The top 10 insurers in the world employ 56% of all AI talent in the insurance industry. That is not a distribution curve; it is a verdict on the state of the industry.

While there are thousands of carriers globally, specialized expertise is aggressively clustered within this tiny elite. Compensation is only part of the story; the real divide is environmental. AI talent has surveyed the broader market and concluded that the vast majority of organizations lack the unified data infrastructure and research-led culture required for serious technical work to survive.

This article is the third in a series. The first, “Why Insurance Telematics Integrations Fail,” (Carrier Management, November 2025) documented technology integration failures. The second, “Why ‘Good Enough’ Is Killing Insurance,” (Carrier Management, January 2026) identified satisficing, the decision-making pattern that produces them. This article examines the consequence that neither piece addressed: what happens to the talent pipeline when an industry’s technology decisions become visible to the people it needs to hire.

Before the diagnosis, note two different prediction problems we should not conflate. The first: Which B2B infrastructure vendors selling to insurance companies will succeed? Carriers have real influence here. Their purchasing decisions shape which vendors survive. The second: Which AI technologies will matter and how they will reshape the competitive landscape. Carriers have almost no influence here. Insurance is one use case among many. For some technology companies, insurance is not even a priority market.

Both things can be true: AI is a bubble and AI will be transformative. That debate is interesting but not useful for organizational decision-making. We cannot accurately predict which AI technologies will matter five years from now. But we can predict how organizations will respond to those technologies based on their structural characteristics. The talent market is already making that prediction.

The Tech Stack as Talent Filter

Recruiters and hiring managers across the insurance industry report a consistent pattern: candidates ask about technology platforms, engineering culture, and innovation commitment early in the hiring process. This is not limited to candidates for technical jobs. Claims professionals, actuaries, underwriters, product managers, and marketing candidates from outside the industry all screen on tech stack. The question they are really asking is not about technology culture in the abstract. It is whether the infrastructure will let them do their job at the level they know is possible.

The actuary who has worked with real-time modeling does not want to rebuild analyses in spreadsheets because the core system cannot export anything else. The underwriter who has seen what integrated data looks like does not want to wait for a PDF, a limitation the first article in this series documented, where telematics data was available but delivered only as a static document. As I wrote in the second article, hiring world-class talent into infrastructure that constrains what they can do is bolting a Ferrari engine to a covered wagon. The candidates can see the covered wagon in the interview.

The language confirms it. “We built” signals engineering culture. “We implemented” signals purchasing culture. “We deployed” signals vendor management. Candidates from companies where technology is the business hear the difference immediately.

The compensation picture makes it worse. U.S. insurance technology salaries run as much as 35%-45% below tech sector equivalents for technology roles. Carriers that cannot compete on compensation or engineering culture have eliminated both levers.

Everyone in the insurance industry knows the next number. Approximately 400,000 insurance professionals will retire or leave the industry over the coming decade. That figure has appeared in so many conference presentations that it has lost its force. But here is what gets lost in the repetition: the generation replacing those hundreds of thousands of professionals screens on technology harder than the generation leaving. The retirement wave is not just a headcount problem. It is an accelerant for everything that follows. (Editor’s Note: The 400,000 number is outdated.)

That 56% concentration is the scorecard. Nearly 13,600 of the industry’s approximately 24,000 AI professionals chose those 10 carriers because they built something worth working on. (Editor’s Note: Employee counts are approximate. They were estimated from a graph showing AI professionals employment numbers for the top 10 carriers in “Evident AI Insurance Index: Key Findings Report 2025,” published in June 2025)

The CIO as Purchasing Agent

When the CIO’s primary function is vendor management, the entire technology organization becomes a procurement operation. Job descriptions say “engineer” or “architect” but the actual work is vendor coordination. The candidate from a technology company figures this out in the first interview.

Here is the uncomfortable part. A CIO who keeps the lights on, manages vendor relationships, and delivers projects on time and on budget is doing the job as defined. The organization evaluated its technology needs and decided vendor management was sufficient. That is a reasonable judgment. It is also the judgment that satisficing theory, the framework from the second article in this series, would predict.

The purchasing-agent CIO excels at the first prediction problem: which vendors to buy from, which platforms to deploy, which integrations to prioritize. The mistake is assuming that competence at vendor selection, where carriers have real influence, translates to competence at evaluating which AI technologies will reshape the industry, where insurance is one use case among many.

Some observers within the industry have argued that existing systems are “modern enough.” That assessment makes sense for the vendor selection problem. The talent market is evaluating a different question entirely: Does this organization have the capacity to absorb technological shifts it cannot control? The two questions have different answers.

The person considering a role at a carrier that operates this way sees an organization where technology is overhead, not strategy. They see a ceiling. They see that the interesting problems have been outsourced to vendors. They see that the 20% of capability that would differentiate the carrier does not exist because no one in the building was asked to build it.

For 25 years, insurance carriers bought most of their core capabilities from vendors, perhaps 80%, using the familiar 80-20 split. That means they built only the remaining 20%: proprietary underwriting models, unique claims workflows, agency-channel logic. That worked when the 80% was the commodity and the 20% was the value. AI inverts the relationship. The differentiating AI applications require exactly the engineering capability that lives in that 20%.

Once a carrier decides that wrapping and reskinning vendor platforms is the strategy, it has decided its technology organization is an integration shop. The people who thrive in that environment are vendor managers, not builders. The builders leave or never arrive. When the carrier needs AI capability that cannot be purchased, there is no one in the building who can build it. The purchasing-agent model has consumed the engineering capacity that would be needed to move beyond it.

The timeline makes this worse. As the first article in this series documented, Chief Information Officer tenure averages less than 5 years in the insurance industry while core system implementations take two to three. The CIO who champions the purchasing-agent model is often gone before anyone measures the consequences. The next CIO inherits the talent gap, misdiagnoses it as a technology gap, and reaches for the same playbook.

The cycle does not just repeat. It entrenches.

The AI Production Gap

The numbers have become familiar: roughly 70% of insurers are piloting AI, 22% have it in production for some workflows, and only 7% have achieved enterprise-scale deployment. What has not become familiar is what those numbers actually describe, because carriers are not making these decisions in a vacuum.

On one side, vendor reality distortion machines are running at full capacity. When a technology company CEO tells an industry to adopt his product or face collapse, that is a sales pitch, not insight. Tech vendors have run this play for decades. Nobody ever got fired for buying IBM, and every technology company since has wanted to be IBM: manufacture the existential threat, position the product as the safe choice, let risk aversion close the sale. The mechanism leverages groupthink: a few early buyers move, the rest follow because the risk of being wrong alone outweighs the risk of being wrong together. The FOMO feeds the satisficing decision from the second article in this series: buy the tool, check the box, announce the transformation.

On the other side, critics will tell you the entire AI movement is a fraud. Push back with a specific example of measurable results and the critic’s position shifts: That is not really AI. The goalpost is designed to move. The first claim gets the attention. The second handles any evidence that contradicts the first. Between the two positions, AI can never succeed by definition. The critic’s business model depends on the controversy, not the resolution.

Most operators work between these two poles of vendor distortions and critics’ pronouncements. Neither extreme should be driving operational decisions. Carriers are making adoption decisions between two noise machines, neither of which is producing useful signal for operational decisions.

The 93% that have not scaled are not failing because AI does not work. They are failing because production AI requires real-time data pipelines, cross-system orchestration, governance built into the architecture, and human oversight capacity. These are engineering capabilities, not vendor purchases. The carriers stuck between piloting and production are stuck because they are trying to buy their way across a gap that can only be built across.

AI agents are about to stress-test everything carriers built or failed to build in the prior decade. Carriers that built the production stack will deploy agents. Carriers that did not will watch.

The human-in-the-loop requirement makes the talent problem explicit. Production AI systems require people who can customize models, evaluate AI outputs, debug failures, and exercise judgment about when to override. These people need both domain expertise and technical literacy. The retirees walking out the door are the domain experts. The technical talent that could partner with them was screened out by the tech stack years ago.

This is the Retirement-AI Double Bind: Domain experts are leaving before AI systems are ready to capture their knowledge. The window is narrow, and carriers that lack engineering talent cannot build the extraction systems in time. Every month of delay is institutional memory that leaves permanently.

The performance gap is already measurable. WTW’s research found that P/C insurers using sophisticated analytics achieved combined ratios six points lower than slower adopters between 2022 and 2024. Six points of combined ratio is not a rounding error.

The Compounding Effect

Each of these problems is serious on its own. Together they create something worse.

The carrier decides the tech stack is “modern enough.” The tech stack signal screens out candidates across all functions. Without engineering talent, the carrier cannot build AI production infrastructure.

What makes this different from a generic cycle is that each stage removes the capacity to reverse the prior stage. You cannot fix the tech stack without engineering talent. You cannot attract engineering talent without a tech stack worth working on. You cannot build production AI without both. Each iteration does not just maintain the problem. It consumes the resources that would be needed to break it.

Organizations deep in this cycle see AI code generation tools and think they have found the escape hatch. They have not. Code generated by people who cannot evaluate, debug, or maintain it is worse than the vendor management problem it was supposed to solve. When vendor software breaks, you call the vendor. When AI-generated code breaks and no one in the building understands it, there is no one to call. AI makes good engineers faster. It does not make non-engineers into engineers.

What Operators Are Doing About It

The cycle is not inevitable. It is predictable for a specific organizational archetype. Operators who recognize the pattern have practical options. Both FOMO-driven adoption and paralysis are failure modes. The path is deliberate engagement.

Play the game. AI is real, it is accelerating, and opting out is not a viable strategy. But FOMO-driven adoption is equally dangerous. Purchasing-culture organizations see the marketing, see the competitor announcements, and make the satisficing decision to buy the tool, check the box, announce the transformation. This is the two-prediction-problem conflation: treating questions about which AI technologies will reshape the industry as vendor selection exercises.

Deploy AI to production and don’t wait for certainty. Build learning into operations. Build workflows and data pipelines that are not dependent on any specific AI model or vendor. Insurance is not the market that will determine which models survive.

Document everything. Proficiency in AI comes from daily use, but the learning compounds only if it is captured.

Document what works, what fails, what the tools can and cannot do. Patterns that are not obvious today may become critical when the architecture shifts. Maintain audit trails, prompt logs, code review records, and architectural documentation.

Make sure you have the right people. Engineering judgment to supervise AI output. Architectural thinking to design systems that AI implements. Security awareness to catch what AI misses. These are human capabilities that become more valuable as AI generates more code, not less.

If the job is vendor management, hire vendor managers and build the best vendor management operation in the industry. If the job is engineering, hire engineers and give them an environment where they can build. The failure is hiring for one and expecting the other.

The instinct may be to reach for upskilling: train existing staff on AI tools and close the gap. But training people to use AI tools inside an organizational model built for vendor management produces better-trained vendor managers, not engineers.

The capability gap is not a training problem. It is an organizational design problem. In this environment, the stakes of getting the diagnosis wrong means some organizations will not be viable at all.

Keep learning. The capability to learn, evaluate, and adapt is more durable than any specific tool adoption. There is serious debate about whether large language models are the enduring AI architecture or a transitional one. Small language models and physical AI suggest the landscape may shift fundamentally. These are shifts carriers cannot influence and did not choose.

Carriers building flexible, model-independent infrastructure will absorb them. Carriers locked into vendor-specific implementations will face another round of the same cycle.

The Verdict

The first article in this series showed that technology integrations fail. The second showed why organizations choose “good enough” instead of fixing them. This article shows the consequence: the talent market has already rendered its verdict on those decisions, and the verdict is self-reinforcing.

“‘We built’ signals engineering culture. ‘We implemented’ signals purchasing culture. ‘We deployed’ signals vendor management. Candidates from companies where technology is the business hear the difference immediately.”

The industry cannot determine which AI technologies will matter. That is the second prediction problem, and insurance is not the market that decides. AI may be a bubble. AI may be transformative. Both can be true, and neither resolves the organizational question.

What the industry can determine is how it responds. That is organizational behavior, and it is predictable from structural characteristics. The purchasing-agent model will satisfice. Engineering cultures will measure and adapt. The talent market will continue to sort candidates accordingly.

The carriers that recognize this pattern have a window. The carriers that do not will discover that the talent market made the decision for them.

AI Disclosure: Research for this article utilized AI tools to discover and verify publicly available data sources and citations. All analysis, interpretation, conclusions and writing are original work by the author based on 20 years of operational experience leading telematics data integration programs at @Road/Trimble and working with major commercial auto insurance carriers.

Featured images: AI-generated (ChatGPT)