Insurance Services Office, the Verisk unit whose advisory loss costs sit underneath most US property-casualty rate filings, was granted US Patent 12,694,109 B2 on July 28, 2026 for a system that decides whether a claims PDF has been altered (USPTO Official Gazette). It runs roughly twenty independent structural checks, collapses them into one suspicion score, and draws a box around the region that changed.

Key Takeaways

  • 38 claims issue on an application filed February 8, 2024, with priority to a provisional filed exactly one year earlier.
  • Nothing in the architecture requires a neural network. Each check emits a weight between 0 and 1 and an opinion score between -1 and 1; the system multiplies the pair and averages across the checks that applied.
  • Industry work puts leakage at 7% to 14% of total claims payouts (Coalition Against Insurance Fraud), covering unvetted vendor invoices, missed deductibles and duplicate payments. The target is soft leakage rather than organized rings.
  • 36% of consumers would consider digitally altering a claim document, rising to 55% among Gen Z, against only 32% of insurers very confident they could identify a deepfake.

Patent Details

Patent numberUS 12,694,109 B2
TitleSystems and methods for detecting, localizing, and visualizing manipulations of portable document format files
AssigneeInsurance Services Office, Inc., Jersey City, NJ
InventorsMatthew D. Frei (Lehi, UT); Cherie Ekholm (Redmond, WA); Shubhi Singh (Ottawa, CA); Zheng Zhong (Seattle, WA); Palak Jain (Draper, UT)
Application18/436,913, filed February 8, 2024
PriorityProvisional 63/444,021, filed February 8, 2023
Pre-grant publicationUS 2024/0265103 A1, August 8, 2024
GrantedJuly 28, 2026 (Official Gazette Vol. 1548 No. 4)
Claims38
CPCG06F 21/565 (file-based threat detection); G06V 30/41 (document analysis); G06Q 40/08 (insurance)

What the Checks Actually Test

The specification states the premise plainly: "PDF files have a great risk of fraud and/or manipulation, which presents a significant security risk" (US 2024/0265103 A1). The format's own integrity features, digital signatures and encryption, are optional and largely unused in claims workflows, so the system infers tampering from what an editor leaves behind rather than from what the author declared.

The roughly twenty check processes fall into four families. Structure and versioning reads trailers and UUIDs revealing multiple incremental saves, header versions that disagree with the document catalog, and cross-reference tables permitting reconstruction of prior generations. Provenance metadata compares creator and producer applications against a blocklist of suspect editors, creation dates against modification dates, and whether the file carries metadata at all. Rendering artifacts catch objects placed on top of original text, subset fonts mixed with the Base 14 set, missing glyphs and incomplete tagging. Content and semantics validates claim dates, signature validity under ISO 32000, and the arithmetic of totals, taxes and line items against reference costs.

That last one is underwriting judgment made mechanical. It recomputes an invoice and tests the line items against reference costs, which is the same test a good bill reviewer applies by hand and almost never applies to every document.

The scoring is deliberately simple. A weight fixes how much a check can move the answer, described as "a static value that can be set using predetermined values, set by users, set by the system, or set using a training process in which ground truth data and labeled data are used." Worked examples put the digital-signature check near 0.9 and the creation-and-modification-date check nearer 0.5. The opinion score is the check's own certainty, from -1 for a file it affirmatively trusts to 1 for one it does not, with null for checks that do not apply. The system multiplies weight by opinion, averages the products, and excludes the nulls.

The classification is worth noting for why it survived examination. G06Q 40/08, the insurance subclass that has drawn the most Alice-era scrutiny, appears only as a secondary. The primary is G06F 21/565, file-based threat detection, where antivirus engines live. The claimed advance operates on trailers, cross-reference tables, font tables and object overlap coordinates, none of which is an abstract idea about evaluating an insurance claim. Patents that put the novelty below the business decision have been clearing the bar while those that put it in the decision have not, a divide the site traced in the Section 101 reset.

What It Is Worth on the Combined Ratio

The translation is worth doing arithmetically, with the caveat that the inputs are illustrative rather than sourced. Take a book paying $1 billion of losses a year at a 65% loss ratio. Leakage at the midpoint of the 7% to 14% range is about $105 million. If document manipulation drives a tenth of that and screening recovers half of what it drives, the indemnity saved is roughly $5.3 million: about 53 basis points of paid loss, or 34 basis points of combined ratio.

That is real, worth having, and an order of magnitude smaller than the fraud headline suggests. The gap between the two matters for how the capability gets sold and how its benefit gets priced into a rate filing, because a carrier crediting a fraud-headline number rather than a leakage-arithmetic number is building an assumption the recovery cannot support.

The exposure concentrates where documents carry the indemnity rather than merely describing it. Auto physical damage runs on repair estimates. Property runs on contents inventories and restoration invoices. No-fault, PIP and workers compensation medical run on billing statements. Those are the lines where a PDF is the payment instrument, and the same lines where the site has tracked AI moving into the recovery side in CCC's subrogation scoring and in coverage agents built explicitly against leakage.

There is a reserving consequence in the timing. Screening that goes live mid-year changes paid severity on the affected diagonals without changing the underlying loss, so a carrier that cannot name the go-live date by line of business will read a genuine leakage recovery as favorable development.

The Ground Truth Is the Asset

None of the individual checks is hard. A competent engineering team can parse a cross-reference table, read a producer string, and compare a header version to a document catalog. The reproducible part of this patent is the part that does not matter.

What is not reproducible is the labeling. The specification gives it away: weights can be set "using a training process in which ground truth data and labeled data are used." Ground truth for document forgery means confirmed-manipulated claim documents at volume, matched to SIU outcomes, across enough carriers and enough document types that the weights generalize past one book's vendor mix. A single carrier's SIU file will not produce that corpus. Verisk's position in the claims data flow does, and that asymmetry rather than the algorithm is the durable asset, the same structure the site described in Verisk's loss-cost pricing power.

The grant adds an exclusion right on top of the data advantage, and management has told the market the line is selling. Elizabeth Mann attributed anti-fraud growth to "good price realization, early renewals, adoption of newer innovations, including digital media forensics" (Q2 2026 earnings call), against Claims segment revenue of $237 million, up 6.1% organically (Verisk Q2 2026 results).

The compounding runs against the carrier. Every flagged-and-confirmed document a carrier feeds back is a training example, and whether the contract lets the carrier retain the labeled outcomes its own claims generate, or whether that ground truth accrues only to the vendor, decides who owns the next version of the weights. A carrier that buys the screening and gives away the labels is funding the improvement of a model it will keep renting, on data it produced.

Further Reading

Sources