From cataloging carrier AI announcements across the top 20 U.S. insurers since 2025, three architectural patterns have crystallized, each with distinct implications for how actuarial teams access and validate AI tools. The trade press covered each carrier’s announcement individually: State Farm’s OpenAI Frontier launch in February 2026, Travelers’ Anthropic partnership in January 2026, Allstate’s ALLIE trademark filing in early 2026. No outlet has mapped these moves as a comparative framework showing why the architecture choice, not just the AI investment, determines what actuaries can build on top of it.
That gap matters because architecture is not a technology decision alone. It defines the vendor governance burden, the model validation requirements under ASOP No. 56, the documentation path for state rate filings, and the long-term switching costs when foundation models evolve. A carrier that chose the platform model in 2026 faces a fundamentally different compliance and procurement trajectory than one that built proprietary. This article maps the three models, evaluates their trade-offs with sourced carrier data, and identifies which architecture type gives pricing and reserving actuaries the best tooling path forward.
The Three Architecture Models
The framework rests on one distinction: where the integration layer sits between the foundation model and the carrier’s insurance operations.
In the platform model, the carrier buys into the foundation model lab’s enterprise platform. The lab provides the model, the deployment infrastructure, the governance controls, and the orchestration layer. The carrier builds applications on top. State Farm and OpenAI Frontier represent this approach.
In the partnership model, the carrier forms a direct relationship with a foundation model provider but builds its own integration and orchestration layer internally. The lab provides the model; the carrier owns the platform that connects it to insurance workflows. Travelers and Anthropic, with the carrier’s internally built TravAI platform, represent this approach.
In the proprietary model, the carrier builds the entire ecosystem in-house: the model layer, the orchestration platform, the application interfaces, and the governance framework. External foundation models may serve as components, but the carrier owns and controls the stack. Allstate and ALLIE represent this approach.
Each model carries a different cost structure, a different speed-to-deployment profile, a different vendor risk posture, and a different set of requirements for actuarial model validation. The sections below examine each with specific carrier data.
Platform Model: State Farm + OpenAI Frontier
State Farm became a launch partner on OpenAI’s Frontier enterprise platform on February 5, 2026, alongside HP, Intuit, Oracle, Thermo Fisher Scientific, and Uber. Frontier is OpenAI’s enterprise-grade platform for building, deploying, and managing AI agents with shared context, structured permissions, and centralized governance. For a carrier managing 96 million policies across 19,200 agent offices, the platform model offered something that a raw API integration could not: built-in audit trails, role-based access controls, and a governance framework maintained by the model provider.
Joe Park, State Farm’s Executive Vice President and Chief Digital Information Officer, framed the operational motivation: “We have a systems problem. Too many disconnected tools.” Frontier connects siloed data warehouses, CRM systems, ticketing tools, and internal applications through a single AI orchestration layer. State Farm’s first deployments include Navi (an agent-facing digital assistant for quote pricing and policy details), a claims virtual assistant pilot automating first notice of loss, and Household Story (a customer intelligence tool providing instant summaries of household concerns with tailored product recommendations).
The platform model’s primary advantage is speed to capability. State Farm did not need to build an AI orchestration layer, design a governance framework, or recruit a platform engineering team. OpenAI provides those components. The carrier focuses on building insurance-specific applications that sit on top of the platform, reducing the time from AI strategy announcement to production deployment.
The primary risk is vendor dependency. State Farm’s AI tools run on OpenAI’s infrastructure, use OpenAI’s models, and operate within OpenAI’s governance framework. If OpenAI changes its pricing, modifies its platform architecture, or deprecates capabilities that State Farm’s applications depend on, the carrier has limited recourse. As we documented in our analysis of OpenAI’s 90% carrier AI stack concentration, the systemic risk of multiple top-10 carriers converging on a single foundation model platform extends beyond any individual carrier’s procurement decision.
For actuaries, the platform model means that model governance is partially outsourced to the platform provider. OpenAI maintains the audit trails. OpenAI manages model versioning. When an actuary needs to document which model version produced a pricing indication for a state rate filing, the documentation chain passes through OpenAI’s platform rather than an internal system the carrier fully controls. That dependency creates a new category of vendor due diligence under ASOP No. 56: validating not just the model’s outputs but the platform provider’s governance of the model lifecycle.
Partnership Model: Travelers + Anthropic
Travelers announced its direct partnership with Anthropic on January 15, 2026, deploying personalized Claude AI assistants to nearly 10,000 engineers, data scientists, analysts, and product owners. The partnership is direct: Travelers works with Anthropic, not through a consulting intermediary. The distinction matters because it means Travelers controls the integration architecture, the data flows, and the deployment schedule.
The carrier built TravAI, a secure in-house agentic AI platform that integrates multiple generative AI tools with internal systems. Over 30,000 Travelers employees have access to frontier models through TravAI, and more than 20,000 professionals use AI tools on a regular basis. Mojgan Lefebvre, EVP and Chief Technology & Operations Officer, reported that “since we started introducing personalized Claude and Claude Code assistants, we have seen significantly elevated levels of engineering excellence and meaningful improvements in productivity.”
What distinguishes the partnership model from the platform model is the locus of control. State Farm operates on OpenAI’s Frontier platform. Travelers built its own platform (TravAI) and connects Anthropic’s models to it. This gives Travelers the ability to integrate multiple model providers without depending on any single lab’s platform architecture. Travelers already demonstrates this: the carrier uses OpenAI for its customer-facing agentic voice system in claims (which has enabled consolidation from four call centers to two) and Anthropic for its engineering and analytics stack. The dual-vendor architecture functions as a model risk hedge, ensuring that no single provider’s outage, price change, or capability regression can disable the carrier’s entire AI operation.
The operational results are substantial. Travelers’ claims call center population has been reduced by one-third. Over 50% of all claims are now eligible for straight-through processing, and customers adopt straight-through processing approximately two-thirds of the time. Personal lines renewal underwriting handle time dropped 30%. The expense ratio improved three full points over the past decade, from 31.5 to 28.5, despite increases in technology spend. Underlying underwriting income is more than four times higher than a decade ago.
Travelers backs this with a $1.5 billion annual technology budget, with nearly half directed at strategic initiatives including AI. CEO Alan Schnitzer framed the positioning: “Over the decade, we developed competitive advantage of an innovation skill set. Now we’re bringing all that know-how to Innovation 2.0 at Travelers, powered by AI.”
The partnership model’s primary risk is cost and talent. Building and maintaining an internal AI platform like TravAI requires a large engineering organization. Travelers employs thousands of technologists. Mid-sized carriers with 200-person IT departments cannot replicate this approach without either massive hiring or engaging consulting firms like PwC to bridge the gap.
For actuaries, the partnership model offers the most flexibility. Because TravAI integrates multiple models, actuarial teams can select the best model for each task: one model for claims document summarization, another for pricing model code assistance, a third for regulatory filing review. The multi-vendor architecture also simplifies ASOP No. 56 documentation, because the carrier controls the integration layer and can maintain its own audit trails independent of any single model provider’s platform.
Proprietary Model: Allstate ALLIE
Allstate took the third path, building ALLIE (Allstate’s Large Language Intelligent Ecosystem) as a comprehensive in-house AI platform. The trademark was filed with serial number 99581554 and published January 8, 2026. ALLIE supports both generative and agentic AI use cases across customer engagement, direct policy sales, and claims processing.
The production metrics are specific. ALLIE drafts over 50,000 claim-related messages daily, with 100% of claims adjuster emails now reviewed or generated by AI. The system handles more than 250,000 monthly customer conversations end-to-end, achieving a 75% first-contact resolution rate. Email drafting time for claims adjusters dropped 70%. Billing inquiries fell 45%, and billing escalations dropped 50%. As we detailed in our deep dive on ALLIE’s build-vs-buy economics, complaints about insurance jargon in communications fell 30% after AI-drafted messages replaced manually written ones.
CIO Zulfi Jeevanjee provided context on the communications transformation: “When these emails used to go out, even though we had standards and so on, they would include a lot of insurance jargon. They weren’t very empathetic. Claims agents would get frustrated, and so it wasn’t necessarily great communication.”
ALLIE also extends into software development, with 15% of new coding now handled by AI. Allstate is piloting an AI sales “sidekick” with approximately 30 employees, with plans to expand to 20,000 licensed sales representatives and call center staff. Direct policy sales through AI are already live in three states.
The proprietary model’s advantage is control. Allstate owns the IP, controls the architecture, and can customize every layer of the stack for insurance-specific requirements. There is no platform provider whose roadmap changes could disrupt operations. There is no model vendor whose pricing could force a renegotiation. The carrier answers only to itself on capability decisions, deployment timelines, and governance standards.
The risks are engineering investment and talent retention. Building a proprietary AI ecosystem at this scale requires a permanent, large-scale engineering organization that competes for talent against technology companies. If key engineers leave, the institutional knowledge required to maintain and evolve ALLIE leaves with them. The system also uses OpenAI’s GPT models for claims communications, meaning the “proprietary” label describes the platform and orchestration layer rather than the underlying foundation models. Even a build strategy relies on buy components at the model layer.
For actuaries, the proprietary model creates the tightest integration between AI tools and actuarial workflows but also the highest documentation burden. Because the carrier controls the entire stack, every model version, every training data update, and every algorithmic change is the carrier’s responsibility to document. State regulators reviewing rate filings that incorporate ALLIE outputs will direct their questions to Allstate, not to an external platform provider. That accountability is both an advantage (the carrier can answer definitively) and a burden (the carrier cannot defer to a vendor’s documentation).
Side-by-Side Comparison
| Dimension | Platform (State Farm) | Partnership (Travelers) | Proprietary (Allstate) |
|---|---|---|---|
| Foundation model access | OpenAI via Frontier platform | Anthropic (engineering) + OpenAI (claims voice) | Internal ALLIE + GPT components |
| Integration layer | OpenAI-provided | TravAI (internally built) | ALLIE (internally built) |
| Vendor lock-in risk | High (single platform) | Low (multi-vendor, internal orchestration) | Low (self-owned), but model-layer dependency persists |
| Speed to first deployment | Fastest (platform provides tooling) | Moderate (internal platform build required) | Slowest (full stack construction) |
| Engineering talent required | Application-layer only | Platform + application teams | Full-stack AI engineering org |
| Annual tech investment | Not disclosed | $1.5B (nearly half strategic) | Not disclosed |
| ASOP No. 56 documentation path | Depends on platform audit trails | Internal audit trails, multi-model validation | Full internal documentation chain |
| Regulatory filing control | Shared with platform provider | Carrier-controlled | Carrier-controlled |
| Key production metric | 19,200 offices on Navi | 30,000+ TravAI users; 50%+ STP-eligible claims | 50,000+ AI-drafted messages/day |
Hybrid Approaches: Chubb and AIG
Not every top carrier fits neatly into the three-model framework. Chubb and AIG each demonstrate hybrid approaches that combine elements of multiple architectures.
Chubb announced in December 2025 what CEO Evan Greenberg called a “fundamental reshaping of the operating model.” The targets are aggressive: approximately 85% of major underwriting and claims processes automated, a similar share of global gross written premium flowing through fully digital or digitally-enabled channels, and a headcount reduction of approximately 20% over three to four years (primarily through natural turnover). Chubb projects run-rate expense savings of roughly 1.5 combined ratio points. With 3,500 engineers globally and expanding hubs in Mexico, Greece, India, and Colombia, Chubb is building substantial internal engineering capacity. But the carrier also invests in algorithmic tools and large language models from external providers, creating a hybrid of partnership and proprietary elements. Chubb’s 2025 financials provide the investment base: $54.8 billion in net premiums written (up 6.6% year over year), P&C underwriting income of $6.5 billion (up 11.5%), and a record-low combined ratio of 85.7%.
AIG operates through AIG Assist, a multi-agent orchestration system that has processed over 370,000 submissions and is on track toward a 500,000 goal by 2030. AIG partnered with Palantir for its Lloyd’s Syndicate 2479, using LLMs to match program data with the syndicate’s risk appetite. The carrier deployed a patent-pending “Auto Extract” capability for document processing alongside a digital twin of AIG’s processes, workflows, and data elements. CEO Peter Zaffino reported results that “exceeded expectations”: Lexington E&S submissions grew 26% year over year, middle market property submit-to-bind ratios improved 35%, and the private not-for-profit segment processes 100% of applicable submissions through AIG Assist. Zaffino noted: “We’re seeing a massive shift in our ability to process a significant submission flow way beyond our expectations without additional human capital resources.” AIG’s approach combines external foundation models (through Palantir and Anthropic) with a proprietary orchestration layer, placing it between the partnership and proprietary categories.
The hybrid examples reinforce a practical reality: few carriers will adopt a pure version of any single model. The framework describes dominant strategies, not exclusive ones. Even Allstate’s “proprietary” ALLIE uses external GPT models for specific functions. Even State Farm’s “platform” approach will likely incorporate internal tooling as the deployment matures. The value of the three-model framework is directional: it identifies where the center of gravity sits in each carrier’s AI architecture decision.
The Consulting Channel as a Fourth Path
Carriers that lack the engineering scale of Travelers or the strategic commitment of Allstate face a practical question: how to access foundation model AI without building a platform or buying a platform subscription. The Big Four consulting firms have positioned themselves as the answer.
PwC’s expanded alliance with Anthropic, announced May 14, 2026, will certify 30,000 PwC professionals on Claude through a joint Center of Excellence. The insurance-specific claim is specific: underwriting cycles compressed from 10 weeks to 10 days, with delivery time improvements of up to 70% across production deployments. Anthropic CEO Dario Amodei cited the insurance use case directly. Separately, Deloitte, Accenture, and KPMG have announced their own model partnerships, collectively putting over a million consulting professionals on a path to Claude or GPT access.
For mid-market carriers with annual AI budgets under $25 million (what BCG research characterizes as the “broad experimentation” tier), the consulting channel offers a faster on-ramp than building TravAI or ALLIE internally. The trade-off is that consulting-delivered AI introduces a third party into the governance chain: the model comes from Anthropic or OpenAI, the configuration comes from PwC or Deloitte, the data comes from the carrier, and the regulatory responsibility belongs to the appointed actuary. That layered accountability creates ASOP No. 56 compliance complexity that pure platform or proprietary architectures avoid.
Microsoft and Cognizant quantified the broader scaling challenge in their February 2026 report on agentic AI in insurance: only 7% of insurers have successfully scaled AI initiatives across their organizations, and 70% of scaling challenges are organizational rather than technical. Celent’s 2026 survey found 48% of insurers running GenAI in production, but the gap between “in production” and “at scale” remains wide. The consulting channel exists precisely to bridge that gap for the carriers that cannot bridge it internally.
Actuarial Implications by Architecture Type
The architecture choice creates specific, different consequences for actuarial teams across four dimensions.
Tool Access and Availability
Platform model actuaries (State Farm) get tools as fast as the platform provider ships them, but only tools the provider builds or enables. If OpenAI Frontier adds a rate filing assistant, State Farm actuaries can use it immediately. If the feature does not appear on the Frontier roadmap, State Farm actuaries must build it themselves on top of the platform, competing for internal development resources with every other department.
Partnership model actuaries (Travelers) have access to multiple foundation models through TravAI, giving them flexibility to select the best model for each actuarial task. The trade-off is longer lead times: every new AI capability requires internal platform integration before actuaries can use it.
Proprietary model actuaries (Allstate) have the most customized tools, purpose-built for insurance workflows. But they also face the longest wait for capabilities that require engineering resources, and they cannot independently access external model improvements without routing through the ALLIE platform team.
Model Validation Under ASOP No. 56
The American Academy of Actuaries has determined that “GenAI is a model; thus ASOP No. 56 applies.” The standard requires actuaries to select models that “reasonably meet the intended purposes,” understand “known weaknesses in assumptions and methods, limitations of the data,” and “take steps to mitigate model risk.” When relying on externally developed models, actuaries must “make a reasonable attempt to have a basic understanding of the model.”
The validation burden varies by architecture. Under the platform model, the actuary must validate both the model’s outputs and the platform’s governance of the model lifecycle, including version control, access permissions, and audit logging. Under the partnership model, the carrier controls the integration layer, giving the actuary direct access to audit trails and version documentation. Under the proprietary model, the actuary has full visibility into the entire stack, but the carrier must maintain all documentation internally with no external provider to share the burden.
ASOP No. 23 on Data Quality adds a parallel requirement: actuaries must evaluate whether AI-obtained data is “appropriate, sufficient, and reasonable” and validate “refreshed GenAI data sources each time they are generated,” because inputs and outputs may continually change. The architecture determines whose infrastructure generates those data refreshes and whose documentation the actuary must review.
Rate Filing Documentation
When a carrier files rates that incorporate AI-derived pricing factors, state regulators will ask how those factors were developed, validated, and monitored. The platform model creates a documentation chain that passes through the external platform provider. The partnership and proprietary models keep the documentation chain internal. For actuaries preparing rate filings in states with active AI oversight (Colorado’s SB 21-169 requirements effective by mid-2026, the NAIC 12-state AI evaluation pilot), internal documentation control provides a faster path to answering regulatory questions than depending on an external provider’s disclosure policies.
Vendor Governance and AM Best Scrutiny
AM Best survey data shows that 68% of insurers outsource AI capabilities, but only 18% track vendor risk systematically. The architecture choice directly affects this gap. Platform model carriers have a single, critical AI vendor to govern. Partnership model carriers have multiple vendors but control the integration. Proprietary model carriers have minimal vendor governance exposure at the platform level but must still govern the component foundation models they incorporate. As carrier AI projects increasingly fail at the audit layer rather than the technology layer, the governance burden of each architecture becomes a material factor in enterprise risk management assessments.
Why This Matters
The three-model framework is not permanent. Platform carriers may build internal capabilities over time. Partnership carriers may simplify to a single vendor. Proprietary carriers may adopt external platforms for specific functions. The current moment is significant because the architecture decisions made in 2025 and 2026 will constrain each carrier’s AI options for the next three to five years. Switching from one model to another requires rebuilding integrations, retraining staff, renegotiating vendor contracts, and re-documenting model governance for every affected regulatory filing.
For the broader industry, the architecture split creates a natural experiment. Within two to three years, we will have enough production data to measure whether platform, partnership, or proprietary approaches produce better actuarial outcomes: faster rate filing approvals, lower loss ratios from improved pricing, reduced LAE from claims automation, and higher AM Best governance ratings. Celent’s finding that 22% of insurers plan agentic AI deployment by year-end 2026 suggests the architecture decision is accelerating across the market, not just at the top five carriers.
The carriers that have not yet chosen an architecture face a closing window. Foundation model labs are building direct carrier sales organizations. Consulting firms are training tens of thousands of practitioners. The gap between AI-deployed carriers and AI-aspirational carriers widens with every quarter. The architecture question is no longer whether to adopt AI. It is which integration model best positions the carrier’s actuarial and underwriting teams for the next cycle of model capability improvements, regulatory requirements, and competitive pressure.