Dual-Vendor AI Stacks Emerge as Carriers Hedge Model Risk

An IA Capital survey published in May 2026 found OpenAI inside roughly nine out of every ten carrier AI stacks, with Google Gemini absent from carrier deployments entirely. For office software that concentration would be unremarkable. For models that price claims and draft underwriting recommendations, the two largest carriers have started treating it as an operational risk and buying against it.

Key Takeaways

• Roughly 90% of carrier AI stacks contain OpenAI on IA Capital's May 2026 survey, and Google Gemini appears in none of them.
• 10,000 Travelers engineers, data scientists and analysts run Anthropic Claude assistants, while the customer-facing AI Claim Assistant runs on OpenAI. The split is by workload risk profile, not by accident of procurement.
• Autonomous agent runtime went from under an hour to 30 hours at AIG since it began working with Claude, which is what makes a multi-agent underwriting file feasible rather than a chain of check-ins.
• The expense ratio moved from 31.5% to 28.5% over nine years at Travelers on rising technology spend, and two vendors now make that improvement harder to decompose into sustainable and one-off parts.
• 30 minutes to two hours per estimate is the time saving Verisk cites for its Model Context Protocol connectors, which decouple the data layer from the model layer and lower the cost of running a second vendor.

Two Architectures, Both Deliberately Split

Travelers has disclosed the clearest dual-vendor architecture in P&C. In January 2026 it partnered with Anthropic to put Claude and Claude Code assistants in front of nearly 10,000 engineers, data scientists, analysts and product owners. A month later it launched the AI Claim Assistant on OpenAI's models and Realtime API.

CTO Mojgan Lefebvre gave the reasoning in a Fortune interview in April 2026: it is too early in the AI journey to do everything with one partner, and you also do not want ten of them. She placed OpenAI at the forefront for conversational capability and Anthropic ahead on analytical, coding and engineering work.

Dimension	OpenAI (Customer-Facing)	Anthropic (Internal Engineering)
Primary use case	AI Claim Assistant: agentic voice for FNOL	Personalized coding and analytics assistants
User population	External customers calling claims	10,000 engineers, data scientists, analysts
Key capability	Real-time voice, conversational fluency	Code generation, model development, documentation
Deployment model	Agentic (autonomous call handling)	Assistive (augmenting human workflows)
Data exposure	Customer PII, claim details	Internal code, proprietary models, institutional knowledge

AIG layers rather than splits. On the Q1 2026 call, Peter Zaffino described using Palantir Foundry to expand AIG's ontology and orchestrate teams of agents for submission ingestion, extraction, risk evaluation, pricing benchmarks and synthesis. Palantir is the workflow backbone; Claude is the reasoning engine inside the agents. AIG Assist runs across eight lines against a targeted 500,000 annual E&S submissions, and in Lexington middle market property produced a 30% improvement in quoting volume, a 55% cut in time to quote and roughly 40% more submissions bound.

The Split Is a Risk Allocation, and It Shows Up in the Expense Ratio

Segmenting by workload is segmenting by consequence of failure.

A model error in the customer-facing claim assistant reaches a policyholder directly, with E&O and reputational exposure attached. An error in an internal engineering assistant is caught in code review before it touches anything external. Splitting those across vendors also narrows the regulatory answer: a question about customer-facing AI decisions involves one vendor and a defined scope.

Availability follows the same logic. Travelers handles roughly 50% of initial loss reports digitally with AI agents on about 35% of low-complexity claims, so an outage at one provider degrades that channel without stopping the engineering organization. A single-vendor carrier loses everything at once. Autonomous runtime at AIG has gone from under an hour to 30 hours, which raises the value of the workflow and the cost of it stopping in the same motion.

The actuarial consequence sits in expense decomposition. Travelers' expense ratio improved from 31.5% to 28.5% over nine years while technology spend rose. For a rate indication, the question is which part of that three points is durable. Customer-facing automation and engineering productivity have different persistence: one scales with claim volume, the other is a one-time step change in development cost. With two vendors carrying two workloads, the savings no longer decompose from the aggregate, and an expense assumption built on the blended figure is assuming a mix that may not repeat.

For ERM and ORSA work the same architecture is a disclosable dependency. A carrier running one provider across underwriting, claims, service and engineering carries a different operational risk profile than one that has deliberately separated them, and the NAIC's Third-Party Data and Models Working Group is drafting a vendor registration framework that will make those dependencies visible.

What the Second Vendor Does Not Fix

Diversification manages a carrier's own concentration. It does nothing about the industry's.

The 90% penetration figure is a correlation structure, not a market share. A model failure, a provider data breach, or a change in terms of service would reach a large share of the market at once. Cyberwrite's chief executive warned in a February 2026 interview with The Insurer that vendor concentration creates accumulation risk. The analogue is geographic concentration in a catastrophe model: the individual exposure is priceable, the correlation is what breaks the aggregate.

The second vendor is also not free. Model governance surface doubles, because each provider has its own update cadence, documentation practice and transparency about capability changes. Travelers absorbed that by building TravAI as an orchestration layer with unified access controls and audit logging beneath both models; AIG pushed it into Palantir Foundry. Both answers require the engineering budget to build or buy an abstraction layer first.

That is the part that does not generalize down-market. Verisk's Model Context Protocol connectors, which route Underwriting Intelligence and XactRestore into Claude conversationally and save an estimated 30 minutes to two hours per estimate, lower the integration cost by separating the data layer from the model layer. A Carrier Management analysis in April 2026 called the requirement a plug-and-play operating model. But a regional carrier running a 20-person IT team still pays two sets of integrations, benchmarks and regression tests against model updates, and the risk it is buying down is smaller than the one Travelers is.