AIG's New Patent Puts a Confidence Score Between the Submission and the Rate

AIG's newest granted patent, US 12,694,704, covers the retrieval step that decides which parts of an insurance submission a large language model actually reads. A weighted blend of semantic and keyword search picks the passages, a validator checks the model's answer against an expected type and range, and a failed check triggers a second, broader retrieval pass before anything is stored. That step matters because an LLM's answer reflects only the chunks retrieval handed it, not the underlying document.

Key Takeaways

• AIG's fourth granted patent in this architecture family. Three prior grants covered table-versus-text separation and chunk indexing; this one adds the retrieval-scoring and re-retrieval-on-failure mechanics sitting directly downstream.
• The validator confirms structural plausibility, not truth. A payroll figure passes every check the patent describes if it is numeric and falls in a plausible range, whether or not it matches the source document.
• Nothing obligates the confidence score to travel downstream. The reliability signal is architecturally separable from the field a rating engine consumes, and the claims do not specify that it moves with the value.
• More than 370,000 submissions had been routed through AIG Assist at Lexington Insurance as of February 2026, against an original 2030 target of 500,000.

Patent Details

Patent Number	U.S. 12,694,704 B1
Title	Information extraction system for unstructured documents using retrieval augmentation with custom query parameters
Filed	January 24, 2025
Granted	July 28, 2026 (Official Gazette week 30)
Assignee	American International Group, Inc.
Portfolio position	Fourth grant in AIG's document-extraction family, after U.S. 12,437,154, 12,437,155, and 12,511,320

How the Retrieval Layer Decides What the Model Sees

Claim 1 separates a document into table chunks and text chunks, indexes both with vector embeddings, generates distance metrics between those embeddings and a query's embedding, generates a separate keyword-matching score, and combines the two through a weighted function (Google Patents). The weights are tunable retrieval parameters, and the designers choose them per field type.

A field where exact terminology matters, an ISO class code, can be weighted toward keyword matching; one better captured by paraphrase toward semantic similarity. The specification also describes hierarchical criteria intended to guarantee "a minimal number of chunks are retrieved," an efficiency goal that trades off directly against retrieval completeness.

The sharpest mechanism is the re-retrieval loop. When a generative AI manager determines that retrieved chunks fail a retrieval criterion, the semantic and keyword searchers redeploy with "a secondary (e.g., second, broadening, etc.) set of search parameters." In plain terms, if the first pass does not find a confident answer, the system tries again with looser criteria before giving up. The loop resolves ambiguity by trying harder, not by surfacing residual uncertainty to whatever consumes its output.

Before a result is stored, a response validator checks the output against an expected type, length, or numeric range held per prompt template. That is a real safeguard against an obviously wrong output, a text string where a dollar figure was expected. It is not a safeguard against a plausible but wrong figure that falls inside the expected range, which is exactly the failure mode that matters for a rating variable.

Validation Confirms Format, Not Accuracy

The validator asks whether this is a number, whether it falls in range, whether it is the right length. It does not ask whether the number matches what the source document said. A payroll figure of $2.1 million extracted from a scanned loss run passes every check whether the true figure was $2.1 million, $2.4 million, or $1.9 million.

Not every error carries equal weight, which is why a blended accuracy statistic understates the risk. A misread mailing address rarely moves a premium calculation; a misread payroll figure, total insured value, or fleet count multiplies directly into a rating variable. Comparable tools report accuracy above 90% on clean fields and materially lower on scanned or hand-corrected documents, and Vertafore's submission agent runs at roughly 87% across a blended mix, close to one in eight fields carrying some error before human review (site coverage).

The arithmetic is direct. A payroll figure of $2.04 million extracted where the source read $2.4 million understates workers' compensation exposure by roughly 15%. Unless a reconciliation step catches it, that understatement flows through to bound premium and into the exposure base a class-plan relativity is later calibrated against. AIG's retrieval architecture reduces the frequency of these errors relative to a naive keyword or embedding-only search. It does not eliminate the category, and no weighting scheme substitutes for a field carrying an explicit error bar into the rating engine.

Reserving is reached through a different door. Cause-of-loss and coverage tagging feeds case reserving and IBNR segmentation, since both depend on classifying claims into the buckets a triangle is built around. A retrieval system that broadens its parameters on a failed first pass will occasionally settle on a plausible-but-wrong classification rather than surfacing ambiguity to an adjuster. With any systematic bias, more often on complex multi-peril claims than simple ones, a reserving actuary reviewing a triangle by accident year sees something that looks like an ordinary mix shift between coverage lines.

It is not a mix shift. It is a classification artifact introduced upstream of the triangle, and standard triangle diagnostics are not built to distinguish the two. An actuary who has not asked whether a claims system's tagging pipeline changed extraction architecture in the period under review is assuming the tagging was stable, which a retrieval-based classifier does not satisfy the way a fixed rules engine did.

Upstream of Every Governance Program

The NAIC's AI model bulletin, now in force in roughly 24 states and the District of Columbia, requires a written AI Systems Program with model validation and testing for third-party tools, and makes the insurer responsible for a vendor system's outcomes regardless of who built it (NAIC).

A RAG extraction system is an awkward fit for that framework, because it neither scores risk nor sets a price. It produces the exposure fields a separate rating model then consumes. Most governance programs are organized around the rating or claims-triage model itself, so an extraction layer filed under "operations" or "document processing" can sit entirely outside the validation scope, even though its error rate moves the same rating variables the validated model downstream is calibrated against.

The American Academy of Actuaries reached the same seam: an actuary relying on a third-party tool remains responsible for understanding its limitations before treating the output as reliable input (American Academy of Actuaries, September 2024). For a retrieval pipeline that reaches the confidence-score distribution, the re-retrieval trigger rate, and whatever reconciliation sampling exists, none of which vendor materials volunteer unasked.

This is now the third extraction-layer story in five months, after Vertafore's agent and EXL's knowledge-graph document patents. All three sit at the same structural position: upstream of a rating or reserving model, producing inputs rather than decisions, and consequently outside the scope most governance programs were built to review.

Scale is what gives this instance weight. Lexington had routed more than 370,000 submissions through AIG Assist by February 2026, and CEO Peter Zaffino told analysts the outcomes had moved from "aspirational" to "beyond our expectations" (Insurance Journal, February 13, 2026). None of that proves the extraction layer is producing bad data. It establishes that the volume is now large enough, and growing fast enough, that a modest error tail compounds across a meaningful share of a major carrier's submission flow before an actuary sees the resulting exposure record.