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 (USPTO Official Gazette, July 28, 2026).

That retrieval step matters because an LLM's answer only reflects the chunks retrieval handed it, not the underlying document. The claim language is the most precise public description available of how a production extraction system at a major carrier is designed to behave, and of the uncertainty signal it generates but never has to pass along.

Executive Summary

The patent, titled "Information extraction system for unstructured documents using retrieval augmentation with custom query parameters," was granted to American International Group, Inc. on July 28, 2026 (USPTO Patent Full-Text and Image Database). It is AIG's fourth granted patent in this architecture family. actuary.info's earlier coverage of AIG's document-extraction patent portfolio tracked three prior grants covering table-versus-text separation and chunk indexing; this filing adds the retrieval-scoring and re-retrieval-on-failure mechanics that sit directly downstream of that indexing layer.

The system's validator is a real engineering safeguard, but it confirms structural plausibility, not truth. A validated, well-formed extraction is not the same claim as an accurate one. 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.

Just as important is what the claims do not require: nothing in the patent obligates the confidence score, distance metric, or re-retrieval flag to travel downstream with the extracted value. The reliability signal is architecturally separable from the field a rating engine consumes.

The scale is already material. Lexington Insurance had routed more than 370,000 submissions through AIG Assist as of February 2026, against an original 2030 target of 500,000. At that volume, even a modest error tail compounds across a meaningful share of a major carrier's submission flow before an actuary sees the resulting exposure record.

For actuaries, the practical response is a short list of questions about confidence-score distributions, re-retrieval trigger rates, and reconciliation sampling, laid out at the end of this article. A carrier that can answer them is treating the extraction layer as the actuarial control point it actually is.

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

The Mechanism: How the Retrieval Layer Decides What the Model Sees

Claim 1 describes a method that 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, US12694704B1). The weights themselves are tunable retrieval parameters.

A Tunable Dial, Not a Fixed Formula

Claim 1's retrieval parameters include a "search type designation" specifying semantic search, keyword search, or both, plus "weighting parameters for a weighted function of the keyword scores and the distance metrics" (Claim 1, US 12,694,704). The system's designers choose that weighting per field type.

A field where exact terminology matters, a policy form number or an ISO class code, can be weighted toward keyword matching. A field better captured by paraphrase, a narrative description of operations, can be weighted toward semantic similarity. The specification also describes hierarchical search criteria intended to guarantee "a minimal number of chunks are retrieved" (Abstract, US 12,694,704), an efficiency goal that trades off directly against retrieval completeness.

The Re-Retrieval Loop

This is the part with the sharpest actuarial edge. When a "generative AI manager" determines that retrieved chunks fail to satisfy a retrieval criterion, the system's semantic and keyword searchers redeploy with "a secondary (e.g., second, broadening, etc.) set of search parameters" (US 12,694,704, specification).

In plain terms: if the first pass doesn't find a confident answer, the system tries again with looser criteria before giving up. The loop is designed to resolve ambiguity by trying harder, not by surfacing residual uncertainty to whatever consumes its output.

The Response Validator

Before a result is stored, a "response validator" checks the LLM's output against an expected type, length, or numeric range stored per prompt template (US 12,694,704, specification). That is a real safeguard against obviously wrong outputs, a text string where a dollar figure was expected. It is not a safeguard against a plausible but wrong figure that happens to fall inside the expected range, which is exactly the failure mode that matters for a rating variable.

The Problem: Validation Confirms Format, Not Accuracy

The distinction actuaries need to hold onto is that the validator asks: is this a number, does it fall in range, is it 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 the patent describes whether the true figure was $2.1 million, $2.4 million, or $1.9 million.

Commercial underwriting AI tools built on comparable retrieval architectures report field-level accuracy above 90% on clean, well-structured fields like named insured and total insured value, with materially lower accuracy on scanned documents or ones carrying handwritten corrections (V7Labs, 2026). A 90% ceiling on the easy fields implies a meaningfully lower one on the hard fields.

Nothing in the claim language requires the stored output to carry a confidence score, a distance metric, or a re-retrieval flag forward to downstream systems. The patent's own language is that the system stores "a response from the large language model to the extraction prompt and the one or more relevant chunks" (Claim 1, US 12,694,704). That preserves an audit trail for the extraction event itself, but says nothing about propagating uncertainty into the record a rating engine consumes.

This is the upstream version of the garbage-in problem. It is not that the model is unreliable; it is that the reliability signal is separable from the extracted value, and the patent does not specify that it travels with it.

Actuarial Implications

Pricing: Where a Mis-Extracted Field Lands

Not every extraction error carries equal weight, which is why a blended accuracy statistic understates the real 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.

Vertafore's comparable Velocity AI submission agent, covered in actuary.info's reporting on MGA submission-intake automation, runs at roughly 87% field-extraction accuracy across a blended document mix (Vertafore, July 2026), meaning close to one in eight fields carries some error before human review.

AIG's patent describes a more sophisticated retrieval architecture than a flat extraction pass, but sophistication in retrieval does not change where the error lands once it clears validation. 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. A construction-type field mis-tagged from a narrative property description does the same thing to a fire class relativity.

The patented retrieval reduces the frequency of these errors relative to a naive keyword or embedding-only search. It does not eliminate the category, and no combination of semantic and keyword weighting can substitute for a field carrying an explicit error bar into the rating engine.

Reserving: When Extraction Drift Masquerades as Mix Shift

The same mechanism reaches reserving through a different door. Cause-of-loss and coverage tagging, if automated with a comparable retrieval pipeline, feeds directly into case reserving and IBNR segmentation, since both depend on classifying claims into the buckets a reserving actuary's triangles are built around.

A retrieval system that broadens its search parameters on a failed first pass will occasionally settle on a plausible-but-wrong classification rather than surfacing genuine ambiguity to a human adjuster. If that happens with any systematic bias, more often on complex multi-peril claims than simple ones, for example, a reserving actuary reviewing a loss triangle by accident year will see 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. A reserving actuary who has not asked whether a claims system's coverage-tagging pipeline changed extraction architecture in the period under review is implicitly assuming the tagging was stable, an assumption a retrieval-based classifier does not automatically satisfy the way a fixed rules engine did.

Governance: Not a Rating Model, But Right Above One

The NAIC's Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023, is now in force in roughly 24 states and the District of Columbia as of the NAIC's Spring 2026 National Meeting (Quarles Law Firm, 2026). The bulletin requires a written AI Systems Program with model validation and testing for third-party AI tools, and it makes the insurer responsible for a vendor's AI system's outcomes regardless of who built the underlying model (NAIC, Insurance Topics: Artificial Intelligence).

A RAG extraction system built to AIG's patented architecture is a genuinely 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 AI governance programs are organized around the rating or claims-triage model itself. 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' 2024 paper on generative AI professionalism considerations addresses exactly this seam: an actuary relying on a third-party tool's output is still responsible for understanding, to a reasonable degree, that tool's limitations before treating its output as reliable input (American Academy of Actuaries, September 2024).

For a RAG extraction pipeline, that due diligence needs to reach the extraction layer's confidence-score distribution, its re-retrieval trigger rate, and whatever reconciliation sampling exists between extracted fields and source documents, none of which a vendor's marketing materials are likely to volunteer without being asked. The practical fix is to fold that data-lineage documentation into the AI Systems Program's third-party-vendor section rather than treating it as a separate operations-only concern, the same fix actuary.info's coverage of MGA submission-intake automation reached from the vendor side of this problem.

Confidence Score vs. Rating Materiality: A Reconciliation Framework

Extracted field type	Typical rating leverage	What a reconciliation program should check
Payroll, total insured value, fleet count	High, multiplies directly into premium	Sample against source document regardless of reported confidence score
Construction type, occupancy, cause-of-loss tag	High, drives class-plan relativity or coverage bucket	Route below-threshold confidence to mandatory human review
Named insured, mailing address, policy number	Low, rarely moves the rate	Periodic audit only, lower sampling priority
Re-retrieval trigger rate by field type	Indirect, signals where the model is least confident	Track over time as a leading indicator of extraction drift

The Pattern: The Third Extraction-Layer Story in Five Months

AIG's patent is the latest instance of a pattern actuary.info's coverage has now tracked across three separate vendors: an extraction or intake layer gets built, marketed, and adopted faster than the governance documentation for it matures. Vertafore's Velocity AI agent, EXL's document-processing patents covered in actuary.info's reporting on EXL's knowledge-graph architecture, and now AIG's own retrieval-scoring patent all sit at the same structural position: upstream of a rating or reserving model, producing inputs rather than decisions, and consequently outside the scope most AI governance programs were built to review.

AIG's scale gives this instance more weight than most. Lexington Insurance, AIG's excess and surplus unit, had routed more than 370,000 submissions through its AIG Assist tool as of AIG's February 2026 investor update, against an original 2030 target of 500,000. CEO Peter Zaffino told analysts the outcomes had moved from "aspirational" to "beyond our expectations" (Insurance Journal, February 13, 2026).

AIG's Q2 2026 results, disclosed five months after that call, showed an 89% calendar-year combined ratio, improved 30 basis points year over year, with the accident-year figure at 88.1%, also improved 30 basis points (AIG Q2 2026 earnings call transcript, The Motley Fool, August 13, 2026). AIG's public risk-factor disclosures already list the pace of AI implementation, and its ability to respond to competitors' AI initiatives, among the risks it flags to investors (AIG Form 10-Q, SEC EDGAR, 2026).

None of those figures prove the extraction layer is producing bad data. They establish that the volume running through a system like this one is now large enough, and growing fast enough, that even a modest error tail compounds across a meaningful share of a major carrier's submission flow before an actuary downstream ever sees the resulting exposure record.

Four Questions to Ask Before Trusting the Output

An actuary evaluating a book of business underwritten through a RAG-based extraction pipeline built on these mechanics has a specific, answerable set of questions worth putting to the underwriting or IT team that deployed it:

1. What confidence-score or distance-metric distribution did the system produce across the cohort being reviewed, and is that distribution retained anywhere an actuary can access it?
2. How often did the re-retrieval loop trigger, and does that rate vary meaningfully by document type or line of business? A high trigger rate on a specific class is itself a signal of where extraction is least reliable.
3. Is there a reconciliation sample comparing extracted fields against source documents, weighted toward high-rating-leverage fields rather than applied uniformly regardless of premium impact?
4. Does the confidence signal survive into the record a rating or reserving model later consumes, or does it get discarded the moment the extraction passes validation?

A carrier that can answer all four is treating the extraction layer as the actuarial control point it actually is. One that cannot is trusting a retrieval score it never asked to see.