AIG’s most important artificial-intelligence claim is not that a model can summarise an insurance submission. It is that multiple agents, coordinated across underwriting and claims workflows, can increase the volume the company handles without a matching increase in staff.
Chief executive Peter Zaffino says early projections once considered aspirational now appear conservative. He describes a “massive change” in AIG’s ability to process submission flow and says the capacity gain has been the largest surprise.
The insurer embedded generative AI in core underwriting and claims processes during 2025. Its internal AIG Assist tool now operates across most commercial lines, while a newer orchestration layer coordinates agents that extract information, retrieve historical cases, support decisions, and connect stages of work.
Capacity Gains Come From the Complete Workflow
Lexington Insurance, AIG’s excess and surplus unit, has targeted 500,000 submissions by 2030. It had already exceeded 370,000 during 2025. Generative systems help extract and summarise incoming material, reducing the time required to understand each opportunity.
The larger economic benefit depends on what happens next. A summary still creates work if an underwriter must move it manually between risk assessment, pricing, and claims systems. Orchestration aims to compress that “front-to-back” process by assigning tasks to specialised agents and passing results through a controlled workflow.
AIG describes the agents as companions that work with teams. They provide real-time information, draw comparisons with historical cases, and challenge an underwriting position rather than silently making the final decision.
Useful performance measures should include submission turnaround, cases handled per underwriter, correction rates, decision consistency, and portfolio outcomes. Higher throughput is valuable only if risk selection remains sound.
Portfolio Integration Tests More Than Summarisation
AIG applied the technology while converting Everest’s retail commercial business. The company built an ontology of the incoming portfolio, aligned it with its own information, and prioritised accounts for renewal. Management says the work was completed in a fraction of the expected time.
Ontology alignment is the difficult part hidden behind the phrase “combine the data.” Two insurers may describe industries, exposures, policies, and claims differently. Before models can compare portfolios, teams must reconcile meanings, identifiers, and confidence in the underlying records.
A similar approach supported Lloyd’s Syndicate 2479, launched with Amwins and Blackstone. Working with Palantir, AIG used language models to assess whether the Amwins programme portfolio matched the syndicate’s risk appetite.
These examples move AI beyond document productivity into portfolio decisions with direct capital consequences. They require clear sources, reviewable mappings, and named responsibility when automated analysis influences which business is pursued.
Orchestration Creates a New Control Layer
Coordinating agents can reduce handoffs, but it also creates dependencies. An extraction error may travel into risk analysis; an incorrect portfolio mapping may influence prioritisation; an agent that retrieves the wrong historical case may make a recommendation look better supported than it is.
The orchestration layer therefore needs permission boundaries, validation between stages, audit trails, and a reliable way to stop or reverse a workflow. Underwriters should see both the recommendation and the evidence that produced it.
AIG’s language about unbiased analysis also deserves careful testing. Historical cases can contain selection patterns and uneven outcomes. Scale does not remove bias; it can distribute it more efficiently unless performance is evaluated across relevant groups and business lines.
The case shows why enterprise agent value is increasingly an integration question. A capable model is only one component. Data definitions, workflow design, human review, and orchestration determine whether the model changes capacity and cost.
If AIG’s reported gains persist without weakening underwriting quality, the result will be significant. The insurer will have shown that agents create value not by replacing judgment, but by preparing more decisions, connecting more evidence, and allowing scarce expertise to operate across a larger book.
