Grant Thornton Survey Exposes the Insurance AI Proof Gap

From reviewing four major AI adoption surveys published in Q1 2026 alone, the Grant Thornton data stands out because it quantifies the gap between stated revenue impact and provable governance controls, something the broader surveys have treated as a simple adoption question. Adoption rates are everywhere: 82% in Sedgwick's claims data, 61% in Datos Insights' carrier production numbers, 78% in Bain's generative AI figures. What those numbers lack is a corresponding measure of whether any of those deployments could withstand scrutiny from a regulator, auditor, or board committee demanding evidence of responsible use.

Grant Thornton's 2026 AI Impact Survey, fielded between February 23 and March 18 with 100 insurance-specific respondents drawn from a total of 950 business leaders, introduces the concept of an "AI proof gap." The term describes the distance between what organizations claim AI does for them and what they can demonstrate under examination. For insurance, that distance is wider than any other sector in the survey, and it carries consequences that reach from regulatory enforcement to actuarial reserve adequacy.

This article dissects the insurance-specific findings, benchmarks Grant Thornton's data against four other 2026 surveys (AM Best, Capgemini, Datos Insights, and Bain), maps the proof gap onto the NAIC's 12-state AI Evaluation Tool pilot, and explains why the actuary's role in AI governance is shifting from optional specialization to unavoidable professional responsibility.

Defining the Proof Gap: 52% Revenue Claims vs. 24% Audit Confidence

The headline numbers from Grant Thornton's insurance subsample tell a story of enthusiasm outpacing infrastructure. Among the 100 insurance executives surveyed (CFOs, CIOs, COOs, and VPs), 52% report that AI has contributed to measurable revenue growth. That figure sits 15 percentage points above the cross-industry average. Sixty-two percent report improved decision-making insights from AI tools. Fifty percent report direct cost reductions. And 62% rate their organization's AI maturity as "scaling across multiple functions," suggesting these are not pilot-stage programs.

Set against that optimism: only 24% describe themselves as "very confident" their organization could pass an independent review of AI governance and controls within 90 days. The remaining 76% acknowledge some combination of fragmented controls, incomplete documentation, or untested incident response plans.

That 28-point spread between revenue attribution (52%) and audit readiness (24%) is the proof gap. It is not an abstract governance concept. It represents concrete regulatory, reputational, and financial exposure for every carrier operating AI systems in underwriting, claims, or pricing without the documentation to defend those systems under examination.

Tom Puthiyamadam, Grant Thornton's managing partner for advisory services, framed the pattern in terms that resonated across the full 950-respondent dataset: "Companies are making tremendous investments into AI and yet, we're not seeing that correlate with an increase in AI accountability." For insurance specifically, where regulatory examination is more frequent and more granular than in most industries, that statement carries particular weight.

Inside the 76%: Where Insurance Governance Breaks Down

The Grant Thornton survey does more than measure confidence levels. It identifies the specific failure modes within the 76% of insurers who cannot demonstrate governance on demand.

Fragmented evidence across teams and tools. Sixty-eight percent of insurance respondents report that AI controls exist but that evidence of those controls is scattered across different departments, platforms, and documentation systems. A model risk management team may maintain validation records. A compliance team may hold bias testing results. An IT team may own data lineage logs. But no single system or process aggregates that evidence into a form that an auditor or regulator could examine as a coherent governance record. The controls exist in isolation; the proof does not.

Board policy without operational implementation. Sixty-one percent of insurance executives say their boards have established AI governance policies. Cross-referencing that with the 24% audit confidence number reveals a 37-point gap between policy adoption and operational readiness. Boards are setting expectations. Operating teams are not meeting them, often because the policies were written at a level of abstraction that does not translate into testable controls for specific AI use cases in underwriting, claims triage, or pricing models.

Governance as the primary failure cause. Forty-four percent of insurance respondents say governance or compliance challenges have directly contributed to AI project failure or underperformance. This is not a risk management concern about what might happen. It is a retrospective acknowledgment that governance gaps have already caused real damage to AI programs.

Regulatory uncertainty as the top scaling barrier. Fifty-six percent of insurers name regulatory or compliance uncertainty as the number-one obstacle preventing them from scaling AI across more business functions. Only 29% cite talent or upskilling gaps as a top barrier, a finding that suggests the technology and people are available but the regulatory framework for deploying them remains too unclear to justify full commitment.

Untested failure response plans. The cross-industry data shows that only 20% of organizations have tested their response plans for AI system failures, despite 73% piloting, scaling, or running autonomous AI. Applied to insurance, this means carriers operating agentic AI systems in underwriting or claims (AIG's 30-hour autonomous agent cycles, for instance) are doing so without having rehearsed what happens when those systems produce errors at scale.

How Insurance Compares: Cross-Survey Benchmarking

The Grant Thornton data does not exist in isolation. Four other major surveys published between late April and early May 2026 provide points of comparison that collectively paint a picture of an industry moving fast on AI deployment while governance infrastructure lags far behind.

AM Best: 63% Have Formal AI Policies, But Data Quality Undermines Them

AM Best's April 2026 special report surveyed over 150 rated insurers and managing general agents. Sixty-three percent report having a formal AI policy in place, a figure close to Grant Thornton's 61% board-level governance policy finding. But only 47% describe their governance processes as "robust." And the data quality problems that undermine governance are pervasive: 45% cite data readiness as their top AI challenge, 43% name security and privacy concerns, and 41% point to legacy system integration.

Sridhar Manyem, AM Best's senior director of industry research, identified the underlying issue: "AI systems can produce unreliable outputs when underlying data is of poor quality, fragmented across legacy systems, insufficiently governed or lacking appropriate context." The message converges with Grant Thornton's fragmentation finding. Governance policies exist, but the data infrastructure underlying AI systems often cannot support the documentation those policies require.

Capgemini: Only 10% Have Scaled AI Successfully

Capgemini's 19th edition World P&C Insurance Report, published May 5 from surveys of 344 senior P&C executives, puts the scaling problem in sharper relief. Only 10% of P&C insurers have successfully scaled AI. Sixty percent remain in exploration or proof-of-concept stages. And 42% track no AI metrics at all, meaning they could not quantify the revenue growth that Grant Thornton's insurance respondents reported even if it existed.

The Capgemini data also reveals a spending allocation problem that feeds governance gaps. Seventy-two percent of carrier AI investment goes to technology and infrastructure. Only 28% goes to change management, training, and organizational readiness, the category that includes governance program development. Kartik Ramakrishnan, Capgemini's CEO for financial services, described it as a "moment of AI truth" for the industry. The carriers Capgemini identifies as "intelligence trailblazers" (the successful 10%) are roughly four times more likely to invest in change management and three times more likely to have deployed explainable AI, both prerequisites for the kind of governance Grant Thornton measured.

Datos Insights: Production Deployments Jump, But Scope Remains Narrow

Datos Insights' 2026 survey of 36 senior carrier technology leaders found that the share of carriers with AI in production grew from 37% to 61% in a single year. On the surface, that looks like progress that might close the proof gap. But Datos also found that most production deployments remain concentrated in document processing rather than the higher-risk underwriting and pricing functions where governance scrutiny is most intense. Underwriting deployment stands at 56%, claims at 50%, and both at limited scope and scale.

The implication: insurers are increasingly running AI in production (Datos), reporting revenue gains from it (Grant Thornton), but concentrating those deployments in lower-risk document handling where governance requirements are less demanding. The proof gap remains widest in the high-risk functions, exactly where regulators focus their attention.

Bain: 78% Adoption, 4% Meaningful Scale

Bain & Company's 2026 survey of 160 global insurers found that 78% have adopted generative AI in some form, but only 4% have scaled it meaningfully across claims operations. The gap between adoption and scale (74 points by Bain's measure) dwarfs even Grant Thornton's 28-point proof gap. But the metrics measure different things. Bain measures operational scale. Grant Thornton measures governance readiness. Together, they describe an industry where three-quarters of carriers have adopted AI, fewer than one in ten have scaled it, and fewer than one in four could prove to a regulator that what they have deployed is governed responsibly.

Comparative Summary

Survey (2026)	Sample	Adoption/Revenue Metric	Governance/Scale Metric	Gap
Grant Thornton	100 insurance execs	52% report AI revenue growth	24% audit-ready in 90 days	28 pts
AM Best	150+ rated insurers	63% formal AI policy	47% robust governance	16 pts
Capgemini	344 P&C execs	60% exploring/POC stage	10% successfully scaled	50 pts
Datos Insights	36 carrier tech leaders	61% AI in production	Document processing focus	Scope gap
Bain	160 global insurers	78% adopted gen AI	4% scaled meaningfully	74 pts

Every survey confirms the same pattern: broad adoption, narrow governance, and a growing gap between what carriers deploy and what they can defend.

The NAIC Intersection: Why the Proof Gap Becomes a Compliance Gap

The proof gap measured by Grant Thornton becomes an acute compliance problem when overlaid on the NAIC's AI regulatory infrastructure. Two NAIC initiatives, already in motion, convert the proof gap from a board-level governance concern into a filing-level compliance requirement.

The 12-State AI Evaluation Tool Pilot

The NAIC's AI Systems Evaluation Tool pilot launched on March 2, 2026, with 12 participating states: California, Colorado, Connecticut, Florida, Iowa, Louisiana, Maryland, Pennsylvania, Rhode Island, Vermont, Virginia, and Wisconsin. The pilot runs through September 2026, with tool updates planned for September and October and formal adoption expected at the NAIC fall meeting in November 2026.

The evaluation tool comprises four exhibits that, taken together, amount to exactly the kind of independent governance examination that 76% of Grant Thornton's insurance respondents said they could not pass. Exhibit A quantifies an insurer's AI usage across functions. Exhibit B applies a governance risk assessment framework. Exhibit C requires detailed documentation on high-risk AI systems, specifically those used in claims decisions, underwriting, pricing, and fraud detection. Exhibit D demands AI data details, including information on reasonable accommodations.

Carriers operating AI in any of those high-risk categories (which is to say, most carriers reporting AI revenue growth) will need to produce the exact kind of aggregated, coherent governance evidence that 68% of Grant Thornton's respondents admit is currently fragmented across teams. The evaluation tool does not create new governance requirements, as the NAIC has been careful to clarify. But it creates a standardized method for regulators to test whether existing governance requirements are being met, and that testing mechanism is the audit equivalent of what Grant Thornton's 90-day confidence question was probing.

The Model Bulletin and Third-Party Oversight

The NAIC's AI Model Bulletin, originally adopted in 2023, has now been implemented by over half of all states. It requires written AI governance programs emphasizing transparency, fairness, and risk management. A model law on third-party AI vendor oversight is anticipated in 2026, potentially including licensing requirements for vendors whose tools carriers depend on.

For carriers whose 68% fragmentation rate reflects heavy reliance on third-party AI tools (and the IA Capital Group survey showing OpenAI in 90% of carrier stacks suggests that reliance is extensive), the third-party oversight framework adds another governance documentation layer. Carriers will need to demonstrate not just that their own systems are governed, but that the vendor systems embedded in their workflows meet the same standards.

The NAIC is also developing "model cards," standardized reporting tools that function like nutrition labels for AI systems, outlining how each system is built, what data it uses, and what risks it presents. These model cards represent a move toward the kind of portable, auditable governance documentation that the proof gap currently reflects the absence of.

The Workforce Readiness Dimension

Grant Thornton's survey includes a workforce readiness component that complicates the proof gap further. Among insurance respondents, only 7% believe their workforce is fully ready to adopt AI. Forty-seven percent describe the workforce as "mostly ready." Thirty-nine percent say frontline employees need the most AI adoption support.

The workforce readiness numbers matter for governance because governance is not a technology problem or a policy problem alone. It is an execution problem that requires people who understand both the AI systems and the regulatory requirements to produce the documentation that proves compliance. A workforce that is "mostly ready" to use AI tools is not necessarily equipped to document, validate, and maintain governance records for those tools.

Cross-industry, the readiness gap is even more pronounced by role. Thirty-nine percent of CIOs and CTOs believe their workforce is fully ready for AI, while only 7% of COOs share that assessment. In insurance, where COOs typically own claims operations and underwriting workflows (the highest-risk AI deployment areas), that 32-point confidence gap between technology leaders and operations leaders suggests that the people closest to governed AI deployments are the least confident in organizational readiness.

Capgemini's data reinforces this pattern. Forty-seven percent of insurance employees report that their workdays have not changed after 18 months of AI tool access. Forty-three percent cite job security concerns. Only 14% are "very clear" on AI's role in their work. A governance program built on policies that the workforce cannot operationalize is a governance program in name only, precisely the kind of program that fails Grant Thornton's 90-day audit test.

The Actuary's Expanding Role in AI Governance

Patterns we have tracked across carrier AI governance frameworks over the past 18 months point to an expanding set of responsibilities for actuaries that goes well beyond traditional model validation. The proof gap creates professional obligations and professional opportunities simultaneously.

From Reserve Specialist to AI Model Validator

ASOP No. 56 (Modeling) already requires actuaries to understand, document, and validate the models they rely on for professional work. As AI models enter pricing, reserving, and underwriting workflows, the scope of that obligation expands. An actuary signing a rate filing that incorporates AI-generated risk scores cannot simply certify the output. ASOP No. 56 requires documentation of the model's design, data inputs, assumptions, and limitations. When Grant Thornton finds that 68% of insurers have fragmented governance evidence, the actuary relying on those fragmented records for a statutory filing faces a professional standards question, not just a corporate governance one.

The SOA's 2026 ASA Job Analysis Survey confirmed this shift, finding that AI skills are now expected in entry-level actuarial positions. The direction is clear: AI model validation is becoming an actuarial competency, not a data science add-on that happens adjacent to actuarial work.

The Governance Gatekeeper Function

Mathew Tierney, Grant Thornton's global insurance practice leader, described the carriers that are ahead on governance: "Insurance companies that are ahead have built AI governance into their operating model across underwriting, claims, pricing and customer experience workflows with specificity that regulators seek." That "specificity" language is significant. Generic governance policies (the kind that 61% of boards have adopted) do not satisfy regulators. What satisfies them is function-specific documentation showing how AI governance applies to the exact underwriting model, claims triage algorithm, or pricing engine under examination.

Actuaries are uniquely positioned to provide that function-specific translation. They understand the regulatory requirements for rate filings and reserve opinions. They understand the statistical properties of the models. And they occupy a professional role with credentialing, standards of practice, and disciplinary oversight that gives their governance attestations weight with regulators. The proof gap is, in significant part, a gap that the actuarial profession is being asked to fill.

Colorado as the Leading Indicator

Colorado's AI Act, with its June 30, 2026 compliance deadline, provides the most concrete near-term example of how the proof gap becomes an actuarial compliance obligation. The act requires deployers of high-risk AI systems (which includes insurance pricing and underwriting models) to conduct algorithmic impact assessments, implement risk management policies, and provide notice to consumers about AI use in consequential decisions. Carriers operating in Colorado who cannot demonstrate governance on demand, the 76% in Grant Thornton's survey, face statutory enforcement risk within weeks.

For pricing actuaries filing rates in Colorado, the AI governance documentation now has a direct bearing on whether a rate filing survives regulatory review. That connection between AI governance and the traditional actuarial function of rate certification is the mechanism through which the proof gap enters the actuarial workflow.

Autonomous AI Amplifies the Proof Gap

Grant Thornton's cross-industry findings on autonomous and agentic AI carry particular relevance for insurance. Seventy-three percent of respondents across all industries report piloting, scaling, or running autonomous AI systems. Ninety-five percent do not permit fully autonomous high-stakes decisions without human review. Sixty percent limit AI agents to moderate-risk task automation. And 43% cite regulatory and compliance uncertainty as their top concern around autonomous AI.

In insurance, autonomous AI is no longer hypothetical. AIG's disclosed 30-hour agent cycles, Chubb's claims automation targeting 85% straight-through processing, and Hartford's agentic underwriting assistants all represent production-scale autonomous AI in regulated insurance functions. Each of those deployments requires governance documentation proportional to its risk level, and the proof gap suggests that documentation is not keeping pace with deployment speed.

Grant Thornton's finding that only 20% have tested response plans for AI failures takes on particular significance in this context. An autonomous underwriting agent that makes an error does not generate a single bad decision. It generates a pattern of decisions across an entire portfolio segment before human oversight detects the problem. The governance infrastructure needed to monitor, document, and respond to that kind of failure mode is qualitatively different from the governance infrastructure for a traditional assisted-decision tool, and it is the kind of infrastructure that the proof gap suggests most carriers have not built.

Closing the Proof Gap: What the Survey Data Suggests

Grant Thornton's cross-industry data contains one finding that offers a roadmap for closing the proof gap. Organizations with fully integrated AI (as opposed to those still piloting) show governance confidence rising tenfold: from 7% at the piloting stage to 74% at full integration. The data suggests that governance maturity is a byproduct of deployment maturity, not a prerequisite for it, but only when organizations invest in governance infrastructure concurrently with deployment.

The carriers in Capgemini's "trailblazer" category, the 10% that have successfully scaled AI, show a similar pattern. They invest four times more in change management, are three times more likely to deploy explainable AI, and are twice as likely to embed AI expectations into job descriptions. Their 21% higher revenue growth and 51% greater share price increase over three years represent the financial return on governance investment.

Grant Thornton's own recommendations for insurers center on three actions. First, evaluate and modernize governance structures with AI use-case classification and monitoring standards. Second, assess operating models for AI-native execution and redesign workflows across underwriting, claims, and service functions. Third, build GRC (governance, risk, and compliance) frameworks that embed regulatory standards, including NIST AI RMF, ISO/IEC 42001, and the EU AI Act, directly into operational workflows rather than maintaining them as separate compliance documents.

For actuarial teams specifically, the proof gap creates immediate professional obligations. Rate filings that rely on AI-generated inputs require ASOP No. 56 documentation that many carriers currently cannot produce. Reserve opinions that incorporate AI-driven claim severity projections need validation records that the 68% fragmentation rate suggests do not exist in aggregated form. And the NAIC's evaluation tool pilot, running through September 2026, will test whether the governance that carriers claim to have in place actually functions under regulatory examination.

Why This Matters for Actuaries

The proof gap is not a technology problem that technology teams will solve. It is a governance, documentation, and validation problem that sits squarely in the actuarial domain. Three implications stand out.

Professional standards exposure. An actuary who certifies a rate filing or signs a reserve opinion based on AI model outputs that lack proper governance documentation faces potential ASOP compliance questions regardless of whether the AI model itself performed well. The proof gap means the documentation trail that supports actuarial professional judgment is incomplete for a majority of carriers using AI in regulated functions.

Demand for AI governance skills. The combination of the NAIC evaluation tool pilot (12 states, September 2026 completion), Colorado's AI Act (June 30, 2026 enforcement), and the EU AI Act (August 2, 2026 enforcement for high-risk systems) creates a concentrated demand spike for professionals who can bridge AI systems and regulatory compliance. Actuaries with model validation experience and regulatory filing expertise are the natural candidates to fill that role, and the SOA's addition of AI competencies to the ASA pathway reflects that expectation.

Competitive differentiation for carriers. Grant Thornton's data shows that the carriers with the strongest governance also report the strongest AI outcomes. The 10x confidence improvement from piloting to full integration, the 4x revenue advantage of Capgemini's trailblazers, and the correlation between governance maturity and sustainable AI ROI all suggest that governance is not a cost center or a brake on innovation. It is a competitive advantage that the majority of carriers have not yet captured because they invested in AI deployment without investing proportionally in the proof infrastructure that makes deployment defensible.

The proof gap will close, either through deliberate governance investment or through regulatory enforcement actions that force the issue. For actuaries, the professional question is whether to lead that process or to react to it after the fact.