increased to cover the cost of a consumer's deductible
- An estimate that is very general and lump sum
- Damage that includes the entire property rather than one area of it
- Billing for incomplete repairs

On the auto side the NICB's own referral data shows what the disputed object is. In a [ForeCAST report dated 9 September 2015](https://www.nicb.org/sites/files/2019-06/Vehicle%20Repair%20Fraud%20(PDF%20File).pdf) on vehicle repair questionable claims, the leading referral reason in the first half of 2015 was Prior Loss/Damage at 66%, ahead of Inflated Repairs at 8%. A single questionable claim can list up to seven referral reasons, so those shares do not sum to 100%.

Prior Loss/Damage is not a fraud question. It is a line question: is this dent attributable to this loss? No claim-level score answers it, because the answer is a property of one line set against one external record - a prior-loss photograph, a prior estimate, an index-bureau hit.

**Figure 2 (Per-line findings) - A partial denial needs a finding on each line, with its source.** Each line gets its own answer and its own source, drawn from the indicator language the NICB publishes for property repair fraud.

**One estimate, four lines tested**

| Line and finding | Tested against |
| --- | --- |
| Roof slope B: creasing predates this storm | Prior loss photos |
| Deductible added back into unit prices | Carrier price list |
| Scope written for the whole dwelling | Loss-area photos |
| Drywall line billed, work not performed | Site re-inspection |

**What a partial denial needs** A finding on each line with the source behind it. WAC 284-30-330(4) says a reasonable investigation may not rely solely on a database, including estimating software and benchmarks.

Source: Illustrative per-line finding set, not a real claim: no amounts, percentages or multiples are implied. The line descriptions follow the indicators in NICB, Indicators of Contractor / Public Adjuster Property Repair Fraud, revised March 2015. The standard for the output is NAIC Model #900 section 4.L, adopted June 1990, and WAC 284-30-330(4), as amended effective 18 October 2026.

(Illustrative example)

Carriers meet the line-level version of this argument when it comes back the other way, in estimate disputes that run to [146 line items priced lower, only 28 missing](/blog/property-claim-estimate-disputes/).

## Cutting a line item is a partial denial, and it has a legal standard

**Is reducing a claim payment a partial denial?**

Paying a claim down because a line looks padded is a partial denial, and regulators treat it as one. Washington's rule, effective October 2026, states that denying or refusing to pay claims in part without a reasonable investigation is an unfair practice, and that a database alone is not a reasonable investigation.

[Washington Administrative Code 284-30-330](https://app.leg.wa.gov/wac/default.aspx?cite=284-30-330), amended effective 18 October 2026, names as a specific unfair claims settlement practice "denying or refusing to pay claims in part or in full without conducting a reasonable investigation", and adds that "a reasonable investigation may not rely solely on the use of a database, which includes, but is not limited to, estimating software and benchmarks gathered from one or multiple databases".

The [NAIC Unfair Claims Settlement Practices Act, Model #900](https://content.naic.org/sites/default/files/model-law-900.pdf), carries the same spine: section 4.F on refusing to pay claims without conducting a reasonable investigation, and 4.L on failing to provide "a reasonable and accurate explanation of the basis" for a denial or an offer of compromise settlement.

Most states enact some version of Model #900; the database sentence is Washington's own.

The asymmetry makes this an executive problem rather than a compliance footnote. Guessing high costs loss ratio, shows up as [claims fraud leakage](/blog/claims-fraud-leakage-pillar/), and nobody outside the company ever sees it. Guessing low costs a market conduct finding and a bad-faith exposure.

Only one of those two errors appears on a fraud dashboard, and it is not the expensive one. What a defensible file has to contain, item by item, is in [the claim file review checklist](/blog/decision-ready-claim-file/).

*Figure: A fraud score answers yes or no. A padded line asks how much.*

| Dimension | Binary fraud flag | Line-item verification |
| --- | --- | --- |
| Question it answers | Is this claim suspicious? | Is this line supported, and by how much? |
| Output | One claim-level score, or a yes/no referral | A finding per line, with a dollar delta and the source behind it |
| Fit for the largest fraud category | Poor. FRISS puts 44% of fraud in exaggeration of legitimate claims, methodology undisclosed; the claim is real, so there is no clean yes or no | Direct. The disputed object is the number on the line |
| How widely deployed | Automated red flags and business rules: 89% of carriers, the most widely used anti-fraud technology (CAIF 2024, n=35) | No industry adoption figure published |
| Known failure mode | High rates of false negatives and false positives, named a top deployment challenge by 57% of carriers (CAIF 2024) | Review cost per line, which is why it is not done at scale today |
| How success is measured today | Fraud detection rate and referral acceptance rate rank first and second among success measures (CAIF 2024) | Dollars corrected. Total loss mitigation ranks near the bottom of the same chart |
| Standing under WAC 284-30-330(4), eff. 18 Oct 2026 | A database or estimating-software benchmark alone is not a reasonable investigation for a partial denial | Produces the per-line basis the rule contemplates |
| Standing under NAIC Model #900 section 4.L | No per-line basis to explain | Supplies a reasonable and accurate explanation of the basis for an offer of compromise settlement |
| Who performs it | The detection vendor, then the SIU after referral | The adjuster, by hand, inside the cycle time |

## The scoreboard counts flags, not dollars

**How do carriers measure anti-fraud technology today?**

Carriers measure anti-fraud technology by how many flags it raises and how many referrals the SIU accepts, not by how many dollars came off an inflated claim. In the Coalition's 2024 study, fraud detection rate and referral acceptance rate rank first and second; total loss mitigation ranks near the bottom.

Look at what the [2024 Coalition study](https://insurancefraud.org/wp-content/uploads/2024-State-of-Technology-Study.pdf) says carriers bought. Automated red flags and business rules lead at 89%. Case management and data visualization with link analysis both sit at 77%, reporting capabilities at 74% and entity alert at 71%. Text mining has risen from 40% in 2012 to 63%, and geographic data mapping from 23% to 49%.

The reported benefits follow the same grain: increased speed of detection at 66%, more referrals at 54%, higher quality referrals at 46%. Every one of those counts flags or files, not dollars corrected.

**Figure 3 (What is deployed) - 89% run rules, and 57% name false positives among their top barriers.** Adoption of anti-fraud technology is broad and binary, and the misfire rate sits near the top of the barrier list carriers report themselves.

**Anti-fraud tools in use, 2024**

- Automated red flags and business rules: **89%**
- Case management: 77%
- Data visualization and link analysis: 77%
- Reporting capabilities: 74%
- Entity alert: 71%

**Also in use, both up since 2012**

- **63%** - text mining, up from 40% in 2012
- **49%** - geographic data mapping, up from 23% in 2012

**57%** - name high rates of false negatives and false positives among their biggest challenges in deploying fraud detection technology.

Source: Coalition Against Insurance Fraud, 2024 State of Insurance Fraud Technology Study, sixth edition with Shift Technology. 35 insurer-member responses, fielded 3 Oct - 14 Nov 2024; the report notes lower participation than prior waves because of two hurricanes during fielding. Barriers as listed: limited IT resources 86%, data integration and poor data quality 60%, false negatives and false positives 57%.

A threshold tuned to protect scarce investigator capacity necessarily discards the category FRISS puts at 44% of fraud, methodology undisclosed, because no single padded line clears it.

The capacity arithmetic behind the threshold is not subtle. Manual SIU investigation runs 14+ days per case and one investigator carries 200+ cases, which is why roughly 25% of flagged claims get a full workup at all; those three are Hesper internal benchmarks. The funnel looks the same from the other end: [51,419 reports in, 243 cases opened](/blog/siu-fraud-referral-package-automation/).

What happens to the rest of the flagged queue is in [why 75% of flagged insurance claims are never fully investigated](/blog/why-flagged-insurance-claims-never-investigated/).

## So it lands on the adjuster, inside the cycle time

**Who ends up deciding whether a line item is inflated?**

With no referral and no score to lean on, the inflated line item becomes an adjuster's judgement call made under a clock. J.D. Power puts the average homeowners claim at 40.7 days from first notice of loss to final payment, and every carrier is trying to shorten it.

The 2026 U.S. Property Claims Satisfaction Study, [reported by Claims Journal in March 2026](https://www.claimsjournal.com/news/national/2026/03/18/336327.htm), surveyed 5,093 homeowners who had filed a claim within the previous nine months. Average time from first notice of loss to final payment: 40.7 days, down 3.4 days year over year.

Six weeks end to end, and shrinking by design. Inside that window an adjuster reads the policy, scopes the damage, agrees a number and pays it, and somewhere in there decides whether line 47 is supported. The adjuster is asked to move faster and catch more on the same file, and the line item is the first thing that gets conceded.

> A padded line item is not a confession. It is an opening offer, and you do not answer an opening offer with a fraud flag. You answer it with a number and the source behind the number.
>
> - Hesper AI product research

## What a line-item instrument has to produce

**What does a line-item instrument have to return?**

A line-item instrument returns a finding per line, not a score per claim: this labour operation, this contents item, this treatment code, supported or not, by how much, and against which source. That output is what a partial denial needs under NAIC Model #900 and what a fraud score cannot supply.

Work backwards from the regulation and the specification writes itself. A per-line cut needs a per-line basis: sourced, recorded when the check was made, and legible to somebody who was not there.

A detection vendor's score stops at the flag. An estimating database stops at a benchmark, and WAC 284-30-330(4) has just said a benchmark on its own is not a reasonable investigation. Neither layer was built to return a finding on line 47, and neither claims to be.

FRISS's chapter argues that carriers should build for high-volume opportunistic fraud and low-volume sophisticated fraud at once, with AI-powered triage as the mechanism. FRISS's own data is the public evidence the premise here rests on, and it deserves credit.

The arguments part on one narrow point: triage routes a claim to the right queue, and once it has done its job a partially inflated claim is still a real claim with a disputed amount in front of an adjuster. Triage is a routing answer to a verification problem.

FRISS publishes its own report under the title The State of Insurance Fraud 2026; Hesper's separate research page, [State of Insurance Fraud Detection 2026](/reports/state-of-insurance-fraud-2026/), is a different document with different figures, and nothing from the FRISS report is folded into it.

Hesper AI's agents take every claim from first notice to final recovery. On a claim where the loss is real and the amount is disputed, the agents run 15+ investigation phases in parallel and return each line with its finding and the source behind it, in minutes, not weeks. Adjusters review evidence-backed files instead of building them, and the human keeps every decision.

The same evidence engine which produced the per-line finding also produces the coverage position, the settlement valuation and the subrogation demand, because this is one file moving from first notice to final recovery. That is what moves investigation coverage from roughly 25% of flagged claims to 100% of them: not a faster investigator, but an instrument whose capacity is not capped by headcount.

| Dimension | Manual line-item review | Automated line-item verification |
| --- | --- | --- |
| Time per suspicious file | Manual SIU investigation runs 14+ days per case (Hesper internal benchmark) | Minutes, not weeks (Hesper) |
| Investigation phases | Sequential, one investigator at a time | 15+ phases in parallel (Hesper) |
| Coverage of flagged claims | Roughly 25% get a full workup (Hesper internal benchmark) | 100% of flagged claims (Hesper) |
| Caseload constraint | 200+ cases per investigator (Hesper internal benchmark) | Not capped by headcount (Hesper) |
| Cycle-time context | Homeowners claims average 40.7 days from first notice to final payment (J.D. Power, reported March 2026, n=5,093) | The same window, with more of it spent on the decision than on assembling the file |
| Output for compliance | Varies by adjuster and by file | A documented per-line basis, sourced and timestamped |

Organized rings and fabricated losses need the other instrument - cross-carrier data and network analysis - which is why built-in detection is complementary to FRISS and Shift rather than a replacement. The mistake is not owning a detection stack. It is assuming that because the stack covers fraud, it covers the real claim with a wrong number on it, which FRISS puts at 44% of fraud, methodology undisclosed.

## Key takeaways

- FRISS attributes 44% of fraud to opportunistic exaggeration of legitimate claims, against 29% for fake injury claims and 13% for organized crime, methodology undisclosed.
- A fraud score returns a claim-level verdict, which is the wrong shape for a file where the loss is real and only the amount is disputed.
- Cutting a line item is a partial denial: WAC 284-30-330(4), effective 18 October 2026, says a reasonable investigation may not rely solely on a database including estimating software and benchmarks.
- Carriers measure anti-fraud technology by flags and referrals rather than dollars corrected.
- The defensible instrument is a finding on each line with the source behind it, produced inside the 40.7-day homeowners cycle rather than after it.

## Frequently asked questions

### What is an inflated insurance claim?

An inflated insurance claim is a legitimate loss submitted for more than it is worth. The event happened and the policy responds, but one or more line items are overstated: prior damage folded into a new loss, labour hours that were not worked, or treatment that was not needed. The Insurance Research Council calls this buildup and defines it as the inflation of otherwise legitimate claims.

### How do insurance companies investigate inflated claims?

Differently from how they investigate suspected fraud, because the claim itself is not in question. The work is line-item verification: comparing each item on the estimate, invoice or contents list against an independent source. A claim-level fraud score cannot produce this: it returns a probability for the whole file, not a finding on a line.

### What is claim build-up in auto insurance?

Build-up is the Insurance Research Council's term for inflating an otherwise legitimate auto injury claim, usually through unnecessary treatment or diagnostic procedures rather than a fabricated injury. In the IRC study published in February 2015 using 2012 data, claims with the appearance of build-up accounted for 15% of dollars paid for BI and PIP claims.

### Can an insurance company deny a claim for exaggeration?

A carrier can decline the inflated portion, but doing so is a partial denial with its own legal standard. Washington's unfair claims rule, amended effective 18 October 2026, prohibits denying or refusing to pay claims in part or in full without conducting a reasonable investigation, and adds that a reasonable investigation may not rely solely on the use of a database. In practice: you need per-line evidence, not a benchmark.
