BRENNEN

Industries / Finance and business services / Security guard services / Private investigation

Insurance and fraud cases

Insurance and fraud work is checking claims that may be invented or inflated, usually for an insurer's special investigations unit (SIU), by reviewing the claim file, taking statements and checking documents.

5

Milestones achieved, from 1988 to 2023.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1988

    Fraud examiners organize

    Joseph T. Wells, an accountant and former FBI agent, founded the Association of Certified Fraud Examiners (ACFE) in Austin, Texas, in 1988. By the ACFE's account, the idea came from a 1985 conversation in which the criminologist Donald Cressey told Wells that fraud work, then split between accountants who were not experts in investigating fraud and investigators who were not skilled in accounting, needed to "combine these disciplines". The ACFE says it now has more than 100,000 members in over 180 countries.

    legacy.acfe.comacfe.com

    Credit

    Joseph T. Wells; Donald Cressey

  2. 1992

    Theft and fraud bureaus merge

    Insurers had pooled their car-theft investigators since 1912, and their bureaus were consolidated in 1927 as the National Automobile Theft Bureau (NATB). In 1992 the NATB merged with the Insurance Crime Prevention Institute to form the National Insurance Crime Bureau (NICB), joining vehicle investigation with insurance fraud prevention. In 2002 the NICB set up its first Major Medical Fraud Task Force in south Florida and New York City, working with police and insurers' SIUs.

    claimsjournal.com

    Credit

    the NICB

  3. 1998

    Fraud units made mandatory

    New York's second amendment to Regulation 95 took effect on 1 July 1998. Insurers writing more than 3,000 car, workers' compensation or accident and health policies a year in the state had 120 days to file a plan for detecting, investigating and preventing fraud. The filing form asks whether the insurer has a full-time SIU separate from underwriting and claims, with staff who meet the regulation's education and experience requirements, or a contractor in its place.

    dfs.ny.gov

    Credit

    Charles D. DeRienzo, director of the New York Insurance Frauds Bureau

  4. 2022

    Most insurers model fraud

    In a survey by the Coalition Against Insurance Fraud and the software company SAS, sent to 100 coalition members in October 2021, 80% of insurers said they used predictive modeling to detect fraud, up from 55% in 2018.

    dig-in.com

    Credit

    Coalition Against Insurance Fraud; SAS

  5. 2023

    Claim photos checked across insurers

    In June 2023 Verisk added AI image forensics to ClaimSearch, a claims database to which, by Verisk's count, more than 95% of the US industry contributes. It checks each claim photo for when and where it was taken and whether it has been altered, copied from the web or sent to other insurers. In a Verisk study of 768,000 claim images, 1,967 were duplicates, one of them used in 44 claims, and they were linked to 1,475 claims on which about $5.3 million had been paid. When it was announced, 315 insurers had signed up to contribute images.

    investor.verisk.com

    Credit

    Verisk

Watch for

  1. Deepfake alerts in claims

    Verisk says a new Image Deepfake alert in its Digital Media Forensics product helps claims teams "instantly spot signs of AI-generated or manipulated visuals". The release note gives no accuracy figures.

    Achieved whenan insurer or outside tester publishes detection and false-alarm rates on claim images that Verisk did not choose.

    verisk.com

    Claimed by

    Verisk, June 2026

  2. Fraud flags from first calls

    A June 2026 preprint presents a framework that generates synthetic first notice of loss (FNOL) calls, the call in which a customer first reports a claim, as transcripts and two-speaker audio. It combines speech recognition, speaker matching and AI retrieval in a rule-based risk score that flags reused stories, inconsistencies and the same voice across claims. The only validation is the authors' own, on their synthetic data.

    Achieved whenan insurer reports results on real calls, or another team reproduces them.

    arxiv.org

    Claimed by

    the authors of arXiv paper 2606.28002, June 2026

Open

  1. 01

    File review for SIU work

    200 fictional claim files of the kind an SIU investigator reviews: claim form, recorded statement transcript, police report, repair estimate, invoices, and photos with their metadata. Some carry planted inconsistencies, such as a loss date that does not match a photo's metadata, a statement that contradicts the police report, or an invoice for parts not in the estimate, all listed in an answer key. AI tools and experienced investigators are each asked to list every inconsistency and the page it is on.

    The brief
    Credit · unclaimed
  2. 02

    Claim documents checked against facts

    A tool that pulls the checkable facts out of a document sent in with a claim, such as a repair invoice, a travel cancellation letter or a supplier's receipt, and checks them against public sources: that the business exists in its company register, that a VAT number is valid, that the figures add up, that the flight on a cancellation letter existed on that date. It is tested on 1,000 test documents, half of them with planted errors, each marked as a test document.

    The brief
    Credit · unclaimed

Open 01 · the brief

File review for SIU work

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

200 fictional claim files of the kind an SIU investigator reviews: claim form, recorded statement transcript, police report, repair estimate, invoices, and photos with their metadata. Some carry planted inconsistencies, such as a loss date that does not match a photo's metadata, a statement that contradicts the police report, or an invoice for parts not in the estimate, all listed in an answer key. AI tools and experienced investigators are each asked to list every inconsistency and the page it is on.

Why it is still open

The closest published work, the June 2026 FNOL framework, covers call transcripts and audio, not whole claim files, and says existing methods "rely largely on private, text-only datasets". In Verisk's 2026 survey of 300 claims professionals, 38% said their tools miss too many altered images and documents, and 35% reported false positives.

arxiv.orgclaimsjournal.com

Built with what exists

Claude Fable 5.2 to write the files (see the Note "Three working devices"); ElevenLabs Scribe v3 to transcribe statements recorded by volunteer actors reading scripts (see the Note "Fewer word errors"); a Claude-based reviewer of the kind Enveda describes for drug-making records, which raised issues people had missed (see the Note "Seven more flagged").

  1. First step

    Write 10 files with 5 planted inconsistencies each and have one investigator review them without the answer key.

  2. Done looks like

    200 files with about 1,000 planted inconsistencies; at least 3 AI tools and 5 SIU investigators scored on inconsistencies found, inconsistencies invented and minutes per file.

Open 02 · the brief

Claim documents checked against facts

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A tool that pulls the checkable facts out of a document sent in with a claim, such as a repair invoice, a travel cancellation letter or a supplier's receipt, and checks them against public sources: that the business exists in its company register, that a VAT number is valid, that the figures add up, that the flight on a cancellation letter existed on that date. It is tested on 1,000 test documents, half of them with planted errors, each marked as a test document.

Why it is still open

AIForge-Doc, a February 2026 benchmark of receipts and forms altered with AI editing tools, found that detectors that inspect pixels fail on them: a GPT-4o judge scored 0.509, "essentially at chance". Admiral, a UK insurer, says fabricated documents are sent in with claims "on a regular basis". Searches found no published test of checking a claim document's facts rather than its pixels.

arxiv.orgadmiral.com

Built with what exists

the free Companies House API and GLEIF's free Legal Entity Identifier (LEI) data for company checks; the EU's VIES service for VAT numbers; OpenSky Network's historical flight data; Claude to extract the facts; Gemini Decision 1 to give each check a probability (see the Note "A thousand questions, one pass").

  1. First step

    Make 50 test invoices naming companies from Companies House's free data, plant one error in 25 of them, and check all 50 by hand.

  2. Done looks like

    1,000 test documents in English and two other languages; the share of planted errors caught at a false-alarm rate of 5% or less, compared with two pixel-based detectors on the same documents.