AIG CEO Peter Zaffino on using Palantir to rebuild underwriting and cut portfolio analysis from months to days

AIG's Palantir Deployment

Peter Zaffino, executive chairman of AIG as of this week and outgoing CEO, has spent nine years transforming a company that was broken across underwriting, operations, data, and capital. The pitch for what Palantir enabled is specific: AIG was running portfolio analysis on a 30-to-90-day lag. The goal now is to do it daily.

AIG writes complex commercial risk — shipping, marine, energy, Fortune 500 liability — with a 50/50 split between North America and international, and Japan as its second-largest market. Underwriting at that scale can't be done policy by policy. You have to manage the aggregate, and the aggregate used to be stale by the time anyone looked at it.

“We did the full ontology of AIG, and then we went to look at an acquisition called Everest, which had about $2,000,000,000 of premium. We got Palantir in to work with our team. We could build an ontology of Everest's portfolio on top of ours in four days. Having the ability to assess risk and use quantitative data to make better decisions on a daily basis is the aspiration of the way the company is going.”

The Ontology Bet

The clearest proof point Zaffino offers is the Everest acquisition. AIG had already built a full Palantir ontology of its own portfolio. When it moved to analyze Everest's roughly $2 billion in premium, the team layered Everest's portfolio on top of AIG's existing ontology in four days — a process that would previously have taken months. A side effect of building out that ontology was discovering they didn't need centralized data lakes as much as assumed. Going directly to admin platforms for data, rather than routing through scrubbed repositories, turned out to be faster and good enough.

Zaffino is also direct about where the real value sits in the AI stack. LLMs help extract more from incoming structured and unstructured data and get it into a digital workflow faster. But without an ontology — a working digital twin of the business — you can't direct agents or track how decisions compound across the portfolio. The ontology is the substrate; the LLMs run on top of it.

Deployment Model

Palantir's forward-deployed engineers sit embedded with AIG's business and technology teams, not in a separate innovation lab. Zaffino frames the iteration cycle as 90-day increments rather than annual roadmaps, which he credits for keeping the relationship productive. The starting point for each cycle is a shared view between Karp and two senior Palantir executives, Ryan and Ted, on what AIG is actually trying to achieve — so the engineers are translating toward a known business outcome rather than building in the abstract.

Workforce

Zaffino pushes back on the layoff narrative directly. AIG's stated intention is growth and reskilling, not headcount reduction. He does acknowledge that processes where humans have been functioning as manual LLMs — trained to execute tasks outside a normal workflow — need to go. He frames that as ordinary business hygiene rather than AI-driven cuts. Whether the growth ambition absorbs the efficiency gains is an open question, but Zaffino's position is that the deployment was never designed around a headcount target.