Insurance AI Hits the ROI Wall After Years of Pilots: Which Carriers Are Converting Spend Into Actuarial Results

From benchmarking AI-related expense disclosures across the top 20 P&C carriers' earnings calls over six quarters, fewer than a third can point to a specific, quantified operational improvement. The rest describe AI in aspirational language: "exploring," "piloting," "investing for the future." Alpha FMC's April 2026 Insurance Outlook makes the shift official. The consulting firm, which has more than 1,540 consultants working across global insurance and asset management, finds that the industry has moved beyond technology modernization and experimentation. The focus now, per Alpha FMC global head of insurance consulting Britton Van Dalen, "shifts to operationalisation, ensuring those investments deliver real, measurable improvements in underwriting performance, capital efficiency and customer experience."

That framing puts a fine point on something actuaries have felt in practice: the gap between AI announcements and actuarially measurable outcomes. This article constructs a cross-carrier ROI scorecard using public filings, earnings call transcripts, and expense ratio data to identify which carriers are converting AI investment into results that show up in combined ratios, submission throughput, and expense lines, and which remain stuck in pilot mode.

The ROI Wall: Why 2026 Is the Accountability Year

Several forces are converging to make 2026 the year that boards, investors, and regulators start demanding receipts on AI spending.

Morgan Stanley's implementation cost J-curve. In their P&C insurance AI analysis, Morgan Stanley projects $9.3 billion in AI-driven operating income improvement by 2030, with expense ratios declining approximately 200 basis points across the carriers analyzed. The critical detail is the near-term cost structure. For 2026, Morgan Stanley estimates over $6.0 billion in gross cost savings across their carrier universe, but with only 10% flowing through to operating earnings ($600 million) and $3.0 billion in implementation costs, the net effect is a $2.4 billion drag on operating income. Carriers are in the trough of the J-curve right now, spending heavily while the measurable returns remain thin.

AM Best's 11-year expense ratio baseline. AM Best data shows the P&C industry expense ratio fell from 27.7% in 2014 to 25.3% in 2024, a 2.4-percentage-point decline over 11 years. That decline was driven primarily by lower other acquisition expenses (down 1.9 points), with general expenses contributing just 0.5 points, while commissions and taxes remained flat. The question AI ROI analysis must answer: is the expense ratio improvement accelerating beyond what remote work and operational consolidation already delivered, or is AI spending being absorbed into a trend that was already underway?

Alpha FMC's four-theme framework. Alpha FMC's 2026 Outlook identifies four strategic themes driving insurer technology investment: embedded AI and underwriting discipline, core modernization and technology resilience, data governance and AI-ready foundations, and product innovation and distribution. The first theme is the one that connects most directly to actuarial measurement. Insurers applying AI across underwriting, claims, and actuarial processes must now link underwriting decisions to portfolio outcomes, not just process speed.

SAS's "operating system" forecast. SAS industry experts project that a Fortune 500 insurer will begin phasing out policy administration systems in favor of AI copilots during 2026, and that straightforward claims will settle in minutes via agentic AI. These forecasts describe an industry where AI is no longer supplemental. If the predictions hold, the financial impact should be visible in expense ratios and loss adjustment expense lines within two to three quarters.

The Carrier Scorecard: Four Approaches to AI ROI

Four carriers have disclosed enough specific, quantified AI metrics in public filings and earnings calls to construct a meaningful comparison. Each represents a distinct AI deployment strategy. Dozens of others mention AI in their earnings commentary without disclosing a single number.

AIG: The Submission Throughput Play

AIG's AI strategy centers on Underwriting Assist, a GenAI platform built on Anthropic's Claude and Palantir Foundry that processes excess and surplus lines submissions. The Q4 2025 earnings call, delivered in February 2026, provided the most granular AI performance data of any carrier.

Quantified metrics from the Q4 2025 transcript:

• 26% year-over-year increase in Lexington submission count after deploying Underwriting Assist
• 35% improvement in Lexington Middle Market Property submit-to-bind ratio
• 370,000+ submissions processed by year-end 2025, against a target of 500,000 by 2030
• Expense ratio improved to 31.1%, down 90 basis points year-over-year, with management reaffirming a sub-30% target by 2027
• Everest portfolio conversion accelerated through GenAI evaluation: $65 million in Q4 gross premiums written, $180 million in January at 75% retention

AIG has also expanded Underwriting Assist to seven additional lines of business, with full rollout across North America, UK, and EMEA planned for 2026. The company described development of an orchestration layer coordinating multiple AI agents across front-office, mid-office, and back-office functions.

Actuarial read: AIG's AI ROI story is fundamentally about top-line growth through submission capacity rather than pure expense reduction. A 26% lift in submission count, if underwriting discipline holds, translates directly into premium volume. The 90-basis-point expense ratio improvement is meaningful, though attributing it solely to AI is premature given concurrent restructuring efforts. The submit-to-bind ratio improvement is the most actuarially significant metric: it suggests the AI tools are helping underwriters identify and convert better-quality submissions faster, which should flow into loss ratio performance over a 12- to 24-month lag. AIG holds three patents protecting the specific extraction and traceability methods behind this system, covering markdown-based document processing, chunk-level hallucination detection, and chain-of-thought prompting for complex spreadsheets.

Chubb: The Expense Ratio Compression Target

Chubb's AI story is the most explicitly tied to financial targets of any major carrier. CEO Evan Greenberg's December 2025 investor presentation disclosed a plan to reduce headcount by approximately 20% through a multi-year AI transformation, targeting 85% process automation and projecting 1.5 combined ratio points in expense savings. In Q1 2026, Chubb delivered a P&C combined ratio of 84.0% and highlighted "agentics and large language model capabilities as major areas of strategic focus" across retail and E&S business lines.

What the Q1 2026 earnings call disclosed:

• Nine active AI transformation projects across the organization
• Small commercial and E&S segments identified as the primary AI growth catalyst over a five-year horizon
• Cyber insurance underwriting enhanced with AI vulnerability assessment tools (referenced Anthropic's Mythos and Gemini models)
• Management described progress as "on track" but provided no specific quantified savings or expense ratio improvement attributed to AI

Actuarial read: Chubb is the only major carrier to publicly commit to a specific combined ratio improvement from AI (1.5 points). That target is aggressive but not unreasonable: Morgan Stanley's carrier-level analysis flags Chubb as one of the carriers positioned to gain the most operating margin from AI adoption. The challenge is verification. Q1 2026's 84.0% combined ratio is strong, but Greenberg attributed performance to underwriting discipline and pricing adequacy rather than AI-driven cost savings. Until Chubb isolates the AI contribution in its financial disclosures, the 1.5-point target remains a projection, not a result. The 20% headcount reduction plan, meanwhile, represents the clearest signal that Chubb views AI as an expense ratio tool, not primarily a revenue growth tool, positioning it differently from AIG.

Travelers: The Enterprise Deployment Bet

Travelers' January 2026 announcement of a partnership with Anthropic represents the largest disclosed carrier-to-foundation-model deployment in the industry. Nearly 10,000 engineers, data scientists, analysts, and product owners received personalized Claude AI assistants, while the broader organization of 30,000+ employees gained access to TravAI, Travelers' in-house agentic AI platform integrating multiple generative AI tools with internal systems.

What is publicly known:

• 10,000 staff equipped with personalized Claude and Claude Code assistants
• 30,000+ employees with access to TravAI, a secure in-house agentic AI platform
• Executive VP Mojgan Lefebvre described "significantly elevated levels of engineering excellence and meaningful improvements in productivity"
• Travelers invested substantially in "transformative technology" during 2024-2025
• 2024 full-year underlying combined ratio: 86.2%, with record $4.5 billion in after-tax underlying underwriting income (a 40% increase from 2023)
• 2025 net income of $6.288 billion, up $1.289 billion from 2024

Actuarial read: Travelers has bet on breadth over depth: rather than targeting a single workflow (like AIG's submission processing), it has equipped a large share of its technical workforce with general-purpose AI tools. The financial results are strong, but Travelers has not disclosed a single metric tying specific productivity improvements to specific financial outcomes. "Meaningful improvements in productivity" is not "X% reduction in policy processing time" or "Y basis points of expense ratio improvement." Travelers' Q1 2026 expense ratio of 28.6% is in line with its historical range, showing no obvious AI-driven inflection point yet. The Anthropic deployment is less than four months old at Q1 close, so the financial signal may simply be too early to detect. For actuaries evaluating build-versus-buy decisions, Travelers represents the "buy the platform, deploy broadly, measure later" approach.

Progressive: The Two-Decade Data Compound

Progressive presents a different case entirely. Where the other three carriers are deploying AI as a distinct strategic initiative, Progressive has been compounding data science and telematics investment for over 20 years. Its Snapshot telematics program, powered by machine learning and behavioral analytics, represents the longest-running production AI deployment in personal lines insurance.

Q1 2026 performance:

• 86.4% consolidated combined ratio (0.4 points worse than Q1 2025, driven by 12.5 points of personal property cat load from March severe convective storms)
• $2.8 billion net income, up 9.8% year-over-year
• 9% growth in policies in force to 39.6 million
• $22.19 billion in quarterly revenue

Actuarial read: Progressive rarely frames its technology investment as "AI" in the way Chubb, AIG, or Travelers do. Instead, it describes data-driven underwriting, pricing segmentation, and usage-based insurance. The result is a carrier that consistently operates at combined ratios that peers struggle to match in personal lines, with policy-in-force growth that suggests pricing accuracy rather than market-share buying. Progressive's AI ROI is embedded and compounded, making it nearly impossible to isolate the incremental contribution of any single AI initiative. For actuaries, Progressive demonstrates what the long tail of data science investment looks like: not a sudden inflection, but a durable competitive moat in risk selection and pricing precision that compounds over market cycles.

The Expense Ratio Evidence: Is AI Showing Up in the Numbers?

If AI is delivering measurable performance, the P&C expense ratio is where it should appear first. Claims improvements take longer to materialize in loss development patterns, but operational efficiency gains from automation, document processing, and workflow optimization should reduce general expenses within one to two years of deployment.

Carrier	AI Strategy	Quantified AI Metric	Expense Ratio Trend	ROI Evidence Grade
AIG	Submission throughput via GenAI	26% submission lift; 90bp expense ratio improvement	31.1% (improving toward sub-30%)	B+ (quantified, partially attributable)
Chubb	Expense compression via automation	1.5 CR points projected; no actuals disclosed	Stable at strong levels	C+ (target only, no attribution)
Travelers	Enterprise platform deployment	"Meaningful productivity" (no numbers)	~28.6% (stable)	C (qualitative only)
Progressive	Compounded data science + telematics	Embedded in pricing; not isolated	Industry-leading combined ratio	A- (results clear, attribution embedded)

Morgan Stanley's carrier-level analysis assigns varying automation rates across the industry: standard carriers like Travelers, Allstate, and Progressive cluster around 20-21% agentic AI automation, while specialty providers like Arch Capital, Hamilton, and Everest reach 25-27%. The broker average sits at 25.1%, and the overall carrier average at 21.6%. These rates imply that most carriers are still in early-stage automation, consistent with the thin ROI evidence in public disclosures.

The Pilot Purgatory Problem

Beyond the top four, the picture is stark. Industry research consistently finds that most insurers remain stuck in what consultants and analysts now call "pilot purgatory," where AI projects demonstrate technical feasibility but never reach production scale or generate measurable financial returns.

The data points converge across multiple sources:

• Fewer than half of insurance businesses have deployed AI in even a single function at production scale
• More than four in five insurance companies dedicate at least $5 million annually to AI, with 14% spending more than $50 million, yet finance teams cannot tie AI investments to returns
• Research from multiple sources indicates that 95% of generative AI pilots fail to scale to production deployment
• Infrastructure limitations account for 64% of scaling failures, with cost overruns averaging 380% at production scale versus pilot projections
• Median time from pilot approval to production shutdown: 14 months

McKinsey's insurance AI research identifies a structural cause: insurers underestimate the full range of investment needs, leading to small-scale, fragmented efforts and poor ROI. The pattern is consistent. A carrier runs a proof-of-concept with a vendor. The POC succeeds on a limited dataset. The team requests production funding. The production buildout requires data infrastructure, model governance, integration with core systems, regulatory review, and ongoing monitoring capabilities that were not scoped in the pilot budget. Costs escalate. The project stalls. A new pilot starts in a different business unit.

This cycle has a direct actuarial impact. Every stalled AI pilot represents expense drag without corresponding loss ratio improvement. For carriers in the trough of Morgan Stanley's J-curve, the implementation costs are real and current while the operating income benefits remain projected and future.

Build vs. Buy Economics in Actuarial AI Tooling

The carrier scorecard reveals a deeper strategic question: should insurers build proprietary AI capabilities or buy them from established vendors? The answer has significant implications for where AI ROI materializes on the income statement.

The builders: AIG and Progressive. Both carriers have invested in proprietary technology stacks. AIG built its three-layer architecture combining Anthropic's Claude, Palantir Foundry, and proprietary extraction tools, then protected the methods with patents. Progressive built its telematics and pricing analytics capabilities in-house over two decades. In both cases, the AI capability becomes a durable competitive advantage, but the upfront investment is substantial and the payback period is long.

The buyers: carriers using Verisk, Guidewire, and EXL platforms. Verisk's Synergy Studio, Guidewire's PricingCenter, and EXL's AI patent portfolio (10 patents covering document extraction, knowledge graphs, domain-specific LLMs, and regulatory reporting) represent the vendor side of the equation. These platforms offer faster time-to-deployment and lower upfront costs, but the AI-driven improvements flow into the vendor's product rather than the carrier's proprietary moat. EXL's services model is particularly notable: its patents protect methods deployed within long-term managed engagements, creating structural switching costs that extend beyond typical vendor relationships.

The platform approach: Travelers and Chubb. Both carriers are deploying foundation model capabilities across their organizations, essentially buying the AI engine (Anthropic for Travelers, multiple vendors for Chubb) while building the integration and workflow layers internally. This hybrid approach splits the difference on cost and control but creates a measurement challenge: when the AI tool is general-purpose, isolating its contribution to any specific financial outcome becomes harder.

For mid-market carriers evaluating these options, the build-versus-buy math depends heavily on scale. Morgan Stanley's data suggests specialty providers achieve higher automation rates (25-27%) than standard carriers (20-21%), potentially because their more focused operations allow deeper integration of AI tools into specific workflows. A carrier writing $2 billion in specialty E&S business may extract more measurable ROI from a purpose-built AI tool than a multi-line carrier deploying the same tool across a dozen business units.

What "Measurable Performance" Means Through an Actuarial Lens

Alpha FMC's framework calls for "measurable improvements in underwriting performance, capital efficiency and customer experience." For actuaries, translating that into concrete metrics requires mapping AI capabilities to the specific financial lines they should affect.

Loss ratio improvement. AI-assisted underwriting should improve risk selection, which shows up as better loss ratios on written business over a 12- to 36-month development period depending on tail length. AIG's submit-to-bind ratio improvement is an early indicator: better-quality submissions being identified and bound faster should eventually appear as favorable loss development on those accident years. Industry research suggests AI can improve loss ratios by approximately 3 percentage points through better use of unstructured and previously inaccessible data, though carrier-specific evidence remains thin.

Cycle time compression. Quote-to-bind time is collapsing in carriers with mature AI deployment. Industry benchmarks suggest reductions from days to minutes in standardized risks, with straight-through processing rates jumping from 10-15% to 70-90% in the most advanced implementations. For actuaries, cycle time improvement matters primarily through its effect on premium volume (more quotes processed means more policies bound if hit ratios hold) and expense efficiency (less human touch per policy means lower per-policy general expense).

Submission throughput. AIG's 26% submission increase and trajectory toward 500,000 annual submissions by 2030 represent the clearest example. In E&S lines where submission volume directly constrains growth, AI-driven throughput is a revenue accelerator. The actuarial question is whether increased volume comes with maintained underwriting discipline or whether speed introduces adverse selection through reduced human review time per risk.

Expense ratio decomposition. AM Best's 11-year data showing a 2.4-point decline in P&C expense ratios establishes the pre-AI baseline. Of that decline, 1.9 points came from other acquisition expenses (largely attributable to remote work and operational consolidation) and only 0.5 points from general expenses. AI's incremental contribution needs to show up in the general expense line or the loss adjustment expense line to be distinguishable from trends already in motion. Morgan Stanley projects an additional 200 basis points of expense ratio improvement from AI by 2030, which would roughly double the rate of improvement seen over the prior decade.

Reserve adequacy signal. AI tools that improve claims triage, medical bill review, or litigation management should eventually produce more accurate initial reserves and less adverse development. This is the hardest AI benefit to measure because reserve adequacy judgments require multiple development periods to validate. No carrier has yet disclosed AI-specific reserve development data, though Travelers' explicit provision for "uncertainty" in AY 2025 IBNR alongside AI deployment creates an interesting natural experiment.

Why Most Insurtech AI Vendors Lack Published Actuarial ROI Evidence

The vendor ecosystem supporting carrier AI deployments presents its own measurement gap. Established platforms like Verisk, Guidewire, and Earnix have market position and product capabilities, but published, carrier-verified actuarial ROI evidence is scarce. Newer entrants, from Shift Technology in fraud detection to Akur8 in pricing optimization to hyperexponential in exposure management, face the same challenge.

Several structural factors explain the gap:

Confidentiality constraints. Carriers that achieve measurable AI-driven improvements are reluctant to disclose specifics because the improvement represents competitive advantage. A carrier that discovers its AI-assisted underwriting model produces a 3-point loss ratio improvement on a $500 million portfolio is not going to publish that finding in a vendor case study. The vendor gets a testimonial quote about "significant improvement" without the numbers that would allow actuarial verification.

Attribution complexity. Most AI tools are deployed alongside other operational changes: new underwriting guidelines, revised pricing models, portfolio repositioning, headcount changes. Isolating the AI tool's marginal contribution requires controlled experiments that production insurance operations rarely accommodate. An actuary reviewing a line of business that deployed an AI pricing tool, revised its territory factors, and exited three unprofitable classes simultaneously cannot attribute the resulting loss ratio change to any single factor.

Development period lag. The actuarially meaningful ROI metrics, particularly loss ratio improvement and reserve adequacy, require 12 to 36 months of earned premium development to validate. A vendor tool deployed in January 2025 is only now producing its first full year of earned premium data. Actuarial-grade evidence requires at least two to three accident years of mature development, placing credible ROI studies in the 2027-2028 timeframe for tools deployed during the 2025 wave.

Small-sample credibility. Many vendor deployments target specific segments or lines of business where the written premium base is too small to produce statistically credible loss ratio comparisons within a single accident year. An AI tool deployed on $50 million of commercial property business needs multiple years of loss-free or loss-heavy experience before a pricing actuary can distinguish AI-driven improvement from normal loss volatility.

The Scorecard Through Q1 2026

Pulling together the carrier disclosures, industry data, and vendor landscape, the insurance AI ROI picture as of Q1 2026 looks like this:

Carriers with quantified AI ROI evidence (partial): AIG (submission throughput, expense ratio trajectory), Progressive (embedded in multi-decade data science compound, visible in combined ratio and growth metrics).

Carriers with AI ROI targets but no attribution: Chubb (1.5 CR points projected, nine projects active, no isolated financial impact disclosed).

Carriers with large AI deployments but no quantified ROI: Travelers (10,000-person Anthropic deployment, "meaningful productivity improvements," no financial metrics tied to AI).

Industry-wide: P&C expense ratios declined 2.4 points over 11 years through 2024, primarily from non-AI factors. Morgan Stanley projects an additional 200 basis points from AI by 2030. The gap between those two numbers, roughly 80 basis points per year of AI-specific improvement needed by 2030, has not yet appeared in aggregate industry data.

What This Means for Actuaries

The transition from AI experimentation to measurable performance creates several practical implications for actuaries across pricing, reserving, and enterprise risk management.

Expense assumptions in rate filings. Pricing actuaries incorporating prospective expense ratio improvements from AI into rate indications face a data problem. The industry-wide evidence supports a gradual downward trend in expense ratios, but the AI-specific contribution cannot yet be isolated from remote work effects and operational consolidation. ASOP No. 29 (Expense Provisions in Property/Casualty Insurance Ratemaking) requires that expense assumptions reflect the actuary's best estimate of future expenses. Actuaries filing rates with AI-adjusted expense loads should document the basis for any assumed AI-driven improvement, which as of Q1 2026 means relying on carrier-specific projections (like Chubb's 1.5 points) rather than observed industry-wide evidence.

Reserve implications of AI-assisted claims. If AI tools are accelerating claims triage and improving initial reserve accuracy, the development patterns in the reserving triangle should eventually shift. Actuaries working with carriers that have deployed AI claims tools should be monitoring whether case reserve adequacy is improving (smaller IBNR relative to case), whether claim closure rates are accelerating, and whether these changes are consistent across lines or concentrated in the segments where AI was deployed.

Vendor risk in outsourced AI. For carriers buying AI capabilities from vendors like EXL, Verisk, or Guidewire, the lack of published actuarial ROI evidence means that vendor selection decisions are being made largely on capability demonstrations rather than proven financial impact. Actuaries involved in model governance under ASOP No. 56 should be asking vendors for the same level of quantified, carrier-verified performance data they would require for any actuarial model entering production use.

Career positioning. The carriers converting AI spend into measurable results share a common trait: actuaries who can work at the intersection of data science, underwriting, and financial reporting. AIG's Underwriting Assist tool, Travelers' TravAI platform, and Chubb's automation program all require actuaries who can measure AI's contribution, not just deploy it. The measurement function, designing attribution frameworks, monitoring development patterns, and validating that AI-driven risk selection improvements are real and durable, is where actuarial expertise is most valuable and hardest to automate.

Looking Ahead: The 2026 Earnings Season as a Measurement Window

The remainder of the 2026 earnings season will provide the next set of data points. Allstate, Hartford, Markel, and Arch Capital have varying levels of disclosed AI investment, and their Q1 and Q2 results will either widen or narrow the ROI evidence gap. Hartford's Q1 2026 results, released April 23, included the first explicit tech-led expense ratio disclosure of the earnings season, quantifying the financial contribution of its agent-facing generative AI rollout.

For actuaries tracking this space, the metrics to watch are not the ones carriers volunteer in their prepared remarks. They are the ones analysts extract during the Q&A: specific expense ratio basis points attributable to AI, specific headcount changes in operations and claims, specific lines of business where AI-assisted underwriting is producing different loss development patterns than traditional workflows. Until those answers contain numbers rather than adjectives, the insurance industry's AI ROI story remains more promise than proof.

Alpha FMC's 2026 Outlook establishes the benchmark: measurable commercial advantage. By that standard, as of Q1 2026, the industry is closer to the starting line than the finish.