Inside AIG’s Agentic AI Underwriting Machine: How Palantir, Claude, and 4 Million Data Points Are Rewriting Commercial Insurance

The insurance industry’s conversation about artificial intelligence has largely been a conversation about efficiency. Faster claims processing, lower administrative costs, automated call center triage. These are valuable outcomes, but they represent a narrow view of what AI can actually accomplish in a business built on risk selection.

A more consequential question is whether AI can expand what insurers are able to underwrite in the first place. Not just processing existing submissions faster, but evaluating more opportunities, with greater consistency, at higher data accuracy, across risk categories that were previously too fragmented or data-poor to assess at scale.

As Tracie Thompson, a former AIG and Aon executive now at insurtech Cytora, recently argued in Insurance Innovation Reporter, AI’s most consequential potential in insurance lies not in cost reduction but in expanding underwriting capacity and making previously uninsurable risks measurable. McKinsey has estimated that generative AI could unlock between $50 billion and $70 billion in additional insurance industry revenue, with underwriting, marketing, and software engineering dimensions showing the highest impact.

That framing matches what we have observed across the E&S market in particular. The excess and surplus lines sector thrives on complexity: risks that are too unusual, too data-fragmented, or too volatile for the standard admitted market. Historically, the constraint on E&S growth has not been risk appetite but underwriting bandwidth. If a carrier receives 500,000 submissions annually but can only thoroughly evaluate a fraction of them, viable risks get overlooked simply because there are not enough hours in the day.

No carrier has bet more visibly on this growth thesis than AIG. Over the past 18 months, the insurer has built what it describes as an “agentic AI ecosystem” through partnerships with Anthropic (the maker of Claude) and Palantir Technologies, deploying generative AI and large language models not on the fringes of its operations but directly into core underwriting and claims workflows across its commercial lines business. The results so far are measurable, specific, and materially different from the vague pilot programs that characterize much of the industry’s AI narrative.

What makes AIG’s approach worth studying in detail is the scope and specificity of its public disclosures. Across four consecutive quarterly earnings calls, a four-hour Investor Day (which featured the CEOs of both Anthropic and Palantir on stage), multiple press releases, and a TIME Magazine profile, AIG has laid out its AI architecture, its workflow mechanics, and its performance metrics with a level of transparency that is rare in the industry. For actuaries, this level of disclosure provides a concrete case study for understanding how AI is actually being deployed in commercial insurance, not in theory, but in production.

This article breaks down exactly how AIG’s AI underwriting infrastructure works, what it has accomplished, and why actuaries across the industry should pay close attention.

The Three-Layer Technology Stack

From tracking AIG’s public disclosures across four consecutive earnings calls, its March 2025 Investor Day, and multiple press releases, a clear architectural picture emerges. AIG’s AI underwriting system operates across three interconnected layers, each handled by a different technology partner or internal team.

Layer 1: Anthropic’s Claude (The Language Intelligence)

At the foundation of AIG’s AI capabilities sits Anthropic’s Claude, a large language model that serves as the primary engine for understanding, extracting, and reasoning over unstructured documents. AIG has been working with Anthropic since the early Claude 2.1 era and, according to statements on its Q4 2025 earnings call (February 11, 2026), has progressed through multiple model generations to Claude 4.6.

Anthropic CEO Dario Amodei, who appeared alongside AIG CEO Peter Zaffino at the 2025 Investor Day, noted that the real differentiator at AIG was not the model itself but the conviction to deploy it into core business processes. He observed on the Nexus Connect summary of the event that AIG moved past pilots and aspirational goals to bet with conviction on underwriting and claims use cases. Anthropic’s subsequent Financial Analysis Solution launch in July 2025 specifically highlighted AIG as a marquee deployment, citing the insurer’s 5x compression in review timelines and accuracy improvements from approximately 75% to over 90%.

Claude’s role in AIG’s stack centers on several capabilities: reading and interpreting the heterogeneous documents that arrive in broker submissions (applications, loss runs, schedules, supplemental questionnaires, financial statements), extracting structured data fields from unstructured text, and providing reasoning that supports underwriting decision-making.

Layer 2: Palantir Foundry (The Ontology and Data Infrastructure)

If Claude provides the language understanding, Palantir’s Foundry platform provides the structural intelligence, the ontology layer that maps relationships across AIG’s business data.

AIG CEO Peter Zaffino described the company’s ambition during an August 2025 earnings call as creating a “digital twin” of the entire business, representing all key data, processes, business logic, and relationships across businesses and functions. The ontology, which AIG has been building since it began its AI work, serves as the foundation for technical reviews, claims personnel insight, and underwriting decision-making.

In practical terms, Palantir’s Foundry creates what the industry calls an ontology: a structured representation of the concepts, entities, and relationships within a domain that allows AI models to understand context rather than just process text. AIG has developed an ontology that enables large language models to access more than four million industry data points, connecting underwriting guidelines, historical loss data, risk characteristics, pricing benchmarks, and portfolio composition into a queryable knowledge graph.

Palantir CEO Alex Karp, who also appeared at AIG’s 2025 Investor Day, praised Zaffino for starting to build the AI ecosystem with what he called a singular and prescient focus on revolutionizing core business activities rather than experimenting at the margins.

AIG’s patent filings reveal exactly how data flows into that ontology. The patents describe a system that extracts structured information from unstructured broker submissions and populates what the patent text calls an “ontological data store” - language that maps directly to the Palantir Foundry integration AIG has described publicly. For a detailed technical walkthrough, see our analysis of Patent #12,437,155.

Layer 3: AIG’s Proprietary Tools (AIG Assist and Auto Extract)

The third layer consists of AIG’s internally developed tools that sit between the LLM capabilities and the underwriter’s workflow.

AIG Assist (formally “Underwriting by AIG Assist”) is the company’s primary generative AI underwriting tool. According to Coverager’s analysis of the tool’s rollout, AIG Assist debuted in late 2024 inside the North America Financial Lines unit and automates the first pass of underwriting by extracting data from broker submissions, structuring unstructured documents, and surfacing key risk information to underwriters. By Q3 2025, AIG was processing 100% of applicable private and nonprofit submissions through the tool. It has since been rolled out to seven additional lines of business, including Lexington Insurance Company, AIG’s primary E&S carrier.

Auto Extract is a patent-pending capability that AIG developed to address one of the most persistent challenges in AI-assisted underwriting: building an accurate, structured data source from a messy population of documents. As Zaffino explained on the Q3 2025 earnings call, Auto Extract uses large language models to pull specific structured information from unstructured text, including documents in multiple formats, websites, and even conversations. This capability essentially solves the data ingestion bottleneck that prevents many insurers from scaling their AI initiatives beyond isolated pilots.

That patent has since been granted. In fact, AIG now holds three granted U.S. patents covering its AI underwriting document processing system, all invented by Lei Zhang and Christopher Cirelli at AIG’s Atlanta Innovation Hub. Patent #12,437,155 describes the core Auto Extract architecture in technical detail: how the system separates tables from text using markdown patterns, creates independent retrieval pipelines for each data type, and populates an ontological data store through LLM-based extraction.

The Underwriting Workflow: From Inbox to Decision

Combining information from AIG’s earnings calls, investor day presentations, and industry reporting, the AI-assisted underwriting workflow at AIG follows a clear progression.

Step 1: Submission Intake and Data Ingestion. Broker submissions arrive in varied formats: PDFs, spreadsheets, scanned documents, emails with attachments, supplemental questionnaires in different layouts. Auto Extract processes these documents using LLMs to identify and pull structured fields (named insured, coverage requested, limit/attachment, loss history, revenue, SIC codes, and more) regardless of the document format. This step alone eliminates what historically consumed a significant portion of underwriter time.

Step 2: Ontology Matching and Risk Classification. The extracted data feeds into Palantir’s Foundry ontology, where it is matched against AIG’s risk appetite parameters, historical underwriting data, and portfolio composition. The system classifies submissions by risk characteristics, identifies gaps in the data (missing loss runs, incomplete applications), and flags submissions that fall outside appetite guidelines.

Step 3: Prioritization and Triage. AIG Assist presents underwriters with a prioritized queue. When underwriters arrive at their desk, as the Carrier Management coverage of Investor Day described, all submissions have been ingested, reviewed, and prioritized. The system can assess propensity to bind, helping underwriters focus on the opportunities most likely to convert to written premium rather than spending equal time on every submission.

Step 4: Underwriter Decision Support. The underwriter reviews the AI-prepared file, which includes structured data, highlighted risk factors, comparable historical accounts, and any flagged concerns. The system does not make binding decisions autonomously. As Zaffino emphasized in his TIME interview (July 2025), the underwriter remains at the center of decision-making, with AI designed to deliver better outcomes and drive operating leverage while keeping experienced underwriters at the core of the process.

Step 5: Portfolio Analytics and Feedback Loop. Every underwriting decision feeds back into the ontology, refining the digital twin and improving future triage and prioritization. AIG’s ontology creates a clear record of any actions taken, which informs business logic and provides the ability to audit agent activities, as Zaffino told Insurance Journal in August 2025. This feedback mechanism is particularly relevant for actuarial pricing: as the ontology accumulates more decision data, it enables increasingly granular analysis of which risk characteristics correlate with successful binds, favorable loss experience, and portfolio-level profitability.

AIG has patented the specific technical mechanisms behind this auditability. U.S. Patent #12,437,154 describes a system that assigns unique identifiers to every document chunk, requires the LLM to report which chunks it used for each extraction, maintains version histories and timestamps, and generates citation lists for regulatory compliance. The patent explicitly references “regulated industries” and the need to demonstrate that data extraction is accurate and unbiased.

The Claims Extension. AIG has also been piloting “Claims by AIG Assist,” which applies similar capabilities to the claims lifecycle. According to Zaffino’s Q3 2025 remarks reported by Reinsurance News, the tool is reducing the time required for claims teams to receive first notice of loss reports and to issue coverage letters. While underwriting has been the primary focus, the claims application matters for actuaries because faster, more consistent claims triage can produce better data for loss development analysis and reserve adequacy studies.

The Back-Office Layer. The AI ecosystem extends beyond underwriting and claims. AIG’s relationship with outsourcing partner Accenture, discussed on the Q4 2025 earnings call, involves shared design of agent-based LLMs for back-office processes, with AIG sharing in both the orchestration design and the resulting savings. This suggests a comprehensive approach to AI deployment across the entire value chain, not a narrow pilot confined to a single business function.

The Numbers: Measuring What Actually Changed

AIG has disclosed specific performance metrics across multiple public forums, making this one of the most quantitatively documented AI deployments in the insurance industry.

Review speed: Underwriting review timelines have compressed by more than 5x in early rollouts, according to statements by Zaffino cited by Anthropic’s financial services partners.

Data accuracy: Accuracy rates within underwriting processes improved from approximately 75% to upward of 90%, as Zaffino first disclosed on AIG’s Q3 2024 earnings call.

Submission volume at Lexington: After deploying AIG Assist, Lexington Insurance recorded a 26% increase in submission counts year over year, with a 35% improvement in the submit-to-bind ratio for middle market property, as reported by Coverager.

Total submissions processed: AIG has already reached more than 370,000 submissions, putting it on track toward the 500,000-submission target set at Investor Day with a 2030 horizon. The acceleration is ahead of schedule. Zaffino remarked on the Q4 2025 call that the opportunity is greater than what he thought at Investor Day.

Financial performance context: In Q4 2025, AIG delivered underwriting income of $670 million, up 48% year over year, with an accident year combined ratio of 88.9%, marking its seventeenth consecutive quarter below 90%. While these results reflect many factors beyond AI, AIG’s management has consistently drawn a direct connection between AI deployment and its ability to grow premium without proportional headcount increases.

100% coverage in Financial Lines: For private and not-for-profit financial lines business, AIG now processes 100% of applicable submissions through its AI-assisted workflow, without adding underwriters, as CDO Claude Wade confirmed at Investor Day.

AIG’s patent filings reveal the specific technical mechanisms behind these improvements. The accuracy gain from 75% to over 90% reflects three innovations described in the patents: separating tabular data from document text so the LLM can process each independently (Patent #12,437,155), building traceability and error detection into the extraction workflow to catch and correct failures (Patent #12,437,154), and using chain-of-thought prompting to reliably process complex multi-table spreadsheets including automated financial unit conversion (Patent #12,511,320).

Beyond Efficiency: AI as a Capital Deployment Engine

Perhaps the most forward-looking aspect of AIG’s AI strategy is how the technology stack is enabling entirely new business structures, not just improving existing ones.

Lloyd’s Syndicate 2479

In December 2025, AIG partnered with Amwins (a major specialty insurance distributor) and funds managed by Blackstone to launch Syndicate 2479 at Lloyd’s of London. This was not a traditional syndicate formation. AIG used Palantir’s Foundry and multiple LLM agents to analyze a broad cross-section of Amwins’ roughly $6 billion in delegated authority premiums, build an ontology of the portfolio, identify risk characteristics aligned with the syndicate’s appetite, and construct a balanced book across lines of business.

The syndicate commenced underwriting $300 million in premium on January 1, 2026, managed by AIG’s Talbot Underwriting. It marked the first time AIG deployed its generative AI capabilities within a third-party capital structure. Zaffino described it as representing the next level of innovation and technical modeling for portfolio underwriting.

McGill and Partners Collaboration

In March 2026, AIG and McGill and Partners announced a strategic collaboration to deploy capacity across up to $1.6 billion in specialty gross premiums written through McGill’s digital broking platform. AIG again worked with Palantir to build an ontology of McGill’s portfolio, enabling near real-time exposure analysis, limit deployment tracking, modeled risk outputs, and loss information.

What makes this deal structurally significant is its application to the subscription market, where large commercial risks are shared among multiple insurers. Lead underwriters price risks and set terms; follow underwriters then decide whether to participate. Traditionally, assembling follow capacity was slow and labor-intensive. AIG’s agentic AI framework can manage follow-capacity decisions in near real time, potentially transforming how the subscription market operates.

The Everest Portfolio Integration

AIG also used its AI capabilities during the integration of renewal rights from Everest’s global retail portfolio. The company built an Everest ontology, essentially a digital twin of the acquired portfolio, and coupled it with AIG’s existing business ontology. This allowed teams to evaluate limits, pricing, attachment points, and conversion strategies, prioritizing accounts for renewal in what Zaffino described as a fraction of the time compared to traditional book-roll processes.

For actuaries, this application is worth noting because portfolio acquisitions and book rolls have historically been resource-intensive, time-sensitive exercises where incomplete data leads to mispriced renewals and higher-than-expected lapse rates. An AI system that can rapidly build a structured representation of an acquired portfolio and match it against the acquiring insurer’s own underwriting data creates genuine strategic advantage in M&A and book transfer scenarios.

The Agent Taxonomy: Knowledge, Adviser, and Critic

AIG’s most forward-looking initiative for 2026, as described on the Q4 2025 earnings call and analyzed by AI News and Coverager, is the development of an orchestration layer to coordinate multiple AI agents across the enterprise.

The company has described three distinct agent types:

Knowledge assistants provide real-time information retrieval, pulling relevant data from AIG’s ontology, policy databases, regulatory filings, and external sources to ensure underwriters and claims professionals have comprehensive context for every decision.

Adviser agents generate insights based on historical use cases. When an underwriter reviews a submission, the adviser agent can surface comparable historical accounts, identify patterns in loss experience for similar risks, and recommend pricing or terms based on what has worked (or failed) in the past.

Critic agents challenge recommendations and decisions. This adversarial layer is particularly notable from an actuarial and governance perspective. Rather than simply confirming AI-generated suggestions, critic agents probe for weaknesses, flag potential concerns, and ensure that decisions are stress-tested before they reach the underwriter.

The critic agent concept has a concrete technical implementation described in AIG’s patent filings. Patent #12,437,154 details a “response validator” component that checks every LLM extraction against expected parameters and a chunk-level verification system that can detect hallucinations by comparing the LLM’s cited sources against the actual content of those sources. This is not post-hoc review; it is structural verification built into every extraction the system performs.

The orchestration layer determines when agents are activated, how they share information, and when human oversight is required. Zaffino has emphasized that the focus for 2026 is shifting from isolated use cases to enterprise scale, noting that orchestrating multiple agents in an orderly way to achieve scale across the enterprise is the primary focus for the year ahead. This effort spans front-office underwriting, mid-office operations, and back-office functions, including work with outsourcing partner Accenture on LLM-based agent deployment.

The Actuarial Implications

AIG’s AI deployment carries significant implications for actuaries across the industry, whether they work in pricing, reserving, enterprise risk management, or consulting.

Pricing and risk selection. When an insurer can process 5x more submissions at 90%+ accuracy without adding headcount, the economics of risk selection change fundamentally. Actuaries building pricing models must consider that AI-assisted underwriting can evaluate a broader universe of risks, potentially improving portfolio diversification while maintaining (or tightening) risk appetite discipline. The improved submit-to-bind ratio at Lexington suggests that AI triage helps underwriters focus on higher-quality opportunities.

Reserving and loss development. Faster, more consistent underwriting may produce cleaner, more homogeneous books of business over time. Reserving actuaries should monitor whether AI-assisted portfolios show different loss development patterns compared to traditionally underwritten business. The ontology’s ability to track every underwriting decision creates a data trail that could enable more granular reserve analysis.

Model governance and validation. AIG’s agent taxonomy, particularly the critic agent concept, offers a template for AI governance that aligns with the principles of ASOP No. 56 (Modeling). The ontology’s capacity to maintain an auditable record of agent actions addresses one of the profession’s key concerns about AI transparency. As we explored in our analysis of The AI Governance Gap in Actuarial Practice, the gap between AI deployment speed and standards development remains significant, but AIG’s approach offers one model for how governance can be built into the architecture itself rather than layered on as an afterthought.

Competitive dynamics. If AIG’s projections hold, the company will be evaluating 500,000+ E&S submissions annually by 2030 and booking $4 billion or more in new business premiums from that pipeline. Competing carriers without comparable AI infrastructure may find themselves at a structural disadvantage in the E&S market, where speed and data quality increasingly determine which submissions get written. Actuaries advising carrier strategy should be modeling the competitive impact of AI-enabled underwriting capacity, not just AI-driven expense savings.

The subscription market transformation. The McGill and Partners collaboration suggests that AI-managed follow capacity could reshape how actuaries think about portfolio construction in subscription markets. If AI agents can manage real-time capacity deployment across a $1.6 billion portfolio, the traditional distinction between lead and follow underwriting begins to blur. Actuaries involved in specialty reinsurance and Lloyd’s syndicate business should track these developments closely.

Career implications for actuaries. AIG’s deployment does not eliminate actuarial roles; if anything, it increases demand for actuaries who can work at the intersection of data science and insurance. The ontology layer requires people who understand risk classification, loss development patterns, and portfolio construction at a deep enough level to validate what the AI is doing. The critic agent concept is essentially actuarial judgment embedded in software. Actuaries who can contribute to the design, calibration, and validation of these systems will be more valuable, not less. However, actuaries whose primary contribution is data aggregation and manual processing face genuine displacement risk, because that is precisely what Auto Extract and AIG Assist are designed to automate.

Broader industry adoption context. AIG is not operating in isolation. Travelers has deployed Anthropic’s Claude to over 20,000 employees for innovation and claims support. Liberty Mutual, Zurich, and other major carriers have disclosed their own AI strategies. But AIG’s approach stands out for its integration depth: rather than deploying AI as a standalone tool for specific tasks, AIG has built an interconnected stack where data flows from document intake through ontology classification to agent-assisted decision-making and back into portfolio analytics. That end-to-end architecture, and the measurable performance data backing it, makes AIG the most instructive case study for actuaries trying to understand where the industry is headed.

Patent Deep Dives

AIG holds three granted U.S. patents covering its AI underwriting document processing system. Together, they provide the most technically detailed public description of how any major insurer is using large language models in production underwriting. Each article below analyzes one patent in depth, mapping the technical claims to AIG’s public disclosures and drawing out the implications for actuarial practice.

• Patent #12,437,155: How AIG Separates Tables from Text to Feed Its AI Underwriting Engine - The foundational architecture for independent tabular and textual document processing using retrieval-augmented generation.
• Patent #12,437,154: The Traceability and Hallucination Detection Layer - Chunk-level source attribution, LLM output verification, version history tracking, and automated citation generation for regulatory compliance.
• Patent #12,511,320: Chain-of-Thought Prompting for Complex Financial Spreadsheets - A step-by-step reasoning methodology for processing multi-table spreadsheets, with financial unit conversion and a hybrid text/vision processing architecture.