Artificial Intelligence at AIG

Headquartered in New York, AIG is a global insurer providing commercial and personal property-casualty coverage across Liability, Financial Lines, Property, Global Specialty, Crop Risk Services, Personal Lines, and Accident & Health. As of December 31, 2025, the company employed approximately 22,100 people across 45 countries and underwrites business in more than 200 countries and jurisdictions through its global network.​

AIG reported $3.1 billion in net income for full-year 2025. General Insurance recorded $23.7 billion in net premiums written, while its combined ratio improved by 1.7 percentage points year over year, from 91.8% to 90.1%. Global Commercial net premiums written increased 4% on a reported basis to $17.4 billion, supported by 9% growth in new business.

AIG has identified AI deployment across underwriting and claims as a key strategic focus, stating in its 2025 Annual Report that it is deploying and scaling agentic AI solutions to speed processes and improve decision-making in these areas. In 2025, the company significantly advanced this strategy through expanded partnerships with Palantir, Anthropic, AWS, and Google, embedding AI capabilities across underwriting and claims.

This article examines two use cases illustrating how AIG is applying this investment within its operations:

• Accelerating Underwriting Submission Triage through AIG Assist — AIG uses large language models to read, prioritize, and summarize incoming submissions, enabling underwriters to evaluate more opportunities without compromising underwriting discipline.
• Speeding Capacity Deployment through Portfolio Ontologies with Palantir — AIG uses machine-readable representations of its insurance portfolios to enable AI agents to evaluate risk and support faster capital allocation decisions, with potential applications across its operations and third-party capital and distribution partnerships.

We begin with AIG Assist and the application of generative AI to underwriting submission triage.

Accelerating Underwriting Submission Triage through AIG Assist

Complex commercial underwriting creates a bottleneck that additional headcount alone cannot eliminate. AIG’s investor materials indicate that manually reviewing a complex commercial submission can take three to four weeks. They also reveal that Lexington, AIG’s excess and surplus lines business, bound policies for only about 2% of the roughly 300,000 new-business submissions it received in 2024. Hiring more underwriters may increase review capacity, but it does not fundamentally reduce the time required to assess each submission or identify the opportunities most likely to warrant coverage.

Video: AIG Underwriter Assistance in Action | AIG at AIPCon 7 (Source: Palantir)

A commercial insurance submission typically arrives as an unstructured collection of broker cover letters, statements of value, loss runs, and supplemental applications, all in inconsistent formats. Underwriters have traditionally had to read and reconcile these documents manually before deciding whether — and on what terms — to quote the risk. AIG’s investor materials illustrate the resulting funnel at Lexington, its excess and surplus lines business: approximately 300,000 new-business submissions in 2024 resulted in about 6,700 bound policies, equivalent to a bind rate of roughly 2% and approximately $1 billion in new-business premium.

AIG’s 2030 ambition for Lexington — 500,000 submissions, a 6% bind rate, and $4 billion in new-business premium — requires the company both to process more submissions and convert a greater proportion of them. Expanding the underwriting team could increase review capacity, but it would not fundamentally change the time required to assess each submission. AIG developed AIG Assist to address that constraint.

Screenshot: AIG’s illustration of how AIG Assist synthesizes, prioritizes, and prepares submission data for underwriter review (Source: AIG Investor Day 2025 presentation, March 31, 2025, PDF p. 117)

​AIG Assist includes a patent-pending capability called Auto Extract, which uses large language models to extract structured data from unstructured submission documents across varied formats. The tool then assesses incoming submissions against the relevant business’s stated risk appetite, prioritizes them for review, and produces curated summaries for underwriters. This changes the initial triage process from reviewing submissions in the order received to focusing on opportunities that appear to fit the insurer’s underwriting criteria.​

In 2026, AIG described the next phase of the technology being developed with Palantir and Anthropic, as a multi-agent underwriting system. On AIG’s Q1 2026 earnings call, CEO Peter Zaffino outlined a proposed architecture in which purpose-built agents could perform submission ingestion and data extraction, evaluate risks against underwriting guidelines, benchmark pricing against portfolio targets, and synthesize their findings for an underwriter.​

AIG is also developing an orchestration layer to coordinate these agents and their handoffs. The company expects the system to supplement underwriters’ analysis while maintaining human oversight of risk assessment, pricing, and coverage decisions. AIG characterized this multi-agent architecture as a system under development rather than a fully deployed production capability.​

AIG Assist’s deployment has expanded in stages rather than launching everywhere at once:​

• AIG Assist entered production in Financial Lines, including its Private Not-for-Profit business, where AIG later reported that the system was reviewing 100% of applicable submissions.
• AIG subsequently began deploying the tool across Lexington’s middle-market property and casualty operations. The company said in November 2025 that it expected to complete deployment across Lexington’s remaining wholesale business by year-end.
• AIG also accelerated its planned deployment across North American, U.K., and EMEA commercial lines by six months.
• On the claims side, AIG has piloted similar document-extraction capabilities intended to shorten the period between receiving a first notice of loss and issuing a coverage letter.

For underwriters, the operational change is a shift from a review queue constrained by the time required to read and reconcile documents to one in which extraction, summarization, and an initial assessment can occur before human review. AIG describes the resulting change to an underwriter’s day-to-day work along several lines:​

• Instead of beginning with an unsorted collection of documents, an underwriter receives a prioritized submission with extracted data and a machine-generated summary.
• Underwriting time can shift away from data entry and document reconciliation and toward risk selection, pricing, coverage, and policy-structure decisions.
• The system is designed to support rather than replace underwriting judgment, with underwriters remaining responsible for consequential risk and coverage decisions.
• On the claims side, a similar approach could accelerate the initial document review between the reporting of a loss and the issuance of a coverage letter.

AIG has supported its broader digital transformation with substantial investment. The company reported in 2024 that it had invested approximately $300 million in data, digital workflows, AI, and talent over the preceding two years, following more than $1 billion in foundational data technology investment over five years. These figures cover AIG’s wider technology program rather than AIG Assist alone.

​In an early AIG Assist deployment, AIG’s investor materials state that submission turnaround fell from three to four weeks to less than one day. The proportion of applicable submissions reviewed increased to 100%, rather than a filtered subset. Zaffino also reported that data-extraction accuracy improved from approximately 75% to more than 90%, alongside a substantial reduction in processing time.

​AIG continued to report progress through subsequent quarters. By year-end 2025, Lexington had received more than 370,000 submissions — up 26% year over year and representing substantial progress toward, but not completion of, its 2030 ambition of 500,000. AIG also reported a 35% improvement in the submit-to-bind ratio for Lexington’s middle-market property business following the rollout of AIG Assist.​

On AIG’s Q1 2026 earnings call, Zaffino said the tool had helped Lexington’s middle-market property business quote 30% more submissions, reduce underwriters’ time to quote by 55%, and increase the number of submissions bound by approximately 40%. These figures suggest that the system is affecting both processing capacity and commercial conversion. However, AIG has not disclosed the underlying volumes, measurement periods, or contribution of other operational changes in enough detail to isolate the effect of AIG Assist.​

AIG reported these performance figures in earnings calls and investor materials, and they have not been independently validated. The 2030 submission, bind-rate, and premium figures also remain ambitions rather than achieved results. Nevertheless, the disclosures show a clear deployment pattern: AIG introduced the technology in selected underwriting businesses, extended it across additional lines, and reported improvements in review speed, submission coverage, and conversion as deployment progressed.

Speeding Capacity Deployment through Portfolio Ontologies with Palantir

AIG’s second AI use case addresses a broader problem than accelerating the review of individual submissions: creating a structured, queryable representation of an entire portfolio so that AI systems can evaluate exposures and support underwriting and capacity-deployment decisions.

​AIG refers to this representation as an ontology. Built on Palantir’s Foundry platform, it integrates information such as insured risks, exposures, policy limits, attachment points, modeled losses, and underwriting rules. Rather than developing a separate account-level integration for every portfolio, AIG has applied the same ontology-building approach across several increasingly complex commercial arrangements.

​

Screenshot: Editorial illustration of AIG’s reported ontology-based capacity-deployment workflow; not an image of AIG’s production interface. Sources: AIG and Palantir public announcements.

AIG has disclosed three applications of its ontology-based approach:

Everest renewal-rights transaction

• In October 2025, AIG agreed to acquire renewal rights for most of Everest Group’s global retail commercial insurance portfolios, representing approximately $2 billion in premium.
• Everest retained the liabilities and claims-administration responsibilities associated with its existing policies, while AIG gained the opportunity to offer coverage to eligible accounts at renewal.
• To support the transaction, AIG developed an “Everest ontology” — a digital model of the portfolio that enabled its underwriters to evaluate account limits, attachment points, and pricing and determine how the acquired business could fit within AIG’s existing portfolio.

Lloyd’s Syndicate 2479

• AIG subsequently worked with Palantir, Amwins, and funds managed by Blackstone to establish Syndicate 2479, a special-purpose vehicle at Lloyd’s managed by Talbot Underwriting.
• The syndicate was established to underwrite $300 million in premium from a diversified portion of Amwins’ approximately $6 billion delegated-authority portfolio beginning January 1, 2026.
• AIG used Palantir Foundry to validate its portfolio analysis before launch and developed an ontology that enabled large language models to access more than 4 million industry data points.
• Importantly, AIG described the portfolio analysis as already completed, while positioning the use of multiple agents to retrieve data, evaluate risk characteristics, and test programs against the syndicate’s risk appetite as a longer-term capability.

McGill and Partners collaboration

• In March 2026, AIG announced a collaboration with specialty broker McGill and Partners under which AIG expects to provide 25% capacity across up to $1.6 billion of McGill’s specialty gross premiums written.
• After analyzing the portfolio, AIG developed underwriting criteria to support real-time underwriting via McGill’s digital broking platform.
• AIG and Palantir also built an ontology intended to provide near-real-time information on exposures, deployed limits, modeled risk, and losses, enabling AIG to monitor portfolio performance and manage its capacity on an ongoing basis.

For portfolio and capacity managers, the intended workflow change is similar to the change AIG Assist brings to submission triage, but operates at the level of an entire book of business. AIG and its partners describe the resulting capabilities as follows:

• Exposure, limit, modeled risk, and loss data can be integrated into a shared portfolio model rather than assembled manually from multiple systems for each review.
• Accounts and programs can be evaluated consistently against defined underwriting criteria and portfolio-level risk appetite.
• Managers can monitor the effect of deploying additional capacity as portfolio data changes.
• AI agents could retrieve and evaluate relevant information across the portfolio, a capability AIG has characterized as a longer-term goal rather than a current production capability.

The sources do not establish that these capabilities eliminate account-level review or autonomously allocate capital. Instead, AIG presents them as decision-support infrastructure intended to make portfolio analysis faster, more consistent, and more responsive to changing exposure data.​

The three arrangements also use different insurance and capital structures. In the Everest transaction, AIG obtained renewal rights and said it could write qualifying policies on its existing balance sheet without requiring incremental capital. In Syndicate 2479, Amwins and funds managed by Blackstone provide third-party capital for a portfolio managed by AIG through a Lloyd’s vehicle. In the McGill collaboration, AIG provides capacity to qualifying risks distributed through the broker’s platform.

​Palantir supplies the Foundry technology for organizing and analyzing portfolio data. AIG contributes its underwriting criteria, portfolio analysis, and risk appetite. Amwins and McGill provide access to distribution and portfolio data, while Amwins and Blackstone also provide capital to Syndicate 2479. Talbot serves as Lloyd’s managing agent for that syndicate.

​Zaffino has described the objective as combining AI-enabled portfolio insights with underwriting discipline to deploy capacity more quickly. Palantir CEO Alex Karp has similarly positioned the Foundry deployment as a way to support new partnership structures and operating efficiencies.

​This use case is at an earlier stage than AIG Assist and has less disclosed performance data. Syndicate 2479 had operated for only a short period when AIG first discussed its progress publicly, while the McGill collaboration was announced in March 2026. AIG and its partners have not yet disclosed loss ratios, retention rates, capacity utilization, or other underwriting-performance measures attributable to the ontology-based process.​

The available evidence, therefore, establishes the structure and intended operating model, rather than the long-term financial results. AIG has demonstrated that it can apply a repeatable ontology-building method across multiple portfolios: first, to analyze renewal opportunities acquired from another insurer; then, to support a third-party capital vehicle; and finally, to manage capacity through a digital brokerage relationship. Whether that approach improves underwriting profitability or portfolio performance at scale remains to be demonstrated.

​