From comparing carrier filings and earnings scripts, the strongest AI signal often appears first in risk controls and process language before it becomes visible in loss or expense ratios. AIG's first-quarter 2026 package is a clean example: the May 2026 Form 10-Q ties technology execution, AI laws, data security, risk assessment, and insurance liabilities into the company's forward-looking risk language, while the 2026 proxy gives investors the board-level governance detail, including GenAI use in underwriting, claims, and data extraction, the AIG Global Artificial Intelligence Policy, an AI Advisory Council, a Steering Group, a Working Group, and an AI Center of Excellence. The Q1 operating figures were strong, with General Insurance underwriting income of $774 million and an 87.3% combined ratio, improved from 95.8% a year earlier. The governance wording matters because it sits next to those operating results. This is no longer a conference-stage technology story. It is becoming a filing artifact that actuaries can benchmark.

The distinction is not semantic. Investor-day demos describe a future operating model. Securities filings define what management, counsel, auditors, directors, and investors are prepared to stand behind. When an insurer says in a proxy that its board maintains direct oversight over AI strategy and related risks, and that GenAI is being applied to underwriting and claims processes, the disclosure becomes part of the company's control narrative. When the 10-Q then lists the company's ability to implement technological advancements, including AI, and respond to competitors' AI initiatives among risk factors, the operating thesis and the risk thesis meet in the same public record.

For actuaries, that convergence is the point. Pricing models, claim triage systems, submission extraction tools, reserving data pipelines, and expense assumptions do not become investor-relevant only after they move a ratio by 100 basis points. They become investor-relevant when management says the process is material enough to govern centrally, route through escalation procedures, and describe in filings. AIG has now put enough detail in public documents to make insurer GenAI governance a board, audit, and investor issue.

AIG's Filing Language

AIG's Form 10-Q for the quarter ended March 31, 2026 was filed on May 1, 2026. The document does not read like a product brochure. It lists AI inside a broader set of risk factors: technology system disruption, data security, the ability to implement technological advancements, competitor AI initiatives, changes in privacy, data protection, cybersecurity and AI laws, reliance on third parties, the ability to assess risk and estimate related losses, and changes in judgments or assumptions concerning underwriting and insurance liabilities. That grouping is important. AIG is not treating AI as a separate digital project that can be quarantined in innovation language. It is tying AI to the machinery that produces underwriting decisions, loss estimates, controls, and third-party dependencies.

The more detailed AI governance disclosure appears in AIG's 2026 proxy statement, filed March 31, 2026. Under the board oversight discussion, AIG says the board maintains direct oversight over AI strategy and related risks and receives periodic updates from management on the company's ability to build, use, and access AI responsibly and in a way designed to meet regulatory requirements. The company states that it currently uses AI systems in the business, including GenAI applied to aspects of underwriting and claims processes and data extraction in certain lines of business. That is a meaningful operating boundary. Underwriting, claims, and extraction are exactly where model outputs can alter pricing inputs, claims severity workflows, operational loss adjustment expense, and audit trails.

AIG also names the policy infrastructure. The AIG Global Artificial Intelligence Policy defines requirements for governance and management of AI systems and provides a framework to manage and control how the technology may be used in the company's environment, including internally and externally hosted or developed systems. The proxy describes an AI Advisory Council led by the Chief Digital Officer with senior personnel from Legal, Technology, Data, Enterprise Risk Management, Digital, Business Operations, Procurement, and other units as needed. It also describes a Steering Group responsible for escalation and remediation of AI system risks, a Working Group responsible for evaluating and approving new AI use cases and technologies, and an AIG AI Center of Excellence that leads development and deployment in accordance with council and leadership priorities.

That structure is more specific than the usual statement that a carrier uses analytics. It creates named governance actors, approval gates, escalation paths, and a policy object. Actuaries working with models that touch AIG's underwriting or claims processes should read that as a control environment. A pricing indication that depends on AI-extracted submission fields, a claims severity study that uses AI-triaged claim notes, or a reserve review that relies on workflow outputs from an AI-enhanced claims system should be mapped to those governance artifacts before the work reaches audit or regulatory review.

Operating Results Beside the Controls

AIG's Q1 2026 results give the governance language a practical backdrop. The company reported General Insurance net premiums written of $5.6 billion, up 24% on a reported basis and 18% on a constant-dollar basis. General Insurance underwriting income was $774 million, more than tripled from $243 million in the prior-year quarter. The combined ratio improved to 87.3% from 95.8%, an 850 basis point improvement, while the expense ratio improved to 29.3% from 30.5%. The accident year combined ratio, as adjusted, improved to 86.6% from 87.8%, and the accident year loss ratio, as adjusted, remained 57.3%.

Those numbers should not be over-attributed to AI. AIG itself says the underwriting income improvement was driven by lower catastrophe-related charges, improved accident-year underwriting results, and higher favorable prior-year development. Catastrophe-related charges were $180 million, or 3.0 loss ratio points, compared with $525 million, or 9.1 points, in the prior-year quarter. Favorable prior-year development was $132 million, compared with $64 million. The visible financial bridge is catastrophe activity, development, mix, reinsurance program changes, and expense management, not a clean AI effect.

That makes the governance disclosure more useful, not less. If AI is already producing a clean ratio attribution, the actuarial question becomes measurement. Before that point, the actuarial question is control design. AIG's filings show a carrier telling investors that AI affects underwriting and claims processes while the financial statements still present ordinary underwriting drivers. This is the stage when actuarial teams should define baseline metrics, not after management begins claiming ratio lift.

Disclosure item AIG Q1 2026 source Actuarial reading
General Insurance combined ratio improved to 87.3% from 95.8% May 2026 Form 10-Q and Q1 earnings release Do not assign the improvement to AI without a separate causal bridge.
Expense ratio improved to 29.3% from 30.5% May 2026 Form 10-Q Track whether underwriting and claims AI affects acquisition, general operating, or LAE assumptions.
GenAI applied to underwriting, claims, and data extraction March 2026 proxy statement Map use cases to model governance, data quality, and reliance documentation.
Global Artificial Intelligence Policy and AI Advisory Council March 2026 proxy statement Treat policy approval, escalation, and remediation as part of the actuarial control file.

The Peer Benchmark

AIG is not alone in moving AI language into filings, but the pattern differs by carrier. Progressive's 2025 Form 10-K, still the most detailed current Progressive AI risk disclosure available before its May 2026 10-Q, says the company has developed and used machine learning, other forms of AI, predictive models, algorithms, and automated processes for many years, and that the growing development and use of GenAI presents additional risks. Progressive identifies flawed, insufficient, or biased datasets, unfairly discriminatory outcomes, undermined predictions and decisions, brand and reputation effects, vendor AI use, intellectual property uncertainty, and new AI-focused regulations. It also notes that nearly half of insurance departments had adopted the NAIC AI model bulletin at the time of the filing.

Progressive's Q1 2026 Form 10-Q supplies a financial comparator. The company describes underwriting margin and combined ratio as primary measures of underwriting profitability. For Q1 2026, Progressive reported a companywide underwriting profit margin of 13.6%, compared with 14.0% in Q1 2025, and said the loss and LAE ratio and underwriting expense ratio were relatively stable year over year. Progressive also disclosed advertising spend of $1.5 billion, 20% higher than the prior-year quarter, with the current-period impact on the expense ratio partly offset by earned premium growth. That is the kind of detail actuaries need before making any AI expense claim: the ratio moved inside a live system with advertising, premium growth, acquisition costs, and claims mix all moving at once.

Chubb's Q1 2026 filing is less explicit on GenAI governance than AIG's proxy, but it gives a sharp performance benchmark. Chubb reported a P&C combined ratio of 84.0%, improved from 95.7%, with a loss and loss expense ratio of 55.6%, improved from 67.8%. The company said the lower combined ratio reflected lower catastrophe losses, while the current accident year combined ratio excluding catastrophe losses was 82.1%, relatively flat against 82.3% a year earlier. Chubb's forward-looking risk language also refers to the ability of technology resources, including information systems and security, to perform as anticipated, the ability to increase use of data analytics and technology as part of business strategy, and the ability to adapt to new technologies.

Travelers' 2025 Form 10-K places AI in the competitive and operational context. It lists the ability to keep pace with changes in technology and information systems, including artificial intelligence, and the ability to use data and analytics to make decisions among competitive factors. It also defines combined ratio as an indicator of underwriting discipline, efficiency in acquiring and servicing business, and overall underwriting profitability, with a ratio under 100% generally indicating underwriting profit. That definition is plain, but it matters. If AI systems change underwriting discipline or service efficiency, actuaries will eventually need to show where that effect enters the ratio, not merely that the company has better tools.

The peer set shows three disclosure styles. AIG is moving from strategic AI narrative to named board and policy governance. Progressive is emphasizing GenAI risk factors, bias, vendors, intellectual property, and regulatory uncertainty. Chubb and Travelers are embedding technology and data analytics in broader business risk and performance narratives. None of those approaches is wrong. The actuarial task is to translate each into evidence: use-case inventories, model lineage, data provenance, human review standards, outcome monitoring, and ratio bridges.

Why Underwriting and Claims Mentions Matter

Underwriting and claims are not generic process labels. In commercial insurance, underwriting AI can affect submission triage, appetite fit, account prioritization, coverage comparisons, exposure extraction, referral routing, pricing support, and portfolio selection. Claims AI can affect notice routing, severity flags, litigation indicators, fraud detection, reserve recommendations, subrogation identification, medical bill review, repair estimates, and adjuster workload. Each of those points can alter the data actuaries later treat as observed experience.

Data extraction deserves its own mention. AIG's proxy ties GenAI not only to underwriting and claims processes but also to data extraction in certain lines. Extraction is often framed as administrative automation, yet it can be a model-risk inlet. If policy terms, limits, deductibles, locations, construction attributes, payroll classes, medical narratives, legal allegations, or claimant characteristics are extracted from documents and then used downstream, extraction error becomes actuarial data error. ASOP No. 23 already requires attention to data quality, reasonableness, and limitations. AI extraction increases the need to document how fields were captured, validated, corrected, and versioned.

The reserving implications are concrete. If a claims organization uses AI to flag severe claims earlier, case reserves may strengthen sooner even if ultimate loss does not change. That can alter paid-to-incurred relationships, case reserve adequacy, and development patterns. If claim notes are summarized before actuarial review, the summarization process may hide uncertainty or bias severity cohorts. If an AI system speeds low-severity settlement, closing rates and settlement lags can shift. The actuary who sees a development change in 2027 may be reading the delayed effect of a 2026 claims workflow control.

Pricing implications are just as direct. If underwriting AI changes the distribution of bound risks by suppressing weak submissions, increasing appetite consistency, or improving exposure capture, indicated loss ratios may improve for reasons that are partly selection and partly measurement. If data extraction improves classification accuracy, historical experience may not be comparable with newly captured exposure data. If referral rules change under a governed AI process, manual underwriter judgment is still present, but the population reaching manual judgment has changed.

NAIC Governance Pressure

The NAIC's Big Data and Artificial Intelligence work gives state regulators a framework for asking questions that look very similar to the questions auditors and investors will ask. The NAIC artificial intelligence topic page, last updated April 3, 2026, says AI is used in insurance areas including underwriting, pricing, customer service, claims handling, marketing, and fraud detection. It also states that insurers remain responsible for compliance with insurance laws, standards, and consumer protection rules when using AI, including fairness, accuracy, and avoiding unfair discrimination.

The NAIC survey figures show why this is no longer hypothetical. Out of 193 private passenger auto insurers responding, 88% reported that they use, plan to use, or plan to explore AI or machine learning models. Among 194 homeowners insurers, the figure was 70%; among 161 life insurers, 58%; among 93 health insurers, 92%. The NAIC also reports P&C use cases in renewal evaluations, inspections to verify policy characteristics, risk scoring, rate factor relativities, accident image analysis, ultimate claim settlement estimates, and fraud detection. Those are actuarial use cases, even when the system owner sits in underwriting, claims, or technology.

The NAIC Model Bulletin on the Use of Artificial Intelligence by Insurance Companies was adopted in December 2023. The NAIC says the bulletin reminds insurers that decisions or actions made or supported by AI must comply with applicable insurance laws and regulations, sets expectations for AI governance, and identifies information departments may request during investigations or examinations. The Big Data and Artificial Intelligence Working Group is also developing an AI Systems Evaluation Tool for market conduct, financial analysis, or financial examination contexts, including information about use, governance, risk mitigation, and third-party models.

AIG's disclosure reads cleanly against that supervisory direction. A global policy, an advisory council, escalation and remediation responsibilities, use-case review, and a center of excellence are exactly the kind of artifacts regulators can ask to see. For actuaries, the practical question is whether the actuarial file points to those artifacts or sits apart from them. A rate filing, reserve memo, or model validation package that references AI-supported data but cannot identify the approving governance body is weak. A file that connects the actuarial reliance to policy approval, model inventory, human review, monitoring, and remediation is much stronger.

Actuarial Documentation Before the Next 10-K

The right preparation is boring in the best sense. Actuarial teams should maintain an AI use-case inventory for any tool that affects pricing, underwriting selection, exposure capture, claims triage, case reserving, settlement, fraud scoring, loss development data, or expense allocation. Each use case should identify the business owner, model owner, vendor, data sources, output users, human review point, approval body, validation date, monitoring metrics, known limitations, and remediation path. If the carrier has a global AI policy, the actuarial inventory should use the same categories and identifiers.

Model validation should cover more than predictive accuracy. For GenAI-supported extraction and summarization, validation should include field-level error rates, false confidence patterns, prompt or instruction versioning, document-type coverage, reviewer override rates, drift in source document mix, and downstream sensitivity. For underwriting decision support, validation should include referral changes, bind-rate shifts, appetite exceptions, protected-class proxy review where applicable, and segmentation effects. For claims support, validation should include severity lift charts, reserve change timing, litigation flag precision, cycle-time effects, reopen rates, and adjuster override behavior.

Expense assumptions need the same discipline. Many AI business cases promise lower underwriting expense, lower LAE, or faster cycle times. The actuarial file should separate gross productivity from net expense impact. A tool that saves adjuster hours may not reduce LAE if it increases review layers, vendor fees, quality assurance, legal review, or technology amortization. A tool that improves submission handling may increase acquisition expense if it supports growth through higher broker volume. AIG's Q1 expense ratio improved, but the filing does not say AI caused it. That restraint is a useful benchmark.

Vendor reliance also needs a documented boundary. AIG's policy disclosure explicitly covers internally and externally hosted or developed AI systems. That phrase matters because many carrier AI deployments use a mixed stack: internal data, third-party platforms, cloud infrastructure, model vendors, consulting partners, and business-unit overlays. Actuaries do not need to own every layer, but they do need to know which layer produced the actuarial input and what assurance exists for that layer.

Why This Matters

AIG's filings show the next phase of insurer GenAI governance. The story is not that one carrier mentioned AI. Many have. The story is that AIG has placed underwriting, claims, extraction, global policy governance, board oversight, risk escalation, and Q1 operating performance in a public disclosure set that investors can read together. That gives actuarial teams a benchmark for what credible AI control language looks like before ratio attribution becomes visible.

This continues a trend we have seen across recent insurance AI disclosures: the governance layer is maturing faster than the public ROI layer. Verisk has described contracting friction around AI governance. The NAIC is moving toward evaluation tools. Progressive is warning about GenAI bias, vendors, intellectual property, and regulatory expectations. Chubb and Travelers frame technology and analytics as competitive and operational capabilities. The common thread is control evidence.

Actuaries should not wait for management to ask for an AI ratio bridge in the annual report. The work starts earlier: define the baseline, map the workflow, preserve pre-deployment experience, identify the governance owner, document the model or extraction process, and monitor whether experience changes because risk changed, measurement changed, or the portfolio changed. That is the difference between an AI story and an actuarial control file.

Sources

Further Reading