An IA Capital Group survey of 36 senior carrier technology leaders, published May 6, 2026, found OpenAI present in approximately 90% of carrier AI technology stacks and Google Gemini in none of the production deployments.

The concentration arrives as carrier AI crosses into production, with deployment up from 37% to 61% of respondents in a year. An industry whose business is pricing correlated exposure has built its operating layer on a single point of correlation.

Key Takeaways

  • Roughly 90% of surveyed carrier AI stacks include OpenAI, whether through the API directly or embedded inside vendor tools, while Google Gemini appears in zero production deployments.
  • 70% of carriers spent under $500,000 on AI in the past year, against a switching cost estimated near $409,000 for a single workload, which makes migration uneconomic before the merits are considered.
  • Only 8% believe they lead peers on AI while 70% expect competitive advantage within three years, a gap that defers vendor selection to whoever has the widest adoption base.
  • A vendor model update reaches every carrier at once. Where AI outputs feed claims triage or reserving, the resulting behavioural shift is directionally correlated across the market rather than idiosyncratic.
  • Exhibit A of the NAIC pilot will make this a regulatory dataset, since it asks 12 states' carriers to quantify AI usage including vendor-embedded models.

What the Survey Found

The survey covers 36 senior technology leaders at national and regional carriers. It is not a census, but the findings are consistent and the seniority is high.

OpenAI appears in approximately 90% of stacks, counting both direct API use and models embedded inside third-party tools. Gemini appears in none. The path there was ordinary: first-mover advantage on GPT-series models through 2023 and 2024, enterprise sales into financial services, and proofs of concept that moved to production without anyone revisiting the vendor decision.

The supporting figures explain why nobody revisited it. Seventy percent of carriers spent under $500,000 on AI in the past year, a level that buys point solutions rather than parallel vendor evaluations. Only 8% believe they are currently ahead of peers, while 70% expect at least moderate competitive advantage within three years.

Four of the top five production use cases are document processing: reading, extracting, summarizing, classifying. Underwriting deployment stands at 56% and claims at 50%, more limited in scope but the functions where model behaviour carries the most weight.

Commodity Tasks, Non-Commodity Integration

Document processing should be the easiest workload to move, which is what makes the lock-in worth explaining.

Insurance document processing is not generic. The systems ingest ACORD forms, loss runs, policy endorsements, claims files, medical records and regulatory filings, each with its own formatting and extraction requirements. What sits around the model, prompt templates calibrated per document type, parsing logic tuned to one vendor's response formatting, error handling built around that vendor's failure modes, and quality assurance validated against its output distributions, is carrier-specific work that does not transfer.

The cost of rebuilding it is where the survey's spending figure bites.

Cost Category Estimated Range Driver
Engineering rewrite $216,000 ~1,200 hours at $180/hr for API integration, prompt re-engineering, output parsing
Dual-run infrastructure $60,000 Running old and new models in parallel during validation period
Data movement and transformation $25,000 Reformatting training data, evaluation datasets, and test suites
Revalidation and testing $40,000 Model performance benchmarking, edge case testing, regression suites
Risk buffer (20%) $68,200 Unforeseen integration issues, extended timeline overruns
Total per workload ~$409,200

At roughly $409,000 per workload against 70% of carriers spending under $500,000 a year on AI in total, a single migration consumes most of an annual budget. Executives do not price it that way: a Zapier survey of 542 US executives found nearly 90% believed they could transition AI vendors within four weeks and 41% within one week.

For an actuary the more consequential exposure is not cost but correlation. When a vendor updates a model, the update reaches every carrier using the API at the same moment. A change in how ambiguous claims documents are classified, or in probability calibration on a risk score, arrives simultaneously across the market rather than at one carrier's chosen release date.

That produces correlated movement in outputs feeding reserving and adjudication, which is the structure of an accumulation exposure rather than an operational risk. Cyberwrite CEO Nir Perry put it in February 2026 as a failure at one of three or four dominant vendors cascading across hundreds of millions of businesses at once, "creating an accumulation exposure the insurance market has no historical framework to price or contain." The same concentration that a carrier would decline to write is sitting inside its own operations.

The Disclosure Turns It Into a Regulatory Question

What changes the position is not a market event but a filing requirement.

The NAIC's AI Systems Evaluation Tool pilot launched March 2, 2026 and runs through September across 12 states. Exhibit A asks carriers to quantify AI usage across functional areas including vendor-embedded models, which produces the first regulatory dataset on vendor concentration in the industry. When the same vendor appears across most filings, the concentration stops being a trade-press observation.

The other exhibits then apply pressure in sequence. Exhibit B assesses governance, so a single-vendor carrier has to show that its framework addresses vendor dependency, continuity, and model update management. Exhibit C covers high-risk systems in claims, underwriting, pricing and fraud detection, where validation documentation has to address drift monitoring after vendor updates. Exhibit D examines data sources and discrimination risk, and shared models make correlated bias a live question: a bias in the vendor's training data propagates to every carrier using it.

Underneath all four sits the principle that carriers bear full responsibility for third-party systems. A vendor's governance practices, training data provenance and update procedures are effectively subject to insurance regulatory review through its customers, without the vendor being a regulated entity.

The disclosure lands on carriers least equipped to answer it. Grant Thornton's 2026 survey found 68% saying AI controls exist but are fragmented across teams and tools, with only 24% fully confident in those controls; Capgemini found 42% of insurers tracking no AI metrics at all and 55% reporting unclear ROI. A carrier that does not benchmark its current vendor's performance cannot document why it selected that vendor, cannot detect drift after an update, and cannot demonstrate the alternative it would use if the service failed. That is three Exhibit answers missing for the same reason.

Further Reading

Sources