Why Insurance Carriers Are Consolidating Their Generative AI Stacks

U.S. enterprises spent $37 billion on generative AI in 2025, more than three times the $11.5 billion they spent in 2024, with financial services and insurance among the heaviest investors. Much of that capital, though, has produced a fragmented operating reality. Carriers have built copilots inside claims, chatbots inside service, summarization tools inside underwriting, and search assistants inside compliance, each running on its own architecture and its own contracts. Three-quarters of U.S. insurance firms have now deployed generative AI in at least one business function, yet the maturity of those deployments varies sharply. What looked like rapid adoption a year ago is revealing itself as something more difficult: a stack that scales horizontally without consolidating vertically.

Abhishek Kumar, Corporate Vice President at New York Life and a member of the IEEE Computer Society, has spent his career at the boundary between AI product strategy and regulated enterprise architecture. With more than fourteen years of experience across financial services and insurance, his recent focus has been on the platform problem: how a Fortune 100 mutual life insurer with millions of policyholders can adopt generative AI across multiple business units without building each capability twice. The architecture he and his team have shipped now powers AI agents across underwriting, service experience, finance, marketing, field experience, and corporate strategy from a shared foundation.

The Cost of AI Sprawl Across the Insurance Stack

78% of enterprises now struggle to integrate AI with their existing systems, per a 2025 survey of more than 500 large-enterprise U.S. leaders. 28% are already running more than ten AI applications, and 76% have experienced at least one negative outcome from disconnected AI tools, ranging from redundant spend to manual data transfers between systems that were supposed to talk to each other. McKinsey's 2025 analysis of insurance carriers reaches a parallel conclusion: carriers that struggle to scale AI share one trait, which is the inability to create components reusable across business lines. Each unit ends up with its own model selection, its own retrieval logic, its own evaluation harness, and its own integration code. The duplication stays invisible inside any single deployment but compounds quickly at the level of the enterprise.

This is the operating reality Kumar inherited. Inside large carriers, individual product owners had already shipped functioning AI tools, often with strong local results. The problem was structural rather than technical. Underwriting had built one assistant, service had built another, and field advisory had built a third, with no shared retrieval layer, no shared compliance controls, and no shared way of evaluating whether the tools were working in production. Each project became its own twelve-month engineering cycle, repeating most of the work that had just been done somewhere else in the building. Kumar's view was that the next gain for insurance AI would not come from a smarter individual assistant. It would come from removing the cost of building the next twenty assistants.

"The first wave of insurance AI proved that the use cases were real. The second wave has to prove that the infrastructure underneath them is reusable," Kumar says. "If every business unit has to rebuild the foundation, the technology never reaches the scale where it changes how the carrier operates."

From Point Solutions to a Single GenAI Foundation

The structural answer is now visible across the largest enterprise software analysts. Gartner predicts that by 2027, 70% of organizations with platform teams will include generative AI capabilities in their internal developer platforms, turning AI capabilities into self-service primitives rather than per-application builds. A late-2025 TechCrunch survey of 24 enterprise-focused venture capitalists found broad agreement that 2026 will be the year enterprises consolidate AI spending into fewer contracts and fewer winners, after a two-year period of parallel experimentation. The shift mirrors what happened to internal developer platforms between 2018 and 2022, when organizations moved from per-team Kubernetes clusters to shared platform engineering. The same logic now applies to retrieval, prompting, evaluation, and guardrails.

Kumar's response is a centralized retrieval-augmented generation platform that operates as a single foundation for AI agents across the enterprise. He led the product strategy and end-to-end implementation, building a suite of proprietary libraries that allow individual business units to deploy specialized agents on top of a common base. The platform handles knowledge ingestion, retrieval, model selection, and output controls as shared services. Business units bring their domain knowledge, their use cases, and their workflow integrations. Two of the production agents already running on the platform, "UW Assist" for underwriting and "Suitability IQ" for the field and advisory channel, were built using the same underlying libraries, drawing on the same compliance primitives, and evaluated against the same internal benchmarks.

"The mistake most large institutions make is treating each GenAI use case as a one-off product. That is how you end up with twenty assistants and twenty stacks underneath them," Kumar observes. "A platform forces a different question. What does every agent in this company need, and how do we ship that once?"

A Library Model That Lets Business Units Build Independently

The pace at which insurers are moving into generative AI makes the architectural question urgent. Celent's 2025 GenAI-oneers report found that 44% of insurers now have active generative AI deployments, up from 28% in 2024 and 8% in 2023. Underwriting, customer service, distribution, and policy administration each have multiple emerging use cases, and the average carrier is no longer choosing whether to deploy. They are choosing how many systems to deploy in parallel and how those systems will share data, evaluation harnesses, and audit trails. Without a shared library layer, every new agent demands its own integration project. With one, the cost of the next agent is a fraction of the cost of the first.

The library architecture is the part of the platform that lets six business units, Finance, Service Experience, Underwriting, Marketing, Field Experience, and Corporate Strategy, build their own AI agents without writing common infrastructure from scratch. Kumar, a Stevie Award winner in the Product Development / Management Executive of the Year category at the American Business Awards, approached the design with a single product principle: the components every business unit will need should be written once and owned by the platform team, not rebuilt by each business line. Inside the platform, that principle produced a clean split. The retrieval pipeline, the prompt and response orchestration, the response evaluation harness, and the PII redaction layer are written once and exposed as libraries. Each business unit consumes them, configures them for its own domain, and adds the workflow integrations specific to its product. The internal time-to-market for a new AI agent dropped from twelve months to three months, a 75% compression that scales as more business units come onto the platform.

"A platform is not a piece of software. It is an agreement about which problems get solved once and which get solved many times," Kumar notes. "If you get that contract right between the platform team and the business units, the rest of the velocity follows."

Why Cross-Functional Teams Outperform Centralized Builds

Platform success is as much an organizational outcome as a technical one. Deloitte's 2026 research on AI team structure found that cross-functional teams, those working across multiple business units, were 30% more likely than others to report significant gains in efficiency and innovation from using AI. The same study found that larger, mixed-skill teams report markedly higher rates of AI usage and effectiveness compared with smaller, single-discipline teams. The signal is consistent across industries. AI value compounds when product, engineering, data science, and design move together inside a single team, rather than when those disciplines hand work between siloed groups.

Kumar's platform team is structured to that pattern. The build runs across more than fifteen data scientists, AI engineers, data engineers, and UI partners, working in a single product organization rather than across handoffs. His published Forbes piece, How To Choose Enterprise AI Tools That Deliver Real Value, makes the same case from a broader angle: the tool decision is downstream of the operating-model decision, and enterprises that fail to align team structure, data ownership, and evaluation discipline rarely get value from any tool, no matter how capable. The cross-functional model is what allows the platform to ship a new business-unit-specific agent in three months instead of twelve. A data scientist working on retrieval quality sits next to an engineer building the orchestration layer, who sits next to a product manager negotiating the workflow integration with a specific business unit. Decisions that would otherwise sit in three different inboxes get made in the same room.

"You cannot run an AI platform on a service-ticket relationship between teams. The whole point of the platform is that the business unit consuming it is also part of how it gets built," Kumar reflects. "When we ship a new library, the people who needed it are the ones who validated it."

From Shared Retrieval to Shared Execution

The next stage of enterprise AI is already visible in the analyst data. Gartner's 2026 Hype Cycle for Agentic AI reports that only 17% of organizations have deployed AI agents to date, while more than 60% expect to do so within the next two years, the steepest adoption curve recorded for any emerging technology in the firm's CIO survey. Banking and insurance lead the production rate at roughly 47% as of early 2026, with that figure projected to reach 63% by 2027 across the same sectors. The category is shifting from read-only retrieval, where AI systems surface information for a human to act on, toward read-write execution, where AI systems take operational actions inside workflows under defined controls.

That shift is where Kumar is now focused. The same library architecture that powers retrieval today extends naturally into the agentic phase, because the policy enforcement, the audit trails, and the evaluation harnesses already exist as platform primitives. The team is moving its existing AI agents from read-only assistance into read-write workflows that target documented full-time-equivalent savings and productivity gains across the carrier's advisors, agents, and back-office staff. Kumar is also leading proofs of concept with hyperscaler partners on select use cases, informing the enterprise agentic AI strategy from the inside out. The bet is that the platform built for retrieval will compress the next architectural transition the same way it compressed the first.

"The reason platforms matter is that they make the next change cheaper than the last one," Kumar explains. "When agentic systems become the operating layer of an insurance company, the carriers that built the foundation in 2025 will be the ones who actually use them in 2027. The rest will be rebuilding the basement while everyone else is shipping the second floor."