Underwriters are supposed to price risk. Instead, most of their week goes to reading loss runs, chasing brokers for missing exhibits, and re-keying data between systems. At AIG, that ratio has flipped: the carrier now reports its underwriters spend 50%+ less time on data ingestion and submission triage — time redirected to the judgment calls that actually decide what gets bound, and at what price.
Commercial insurance runs on paperwork. A single excess-and-surplus (E&S) submission can carry dozens of pages of loss history, statements of values, broker notes, and prior policy terms — and a senior underwriter has to read all of it before deciding whether the risk is even worth quoting. For decades, that reading bottleneck capped how many submissions one underwriter could realistically carry in a week, no matter how sharp their pricing instincts were. There's a better way, and it's already running in production at one of the industry's largest carriers.
AIG's Lexington unit — its E&S arm — rolled an AI-assisted underwriting workflow across middle-market property submissions. The measured result: a 30% increase in quoted submissions, a 55% reduction in time-to-quote, and roughly 40% more submissions actually bound. Lexington had already logged more than 370,000 submissions by year-end 2025, on a path toward AIG's own stated ambition of 500,000 by 2030 — a volume no underwriting bench could handle by adding headcount alone.
What AIG's AI Underwriters Actually Do All Day
AIG built the system — internally called Underwriter Companion — with Palantir and Salesforce, running on Anthropic's Claude. Claude, accessed through AWS Bedrock, handles the natural-language reasoning: reading an incoming submission, summarizing the exposure, and drafting the next question a senior underwriter would actually ask the broker. That last part matters — the system isn't just extracting fields, it's anticipating the follow-up a 20-year underwriting veteran would raise.
None of this replaces the underwriter's call on price and appetite. It replaces the hours spent turning unstructured PDFs into structured facts and chasing brokers for the exhibit that's always missing — the part of the job that adds friction, not value. If you manage underwriters, this is the distinction worth sitting with: the automation targets the reading, not the deciding.
Inside the Stack: Claude, Foundry, and the Ontology Underwriters Never See
Palantir Foundry and its AI Platform (AIP) supply the part of the system underwriters never directly interact with: a structured ontology mapping policies, exposures, and historical losses, against which Claude's reasoning is grounded and checked. Foundry also orchestrates the flow between systems, so a submission moves through triage, enrichment, and routing without manual handoffs. Salesforce sits on the broker-facing side — intake, correspondence, and status tracking — so brokers see a single, consistent workflow regardless of what's happening underneath.
AIG's CEO, Peter Zaffino, has described the company's next phase as a shift from single-task copilots to a multi-agent orchestration layer: specialized AI agents coordinating with each other on a shared task, rather than each answering isolated prompts. Early Claude-based agents at AIG could run autonomously for under an hour before needing a human check-in. AIG now reports agents operating unsupervised for up to 30 hours on complex, multi-step underwriting tasks — a maturity curve worth watching if your own AI pilots are still measured in single-digit minutes.
The Results: 55% Faster Quotes, 40% More Binds
In Lexington's middle-market property book, the AI-assisted workflow delivered a 30% increase in quoted submissions, a 55% reduction in time-to-quote, and approximately 40% more submissions bound — without a proportional increase in underwriting headcount. AIG frames the ambition plainly: turning what one underwriter could handle into roughly the throughput of five.
Zoom out and the pattern holds industry-wide: insurers using AI-powered underwriting and claims automation are resolving work 75% faster with 30–40% cost reductions in adjacent claims processes, and underwriting expense ratios are projected to decline 15–20% in P&C and more than 25% in life insurance as these systems mature through 2026 and beyond. AIG isn't an outlier chasing a headline — it's an early, well-documented data point in a trend the rest of the industry is racing to catch up to.
The Cost of Waiting Is Already Visible
This isn't a pilot sitting in an innovation lab. AIG is rolling the same architecture out across Glatfelter, AIG Re, and its core commercial property and casualty lines through 2026. Regulators are moving in parallel, not behind: 23 U.S. states plus Washington, D.C. have adopted the NAIC Model Bulletin on AI use in insurance, which requires documented human oversight of any AI-influenced underwriting decision. Carriers that wait to build both the AI workflow and the governance layer around it will be doing two hard things at once, later, and under considerably more regulatory scrutiny than AIG faced building this in 2025–2026.
The Tools Behind the System (And What They'd Actually Cost You)
Nothing in AIG's stack is exotic or unreleased. It's the same category of enterprise tools available to any carrier today, assembled around a genuinely hard integration problem: getting a large language model to reason reliably against a live, governed data model instead of hallucinating against a blank context window.
| Tool | Role in This Workflow | Free Tier? | Paid From |
|---|---|---|---|
| Anthropic Claude (via AWS Bedrock) | Reads submissions, drafts broker follow-up questions, summarizes exposure | No | Usage-based, AWS Bedrock consumption pricing |
| Palantir Foundry / AIP | Ontology of policies, exposures, and losses; cross-system orchestration | No | Enterprise contract, quote-based |
| Salesforce | Broker intake, correspondence, submission status tracking | Limited | From ~$25/user/month, Sales Cloud entry tier |
[REQUIERE VERIFICACIÓN]: AIG's specific per-seat or per-submission cost for this deployment isn't publicly disclosed; the figures above reflect general list pricing for the underlying platforms, not AIG's negotiated enterprise terms.
Who Should Actually Care About This
This is most directly relevant to Chief Underwriting Officers, E&S and specialty-line leaders, and underwriting operations teams at carriers where submission volume is outgrowing underwriter headcount. It's the clearest fit for property, casualty, and specialty lines carrying high document complexity per submission. It's a poor fit for small MGAs or single-line personal auto books, where submission volume and complexity rarely justify a Foundry-scale ontology build. And if your bottleneck is genuinely pricing accuracy rather than document triage, this specific architecture is solving a problem you don't have — worth naming honestly before committing budget to it.
What Every Carrier Should Take From This
Three things stand out from AIG's rollout that apply well beyond one company:
- The reasoning layer and the data layer are separate investments. Claude handles language; Foundry handles the structured, governed ontology it reasons against. Skipping the ontology work and bolting an LLM onto messy legacy data is the most common way these projects underdeliver.
- Governance isn't optional overhead — it's what makes scale possible. AIG's rollout tracks closely with, not around, the NAIC Model Bulletin's human-oversight requirements now adopted in 23 states plus D.C.
- Measure time-to-quote and bind rate, not just "time saved." AIG's most convincing numbers — 55% faster quotes, 40% more binds — are business outcomes, not just efficiency metrics, and that's what makes the ROI case defensible to a board.
Frequently Asked Questions
Is AIG's exact system available to other insurers?
No. Underwriter Companion is a proprietary build combining AIG's own data and workflows with Palantir Foundry and Claude via AWS Bedrock. Other carriers can license the same underlying platforms — Palantir, Anthropic, Salesforce — but would need to build their own ontology and integration layer from scratch.
Does this replace human underwriters?
AIG frames it explicitly as augmentation, not replacement. The stated goal is turning one underwriter's daily throughput into roughly five underwriters' worth of output by removing document triage — not by removing underwriting authority or judgment from the pricing decision.
How long did it take AIG to reach this scale?
The maturity curve was fast by insurance-industry standards: agents that could run autonomously for under an hour at the Claude 2.0 stage now operate for up to 30 hours per task, with the Lexington results reported at year-end 2025 rollout scale — roughly 18-24 months from early pilots to measurable, board-reported results.
Multi-agent orchestration — not single chatbots, but coordinated AI agents handling entire workflows end to end — is becoming the default architecture for AI in commercial insurance through 2026. The tools to build it already exist and are available to any enterprise carrier today. The only variable left is which carriers start building the ontology now, and which ones wait until the gap is no longer closable.