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

Alpha FMC's April 2026 Insurance Outlook declares the experimentation phase over. Global head of insurance consulting Britton Van Dalen frames the shift to "operationalisation, ensuring those investments deliver real, measurable improvements in underwriting performance, capital efficiency and customer experience."

The number that makes 2026 the accountability year is not the payoff. It is the cost. Morgan Stanley models over $6.0 billion of gross AI cost savings across its carrier universe this year, with only 10% reaching operating earnings against $3.0 billion of implementation cost, a net $2.4 billion drag.

Key Takeaways

• $2.4 billion net drag on operating income in 2026 is the trough of Morgan Stanley's J-curve: $6.0 billion of gross savings, $600 million reaching earnings, $3.0 billion of implementation spend, against $9.3 billion of projected improvement by 2030.
• AIG discloses the most attributable figures: a 26% year-over-year rise in Lexington submission count, a 35% improvement in Middle Market Property submit-to-bind, and an expense ratio of 31.1%, down 90 basis points, targeting sub-30% by 2027.
• Chubb is the only carrier to name a combined ratio number, 1.5 points from automation alongside a roughly 20% headcount reduction, and has not isolated any of it in a financial disclosure.
• P&C expense ratios fell 2.4 points over 11 years to 25.3% in 2024, of which only 0.5 points came from general expenses, which is the line where AI has to show up to be distinguishable.
• 95% of generative AI pilots fail to reach production, with infrastructure limits behind 64% of scaling failures and production cost overruns averaging 380% against pilot projections.

The Trough Is the Story

The projections and the current cost sit at opposite ends of the same curve, and only one of them is in this year's numbers.

Morgan Stanley projects $9.3 billion of AI-driven operating income improvement by 2030, with expense ratios falling roughly 200 basis points across the carriers analyzed. The near-term structure is the opposite shape. For 2026 the same analysis estimates over $6.0 billion in gross cost savings, of which only 10%, about $600 million, flows through to operating earnings, against $3.0 billion of implementation costs. Net, that is a $2.4 billion drag.

The baseline makes attribution harder still. AM Best has the P&C industry expense ratio falling from 27.7% in 2014 to 25.3% in 2024, a 2.4-point decline over 11 years, of which other acquisition expenses supplied 1.9 points and general expenses 0.5. Remote work and operational consolidation were already moving that line, so AI has to appear in general expense or loss adjustment expense to be separable.

What the Four Disclosing Carriers Actually Show

Four carriers have put enough quantified AI detail into filings and earnings calls to compare. They are not measuring the same thing.

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)

AIG discloses the most. Underwriting Assist, built on Claude and Palantir Foundry for E&S submissions, coincided with a 26% year-over-year rise in Lexington submission count and a 35% improvement in Middle Market Property submit-to-bind, with over 370,000 submissions processed by year-end 2025 against a 500,000 target for 2030. The expense ratio improved to 31.1%, down 90 basis points, with a sub-30% target reaffirmed for 2027. The submit-to-bind figure is the actuarially interesting one: it points at converting better-quality submissions rather than more of them, which should reach loss ratio over a 12 to 24 month lag.

Chubb has committed to the most specific financial target and shown the least evidence for it. The December 2025 investor presentation set out roughly 20% headcount reduction, 85% process automation and 1.5 combined ratio points of expense savings. Q1 2026 delivered an 84.0% P&C combined ratio with nine active AI projects, and Greenberg attributed the result to underwriting discipline and pricing adequacy rather than AI. Until the contribution is isolated, 1.5 points is a projection.

Travelers made the largest deployment and the fewest measurable claims: nearly 10,000 staff with personalized Claude assistants, 30,000-plus with TravAI access, and "meaningful improvements in productivity" as the disclosed outcome. Its Q1 2026 expense ratio of 28.6% sits in its historical range with no inflection, on a deployment less than four months old at quarter close.

Progressive does not frame the spend as AI at all, describing data-driven underwriting and usage-based insurance across two decades of Snapshot telematics. The 86.4% combined ratio and 9% growth in policies in force to 39.6 million are the result; no single initiative separates out of it.

Morgan Stanley's automation rates put the disclosure gap in context: standard carriers including Travelers, Allstate and Progressive cluster at 20% to 21%, specialty writers such as Arch, Hamilton and Everest reach 25% to 27%, with a 21.6% carrier average against 25.1% for brokers. Most of the industry is early enough that thin ROI evidence is what you would expect.

The Evidence Cannot Arrive on This Schedule

The constraint on the accountability year is that the metrics being demanded mature more slowly than the demand for them.

Loss ratio improvement and reserve adequacy, the two actuarially meaningful measures, need 12 to 36 months of earned premium development to validate, and credible work needs two to three accident years of mature development. A tool deployed in January 2025 has one full year of earned premium. That places defensible ROI studies for the 2025 deployment wave in 2027 and 2028, two years after the boards asking for them.

Attribution is the second constraint and it does not resolve with time. AI tools are deployed alongside revised underwriting guidelines, new pricing models, portfolio repositioning and headcount changes. A line of business that deployed an AI pricing tool, revised territory factors and exited three unprofitable classes in the same year offers no way to assign the resulting loss ratio change. Where a carrier does isolate a real improvement, it has a competitive reason not to publish it, which is why vendor case studies carry testimonials rather than figures.

Credibility is the third. Deployments often target a segment too small to support a credible loss ratio comparison inside one accident year: a tool running on $50 million of commercial property needs several years before an improvement separates from normal volatility.

Meanwhile the failure rate keeps consuming the budget that would fund the measurement. Research across sources puts 95% of generative AI pilots short of production, with infrastructure limits behind 64% of scaling failures, production costs overrunning pilot projections by an average of 380%, and a median 14 months from pilot approval to shutdown. Every stalled pilot is expense drag with no loss ratio improvement behind it, which is the J-curve trough experienced one project at a time. Morgan Stanley's additional 200 basis points by 2030 works out to roughly 80 basis points a year of AI-specific expense improvement, against 0.5 points of general expense movement across the whole prior decade.