What AIG’s AI Underwriting Patents Mean for Carriers Building Their Own Systems

This article is a companion to our AIG patent analysis series, which examines the three granted patents in technical detail. This piece focuses on the competitive and practical implications for other carriers.

The Strategic Signal

When a major insurance carrier patents its AI underwriting methods, it is making a statement that goes beyond technology. AIG now holds three granted U.S. patents covering the core architecture of its AI-assisted document processing system. All three were invented by the same two-person team at AIG’s Atlanta Innovation Hub, all three name Claude (Anthropic’s LLM) as the model used in the system, and all three explicitly describe insurance underwriting as the primary use case.

For carriers, insurtechs, MGAs, and technology vendors building competing AI document extraction systems, these patents raise a practical question: does our architecture fall within AIG’s claims?

This article is not legal advice. Patent infringement analysis requires qualified IP counsel evaluating specific system architectures against specific patent claims. But understanding what AIG’s patents actually cover, and just as importantly what they do not cover, is essential context for any insurance technology leader making build-or-buy decisions about AI underwriting infrastructure in 2026.

How Patent Infringement Analysis Works

Before examining AIG’s specific claims, it helps to understand the basic framework.

Patent infringement requires that someone practices every element of at least one granted claim. The claims are the numbered paragraphs at the end of the patent document. They are the legally enforceable boundaries of the patent, not the abstract, not the detailed description, not the diagrams. The description and diagrams provide context for interpreting the claims, but the claims themselves define what is protected.

Each independent claim describes a specific combination of steps or components. If an accused system performs every step described in a claim, it potentially infringes. If it omits even one required element, it does not infringe that particular claim. This is called the “all elements” rule.

Patents typically contain both independent claims (which stand alone) and dependent claims (which add further limitations to an independent claim). Dependent claims are narrower. If an accused system doesn’t infringe the broader independent claim, it cannot infringe the dependent claims that build on it.

With that framework in mind, here is what AIG’s three patents actually require.

What Each Patent Requires (And What It Doesn’t)

Patent #12,437,155: The Core Extraction Architecture

Full analysis

Claim 1 requires all of these elements together:

• Receiving OCR output that includes tables represented in a markdown language
• Separating that output into a table portion and a text portion using the markdown language
• Forming independent table chunks and text chunks through a chunking methodology
• Identifying relevant chunks based on search criteria related to an LLM prompt
• Storing the LLM’s response to the prompt and relevant chunks
• Generating vector embeddings for the table chunks and storing them in a retrieval index

A system likely falls outside this claim if it:

• Uses a method other than markdown-based separation to distinguish tables from text (for example, using bounding box coordinates from a vision model, or using a proprietary document understanding model that identifies tables without converting to markdown first)
• Does not create independent chunking pipelines for tables versus text (for example, if it chunks the entire document uniformly regardless of content type)
• Does not use a RAG architecture with vector embeddings for retrieval (for example, if it processes entire documents through the LLM without a retrieval step, or uses a fine-tuned model that processes documents directly)
• Uses a completely different indexing approach (for example, knowledge graph-based retrieval rather than vector embedding similarity)

Patent #12,437,154: The Traceability Layer

Full analysis

Claim 1 requires all of these elements together:

• Generating chunks from a document
• Associating document identifiers and page identifiers (or chunk identifiers) with each chunk
• Identifying relevant chunks based on search criteria
• Transmitting a prompt to the LLM that includes both the extraction request AND a separate request for the LLM to identify which chunks it used
• Storing the extracted information with the identifiers of the chunks the LLM reported using

A system likely falls outside this claim if it:

• Does not ask the LLM to self-report which source chunks it relied on (this is the most distinctive element of the claim; many RAG systems track which chunks were retrieved but do not ask the LLM to report which ones it actually used in generating its response)
• Does not store the LLM-reported chunk attributions alongside the extracted data
• Uses a traceability approach that tracks retrieved chunks at the system level but does not incorporate LLM self-reporting into the prompt itself

Patent #12,511,320: The Spreadsheet Processing Method

Full analysis

Claim 1 requires all of these elements together:

• Receiving a document with multiple tables
• Generating prompts that include requests to identify individual tables, extract metadata, and reconstruct tables in a markdown language
• Generating table chunks from the LLM-reconstructed tables
• Generating vector embeddings for those chunks and storing them in a retrieval index
• Identifying relevant table chunks for an extraction prompt
• Storing the LLM’s response

Claim 7 narrows this further to a specific four-step chain-of-thought structure:

• Step 1: Quantify the number of tables
• Step 2: Identify the tables, with explicit reference to the count from Step 1
• Step 3: Extract metadata
• Step 4: Reconstruct tables using the metadata from Step 3

A system likely falls outside these claims if it:

• Processes spreadsheets without using an LLM for table identification (for example, using heuristic or rule-based methods to detect table boundaries, even if an LLM is used later for extraction)
• Does not reconstruct tables in markdown before chunking (for example, if it converts spreadsheet data directly to JSON or another structured format)
• Uses a single-step prompt rather than a multi-step reasoning approach for table identification
• Does not generate vector embeddings of the reconstructed table chunks (for example, if it uses the tables directly without a retrieval step)

What AIG Has NOT Patented

This is equally important. AIG’s patents are specific architectural methods, not broad concepts. The following are not constrained by these patents:

Using AI for underwriting in general. No patent can claim ownership of the concept of applying artificial intelligence to insurance underwriting. These patents cover specific document processing methods, not the broader application.

Using RAG architectures for document extraction in general. RAG is a well-established pattern in AI application development. AIG’s claims are specific to the combination of markdown-based table/text separation, independent chunking pipelines, and the associated retrieval strategies. Carriers using RAG with different ingestion and chunking approaches are not constrained.

Using LLMs to read and extract data from insurance documents in general. A carrier that processes submissions by sending entire documents to an LLM without a RAG framework, or that uses a fine-tuned extraction model rather than a prompt-based approach with retrieved chunks, is operating outside these patents.

Using Palantir, Anthropic’s Claude, or any other specific technology. The patents describe methods that happen to use these tools, but the claims are not limited to specific vendors. Conversely, using the same vendors does not automatically mean you are practicing the patented methods.

Building an ontological data store for underwriting data. The concept of an insurance data ontology predates these patents. What is claimed is the specific pipeline for populating an ontological data store through the patented extraction methodology.

The Vendor Dimension

Many carriers are not building custom AI document extraction systems from scratch. They are deploying vendor platforms such as Indico Data, Hyperscience, Eigen Technologies, or building on open-source frameworks like LangChain, LlamaIndex, or Haystack. This adds a layer to the analysis.

If a vendor’s platform happens to implement an architecture that falls within AIG’s claims, both the vendor and the carrier using the platform could face infringement questions. In practice, enterprise software vendors typically provide indemnification clauses in their contracts that cover patent infringement claims. Carriers evaluating vendor platforms for AI document processing should confirm that their contract includes IP indemnification and should ask vendors directly whether they have evaluated their architecture against AIG’s granted patents.

For carriers building on open-source frameworks, the IP responsibility falls on the builder. LangChain and LlamaIndex provide the building blocks for RAG systems, but the specific architectural choices (how you chunk documents, whether you separate tables from text, how you handle traceability) are made by the implementation team. Those choices are what determine whether the resulting system falls within AIG’s claims.

The Broader IP Landscape

AIG is not the only insurer building an AI patent portfolio, though it appears to be among the first to patent production-level AI underwriting methods in the U.S.

Globally, Ping An Insurance Group holds over 100 generative AI patents, making it one of the most prolific AI patent filers in any industry. However, Ping An’s portfolio is concentrated in the Chinese patent system and covers a broader range of applications beyond underwriting, including healthcare diagnostics, financial fraud detection, and customer service automation.

In the U.S., insurers have historically filed relatively few technology patents compared to banks and fintech companies. AIG’s three-patent filing signals a potential shift. As generative AI moves from experimentation to core infrastructure across the industry, other carriers may begin seeking patent protection for their own novel methods. This could create an increasingly complex IP landscape for insurance AI over the next several years.

EXL, a major insurance services and analytics provider, was recently granted 10 AI-related patents in a single year, including patents covering an insurance-specific fine-tuned LLM and knowledge graph construction for claims processing. While EXL operates as a vendor rather than a carrier, its patent activity indicates that AI intellectual property is becoming a competitive dimension across the insurance value chain, not just among carriers.

Practical Recommendations

For carrier CTOs and technology leaders: Conduct a freedom-to-operate review. If your organization is building or has built an AI document processing system for underwriting, have IP counsel review your architecture against AIG’s three granted patents. This is standard practice when a competitor holds patents in your technology space and costs far less than defending against an infringement claim. Focus the review on the specific elements: markdown-based table/text separation, independent chunking pipelines, LLM self-reported source attribution, and chain-of-thought multi-table processing.

For actuaries overseeing AI implementations: The traceability architecture described in Patent #12,437,154 represents a high-water mark for AI auditability in insurance. Whether or not your organization’s system falls within AIG’s claims, the design principles (chunk-level source attribution, LLM self-reported citation, version tracking, response validation) are worth adopting as governance best practices. These capabilities align with what ASOP No. 56 and the NAIC Model Bulletin are increasingly requiring.

For insurtech founders and vendors: If your product processes insurance documents using RAG-based LLM extraction, these patents are directly relevant to your product architecture. Evaluate whether your table/text handling, chunking methodology, and traceability features overlap with AIG’s claims. Build this evaluation into your product development process now, before a carrier customer or their legal team raises the question.

For enterprise risk management: AI patent risk is emerging as a category within technology and operational risk frameworks. As more carriers patent their AI methods, the potential for cross-industry infringement disputes grows. ERM teams should work with legal counsel to monitor patent activity in insurance AI and assess exposure as part of their technology risk assessments.

A Snapshot in Time: How These Patents Age

Patents last 20 years. Technology cycles in AI are measured in months. That tension is worth examining honestly, because it shapes how much weight carriers should give these patents in their strategic planning.

The claims in AIG’s patents protect specific architectural patterns that reflect the state of the art in LLM-based document processing as of early 2025. Several technology trends are already shifting the ground beneath those patterns.

Multi-modal models are reducing the need for OCR-to-markdown conversion. The first patent’s core innovation is separating tables from text using markdown language patterns generated by OCR. But multi-modal models that can look directly at a PDF image and understand table structure, column headers, and cell relationships without first converting to text are rapidly improving. If a carrier uses a vision model to process documents natively rather than running them through an OCR-to-markdown pipeline, the entire first step of Patent #12,437,155 may not apply. AIG’s third patent acknowledges this direction by including a multi-modal language model fallback path, but the claims still center on markdown-based processing.

Context windows are making RAG optional rather than mandatory. When AIG filed these patents, feeding an entire 200-page submission package to an LLM in a single pass was not feasible. Chunking, indexing, and retrieval augmentation were necessary because models could not hold that much context. As context windows expand to millions of tokens, some carriers may find it simpler to process entire documents without a retrieval step at all. A system that sends a full document to a long-context model and extracts data without chunking, embedding, or retrieval would likely fall outside these claims entirely.

Agentic architectures are evolving beyond retrieve-then-extract. The industry is moving toward systems where AI agents autonomously decide how to process documents, which tools to invoke, and how to validate results. These emergent architectures may look structurally different from the pipeline AIG patented, even if they accomplish similar outcomes.

The strategic value shifts over time. Right now, these patents protect AIG’s current production architecture and signal competitive seriousness to the market. In two years, their primary value may be defensive, preventing competitors from copying AIG’s specific current approach, rather than offensive, blocking the only viable path to AI-assisted document processing. That is still valuable, but it is a different kind of value.

The traceability patent ages best. Of the three patents, #12,437,154 has the longest shelf life. Regardless of how document extraction technology evolves, the regulatory requirement for auditability in insurance is not going away. The NAIC Model Bulletin, state-level AI governance frameworks, and ASOP No. 56 all point toward increasing demands for demonstrable traceability in AI-assisted decisions. The concept of requiring an AI system to cite its sources, storing those citations alongside extracted data, and maintaining version-tracked audit trails will remain relevant even as the underlying extraction technology changes completely. The specific implementation may become one option among many, but the principles have staying power that transcends any single technology generation.

The honest read for carriers is this: these patents are a meaningful constraint on the current generation of RAG-based document extraction approaches. They will become less constraining as the technology evolves beyond RAG toward long-context processing, native multi-modal understanding, and fully agentic architectures. But they will not become irrelevant, because some carriers will continue using RAG-based approaches for years, the patents remain enforceable against anyone who does, and the traceability principles they describe are becoming regulatory expectations independent of any specific technology.