AIG's agentic AI: orchestration, throughput, and the new underwriting workflow
AIG says generative AI is already moving the needle on underwriting capacity, operating cost, and portfolio integration. Recent Investor Day disclosures point to measurable throughput gains and redesigned workflows-signals insurance leaders watch for when impact moves from pilot to production.
Leadership initially called projections "aspirational." Now, the tone has shifted. According to CEO Peter Zaffino, the company is processing submission flow without additional human capital-an outcome that directly improves unit economics.
What changed
AIG reports it embedded generative AI across core underwriting and claims in 2025, expanding adoption through an internal tool, AIG Assist, used across most commercial lines. In excess and surplus, Lexington targeted 500,000 submissions by 2030 and has already surpassed 370,000 in 2025, indicating real volume handling at scale.
The headline isn't just "we use LLMs." It's the throughput. Faster intake and triage, more consistent case handling, and fewer manual loops are showing up as capacity and cycle-time gains-not promises, numbers.
How they're doing it
AIG uses generative models to extract and summarize incoming data, then routes decisions through an orchestration layer that coordinates multiple AI agents. These agents operate alongside teams, surface real-time context, draw on historical cases, and challenge underwriting decisions.
That orchestration compresses the "front-to-back workflow"-from intake to risk assessment to claims handling. Repetitive steps shrink, handoffs drop, and analysis that took hours now fits inside an underwriting session.
Proof points from transactions
During the conversion of Everest's retail commercial business, AIG built an ontology of the acquired portfolio, merged it with its own, and prioritized renewals in a fraction of the time. Ontology work is hard and often underestimated; doing it well is how you avoid months of friction.
For Lloyd's Syndicate 2479 (with Amwins and Blackstone), AIG-working with Palantir-used LLMs to test if Amwins' program portfolio aligned to the syndicate's risk appetite. Management says there's a strong pipeline of SPV opportunities ahead. For context on market structures, see Lloyd's syndicates, and for the data platform partner, Palantir.
Why this matters for insurers
- Capacity without headcount: More submissions processed per FTE changes the math on growth and expense ratio.
- Cycle-time compression: Faster triage, faster quotes, tighter renewals-better broker experience and higher hit ratios.
- Portfolio integration: Ontology-driven mapping accelerates M&A and program onboarding (and avoids long, costly cleanups).
- Decision quality: Agent "companions" nudge consistency, expose precedents, and surface blind spots in real time.
- Risk controls: Claims of "no bias" require measurable audits, override tracking, and model monitoring in production.
How to act in your shop
- Start where volume is high and variance is painful: submission intake, appetite triage, clearance, and first-pass quoting.
- Build (or buy) an orchestration layer: coordinate agent roles (ingest, summarize, classify, compare, recommend) across the workflow.
- Instrument everything: track submissions per FTE, touch time, SLA adherence, quote-to-bind, and override rates from day one.
- Invest in ontology and data pipelines early: map exposures, coverages, industries, brokers, and programs to common definitions.
- Keep underwriter-in-the-loop: require reason codes, show evidence chains, log challenges and overrides for auditability.
- Stand up governance: model validation, PII controls, bias tests, lineage, and human accountability at each decision gate.
- Pilot with clear boundaries: one LoB, one geography, one broker cohort; 90-day targets, weekly reviews, and a rollback plan.
Metrics that prove impact
- Submissions per underwriter FTE and triage SLA (minutes to disposition)
- Quote cycle time and quote-to-bind ratio by segment
- Manual rework rate and agent recommendation override rate
- Loss ratio shift from selection effects; expense ratio change tied to AI-enabled throughput
- Claims FNOL-to-payment time and leakage reduction (where agents assist adjusters)
Risks and watch-outs
- Ontology cost: portfolio mapping and data normalization often take longer than model work-budget for it.
- Hallucinations and drift: require retrieval-grounded answers, confidence thresholds, and human review on material decisions.
- Legacy integration: your policy admin, document stores, and broker portals will be the bottleneck-plan adapters early.
- Change fatigue: underwriter adoption rises when agents show their work and reduce toil on day one.
Next steps
Pick one workflow slice, define two measurable outcomes, and run a bounded pilot with orchestration at the center. The lesson from AIG: the win comes from end-to-end redesign, not a single impressive model.
Related training
Your membership also unlocks:
