American International Group processed over 370,000 excess and surplus submissions in 2025—a number that was supposed to take until 2030 to reach. CEO Peter Zaffino calls it “massive change.” Chief Digital Officer Claude Wade calls it an “agentic AI ecosystem.” What they built isn’t a chatbot upgrade or a document scanner with AI branding. It’s a coordinated system of autonomous agents that make decisions, challenge underwriter assumptions, and compress hours-long workflows into real-time analysis. Here’s how it works and what it actually costs to build.
What Agentic AI Orchestration Actually Means at AIG
The distinction between standard AI automation and agentic AI orchestration matters more than the terminology suggests. Standard automation handles discrete tasks—summarize this document, extract these fields, route this email. Agentic AI orchestration coordinates multiple autonomous agents that make decisions across entire workflows, learn from outcomes, and adapt without constant human reprogramming.
AIG’s system ingests data from over 30 approved third-party sources alongside internal datasets. Large language models handle document classification and extraction. Coordinated agents then prioritize target risks, score binding probability, and route submissions through appropriate workflows—automatically, in parallel, without sequential department handoffs. According to AIG’s 2025 earnings calls, the result is that one underwriter now handles the workload that previously required five people. That’s not an efficiency improvement. That’s a structural change in how insurance operations scale.
The Anthropic and Palantir Partnership
AIG didn’t build this internally from scratch. At their 2025 Investor Day, Anthropic CEO Dario Amodei and Palantir CEO Alex Karp joined Zaffino to discuss the collaborative architecture. Anthropic provides the foundational large language models that power agentic reasoning. Palantir handles data orchestration and ontology building—the structural layer that allows different agents to share context and coordinate effectively across disparate data sources.
That partnership structure matters for anyone evaluating similar implementations. The combination of frontier LLM capabilities with specialized data infrastructure is what makes the coordination layer functional. Either component alone produces more limited results—a lesson embedded in AIG’s $1 billion technology investment decision.
The ontology layer deserves particular attention because it’s where most enterprise AI implementations fail quietly. Without a shared data structure that different agents can reference consistently, coordination breaks down—agents make conflicting decisions based on incompatible data representations. Palantir’s ontology work gives AIG’s agents a common language for describing risks, policies, and submissions across the 30-plus third-party data sources the system integrates.
That foundation is what allows the orchestration layer to function at the submission volume AIG is processing. Building it requires specialized expertise that most insurers don’t have internally, which is why the partnership model is architecturally important rather than just a vendor preference.
AIG Assist: How the System Works in Practice
AIG’s flagship tool, AIG Assist—also called AIG Underwriter Assistance—rolled out across most commercial lines throughout 2025. The workflow operates in four coordinated phases that run simultaneously rather than sequentially.
Data ingestion agents automatically categorize and extract information from submissions: broker communications, financial statements, third-party risk assessments. Risk prioritization agents then analyze submission quality, binding probability, and profit potential to route high-value opportunities to underwriters first. During actual underwriting reviews, decision support agents surface relevant historical cases, flag potential blind spots, and provide real-time pricing context. Portfolio management agents run continuous simulations to model risk-adjusted returns and identify diversification opportunities across the entire book of business.
The Everest portfolio integration demonstrates what this looks like under real M&A pressure. When AIG acquired Everest’s retail commercial portfolio, AI agents built an ontology that merged disparate datasets and prioritized renewals in weeks rather than the months that manual data mapping typically requires. That compression of integration timeline is where the $1 billion technology investment starts showing measurable ROI beyond the per-submission metrics.
The Propensity-to-Bind Advantage
One capability that doesn’t appear in standard automation is propensity scoring—predicting which submissions are likely to convert to policies before underwriters spend time on them. AIG’s orchestration layer analyzes historical binding patterns, broker relationships, and submission characteristics to score each opportunity. Underwriters focus attention on high-conversion submissions while agents handle initial screening on lower-probability cases.
In competitive insurance markets where broker relationships and cycle time directly affect binding rates, this isn’t just an efficiency play. It’s strategic resource allocation that affects which business AIG wins. Faster triage, faster quotes, and faster renewals improve broker satisfaction in ways that compound over renewal cycles.
The Implementation Lessons Other Insurers Are Missing
AIG’s experience points to three decisions that most insurers evaluating agentic AI get wrong. The first is sequencing—deploying isolated AI tools before building coordination infrastructure. AIG’s approach prioritized the orchestration layer first, which is what allows individual agents to produce compound value rather than additive value. An underwriting tool and a portfolio tool running independently produce two sets of improvements. The same tools coordinated through an orchestration layer produce decisions that account for portfolio-level implications at the underwriting stage.
The second mistake is underweighting data foundation requirements. AIG’s integration of 30-plus third-party sources isn’t a feature of their system—it’s a prerequisite for autonomous decision quality. Agents making risk prioritization decisions with incomplete data produce worse outcomes than experienced underwriters working with the same incomplete data. The data infrastructure investment comes before the AI investment, not alongside it.
The third is governance sequencing. Insurers that deploy agentic capabilities and then retrofit governance frameworks face regulatory exposure during the gap. AIG built human oversight protocols, audit systems, and bias monitoring as part of the initial architecture rather than compliance additions. That sequencing is what makes their 370,000-submission processing volume defensible in regulated markets—not just technically possible.
AIG’s Lexington Insurance unit processed 370,000-plus excess and surplus submissions in 2025, surpassing the original 2030 target of 500,000 ahead of schedule. According to Zaffino’s investor presentations, agentic AI orchestration delivers 2X-5X faster end-to-end underwriting across commercial lines, 100% review coverage of financial lines submissions without additional underwriting staff, and real-time portfolio optimization through large-scale risk simulations.
What makes these metrics credible rather than marketing is context. AIG is processing real submissions at production scale while most competitors are still running pilots on isolated workflows. The gap between pilot performance and production performance in enterprise AI is where most implementations fail—AIG has crossed that threshold. Their excess and surplus operations now target $4 billion in new premiums by 2030, with submission processing capacity as the enabling constraint rather than underwriting headcount.
Implementation Challenges AIG Had to Solve
AIG’s results don’t mean the implementation was straightforward. Agentic AI systems face the same failure modes as other AI applications—hallucinations on edge cases, potential bias in risk scoring, regulatory compliance requirements that vary by state, and the need for transparent audit trails on automated decisions that affect policyholder outcomes.
AIG addresses these through human oversight protocols that require underwriter approval for decisions above defined risk thresholds. Audit systems track agent decision patterns and flag anomalies for review. Bias monitoring checks for disparate treatment across risk categories. Regulatory compliance tracking logs the reasoning behind automated decisions for markets where insurance commissioners require it.
The governance architecture isn’t peripheral to the technology—it’s what makes autonomous decision-making legally defensible in a regulated industry. Insurers that deploy agentic AI without this layer face regulatory exposure that can shut down implementations faster than any technical failure would.
Where Agentic AI Orchestration Falls Short
AIG’s system excels at standard submissions that fit historical patterns. Complex commercial risks with genuinely novel circumstances still require human expertise that the current generation of agents can’t reliably replicate. Edge cases that don’t match training data produce degraded outputs—which in insurance means potential mispricing or coverage gaps that create downstream liability.
Regulatory fragmentation creates hard limits on autonomy. Insurance commissioners in different states have varying requirements for automated decision-making. What’s permissible in one market may require human sign-off in another, which means the orchestration layer needs jurisdiction-aware routing logic on top of the core AI architecture. That’s implementation complexity that doesn’t appear in vendor demos.
The capital requirements also constrain adoption. AIG’s $1 billion technology commitment isn’t a realistic entry point for regional carriers or specialty insurers. Smaller organizations evaluating similar approaches need to scope implementations to specific high-volume, standardized processes rather than enterprise-wide deployment—which produces proportionally smaller results even when technically successful. The technology works at AIG’s scale. Whether it produces comparable ROI at smaller scale is a different question that the current evidence doesn’t answer.
There’s also a talent dependency that AIG’s public communications understate. The governance frameworks, ontology architecture, and agent coordination logic require engineers and data scientists who understand both insurance operations and enterprise AI systems—a combination that’s genuinely scarce in the current labor market. Insurers that underestimate this talent requirement discover it during implementation when timelines slip and costs escalate beyond initial projections. AIG’s partnerships with Anthropic and Palantir partly address this by externalizing specialized expertise, but that approach carries its own dependency risks if key partnerships change or vendor priorities shift.
Frequently Asked Questions
How does agentic AI orchestration differ from standard insurance automation?
Standard automation handles discrete tasks like data entry or document routing. Agentic AI orchestration coordinates multiple autonomous agents that make decisions, learn from outcomes, and adapt across entire workflows without constant human reprogramming. AIG’s system handles intake, risk prioritization, underwriting support, and portfolio management simultaneously—not sequentially through separate tools.
What specific results has AIG achieved with agentic AI orchestration?
AIG processed over 370,000 excess and surplus submissions in 2025, surpassing a target that was originally set for 2030. According to CEO Peter Zaffino’s investor presentations, the system delivers 2X-5X faster end-to-end underwriting and 100% review coverage of financial lines submissions without additional underwriting staff.
Which technology partners did AIG use to build their AI orchestration system?
AIG partnered with Anthropic for foundational large language model capabilities and Palantir for data orchestration and ontology building. At AIG’s 2025 Investor Day, Anthropic CEO Dario Amodei and Palantir CEO Alex Karp joined CEO Peter Zaffino to discuss the collaborative architecture. The partnership approach—combining frontier AI with specialized data infrastructure—reflects a deliberate decision not to build core capabilities internally.
Can smaller insurers implement similar agentic AI systems?
AIG’s $1 billion technology investment and enterprise-scale partnerships represent requirements that most regional carriers can’t match directly. Smaller insurers evaluating agentic AI should focus on specific high-volume, standardized processes—submission triage, renewals processing, or claims routing—rather than enterprise-wide orchestration. The ROI at smaller scale hasn’t been demonstrated at the same level as AIG’s implementation.
What governance requirements does agentic AI orchestration need in insurance?
Insurance regulators require transparent audit trails for automated decisions affecting policyholders. AIG’s governance framework includes human oversight protocols for decisions above risk thresholds, audit systems that track agent decision patterns, bias monitoring across risk categories, and jurisdiction-aware compliance tracking. This governance architecture is what makes autonomous decision-making legally defensible in regulated insurance markets—not optional infrastructure.
