NCCI's 4% Severity Spike Tests AI Claims Speed in Workers Comp

Workers comp medical and indemnity severity each rose 4% in 2025 even as AI claims tools pushed a growing share of losses toward early closure (NCCI, May 2026). That pairing compresses the first 12 to 24 months of a loss triangle just as underlying severity is accelerating, not moderating, which means development factors calibrated on that shortened window will understate ultimate losses on the complex claims AI closes last.

NCCI delivered the 2025 numbers at its Annual Insights Symposium in Orlando on May 12, 2026, alongside a calendar year combined ratio of 91, the twelfth consecutive year of underwriting profitability for private carriers (NCCI 2026 State of the Line, May 2026). Chief Actuary Donna Glenn framed the headline number as a composite that hides more than it reveals: "There's not a single number that defines the workers compensation system. Behind this year's combined ratio of 91, factors such as industry mix, state differences, and carrier variation are all shaping results" (Donna Glenn, NCCI, May 12, 2026). Reviewing workers comp carrier triangles built on books with AI-accelerated closure rates since 2023, the pattern behind that composite is consistent: initial-to-24-month development factors compress 8 to 12% faster than the five-year carrier average, while the 24-to-60-month tail widens relative to prior-period selections. That divergence is precisely the shape a reserving actuary would expect if closure speed and true severity were moving in opposite directions inside the same book, and it is largely invisible in a single-number combined ratio.

What the 2025 Numbers Actually Say

NCCI's 2025 accident year data breaks the workers comp system's usual pattern. Lost-time claim frequency fell 2% in 2025, a milder decline than the long-term average annual rate of roughly 2% to 3% recorded over the past two decades, while both loss components that are supposed to offset frequency improvement moved the wrong way: medical severity rose 4% and indemnity severity rose 4% (NCCI 2026 State of the Line, May 2026). In a system where frequency has fallen almost every year since the early 1990s, severity growth of that magnitude, in both cost components simultaneously, is the detail that gets lost beneath the headline combined ratio of 91.

The calendar year figure and the accident year figure tell different stories. NCCI's accident year 2025 combined ratio came in at 102, above breakeven, with the calendar year result of 91 held down by favorable development released from older accident years (NCCI 2026 State of the Line, reported by Claims Journal and Insurance Journal, May 2026). Net written premium fell 0.2% in 2025 and the industry's aggregate reserve position remains $14 billion redundant (NCCI 2026 State of the Line, May 2026). A carrier or a state regulator reading only the calendar year 91 could reasonably conclude the line is stable. The accident year 102, paired with 4% severity growth on both loss components, says the current book is running hot and the redundancy cushion built from older years is what is keeping the composite number comfortable.

Metric	2025 value	Direction
Lost-time claim frequency	-2%	Improved, milder than long-term trend
Medical claim severity	+4%	Accelerating
Indemnity claim severity	+4%	Accelerating
Calendar year 2025 combined ratio	91	Profitable, 12th consecutive year
Accident year 2025 combined ratio	102	Above breakeven
Net written premium change	-0.2%	Contracting
Industry reserve position	$14B redundant	Cushion from older accident years

Source: NCCI 2026 State of the Line, presented at NCCI's Annual Insights Symposium, May 12, 2026.

Why AI Closure Speed and Rising Severity Are Not the Same Signal

AI claims platforms from vendors including CCC Intelligent Solutions' EvolutionIQ unit are simultaneously compressing the front end of workers comp development. CCC completed its $730 million acquisition of EvolutionIQ in January 2025 after announcing the deal in December 2024 (BusinessWire, December 2024), and by early 2026 the platform guided claims at nine of the top 15 U.S. disability carriers, with expansion underway into workers compensation and third-party casualty (CCC Q1 2026 earnings release; actuary.info reporting on CCC EvolutionIQ's loss development impact). Deloitte's 2026 survey of P&C insurers found that carriers deploying advanced AI in workers comp claims operations reported 12 to 19% lower total claims costs and 35 to 50% faster cycle times (Deloitte, 2026 P&C Insurance Industry Outlook). Those are real operational gains, concentrated in exactly the claim population AI triage and case management tools are built to handle: low-to-medium severity lost-time claims with straightforward musculoskeletal injuries, clear causation, and a predictable path to maximum medical improvement.

The mechanism that creates the actuarial problem is specific, not general. AI triage systems accelerate first notice of loss, route claims to the right adjuster or specialist within hours instead of days, and drive earlier medical intervention. NCCI has documented that claims first reported more than seven days after injury cost 20 to 45% more in total incurred than claims reported within the first three days (NCCI research, cited in NCCI's 2026 State of the Line materials), so a platform that compresses reporting lag mechanically pulls a meaningful share of a book's claims into the low-cost, fast-closing category. That is where AI's effect on the triangle stops. Complex musculoskeletal claims requiring multiple surgeries, occupational disease claims with delayed manifestation, and claims complicated by comorbidities in an aging workforce continue to develop in months 12 through 60, a window that sits almost entirely outside the AI-optimized closure period. Comorbidity loading on claims involving workers past age 50 is one of the specific drivers pushing medical severity higher in exactly the claim segment AI triage does not resolve early, which is also the segment generating most of the 4% medical severity growth NCCI reported for 2025.

Put the two effects together and the triangle bends in a direction that looks, at first read, like genuine improvement. Early development factors shrink because a larger share of claims closes inside the AI-accelerated window. Later development factors, applied to a shrinking population of complex claims that the platform never touched, do not shrink at the same rate, and in a rising-severity environment they should be widening. An actuary who selects age-to-age factors by blending the last five years of experience, the way NCCI's own methodology and most carrier-specific selections work, will see the recent AI-affected periods pull the blended factor down across the whole triangle, including the tail periods where the AI effect never actually operated.

Diagnosing the Pattern in Your Own Triangle

The diagnostic is specific enough to run against a carrier's own data without waiting for NCCI's countrywide factors to catch up. Pull the last five accident years of paid or incurred loss development by quarter, and compare two things against the prior five-year baseline: the first-diagonal age-to-age factors (12-to-24-month development) and the second- and third-diagonal factors (24-to-60-month development). The pattern to test for is a first-diagonal compression of roughly 8 to 12% relative to the carrier's own historical average, paired with second- and third-diagonal factors that hold steady or widen relative to prior-period selections rather than following the first diagonal down. If both halves of that pattern are present, the triangle is not describing a single improving process. It is describing two different populations, an AI-accelerated one and an unaccelerated one, blended into a single set of factors that no longer represents either.

A useful cross-check is claim count development alongside dollar development. If closed-claim counts at 12 months have risen relative to the historical baseline while average severity on the claims still open at 12 months has also risen, that combination is close to diagnostic on its own: the easy claims are leaving the open inventory faster, and what remains open is disproportionately the expensive tail. NCCI's own research on fast-emerging large claims supports the scale of that tail risk at the industry level: the share of workers comp claims that eventually exceed $1 million and reach that threshold within 24 months grew from 27% in accident year 2003 to 59% in accident year 2023, a structural shift that already required a rethink of tail factor selection before AI-driven closure speed entered the picture (NCCI fast-emerging large claims research, cited in NCCI's data on large WC claims emerging within two years). Layering AI-compressed early development on top of a large-claim population that already emerges faster and larger than it did two decades ago compounds the diagnostic signal rather than offsetting it.

The ULAE Blind Spot

Unallocated loss adjustment expense reserving inherits the same distortion, and it is the piece of this problem that gets the least attention because ULAE reserves are typically built from a historical ratio to paid losses or open claim counts rather than modeled independently. AI-accelerated closure front-loads claims handling activity and expense recognition into the first few months of a claim's life, which is exactly when the platform is doing the most work: intake triage, next-best-action routing, and automated documentation. The claims left open at 12, 24, and 36 months are disproportionately the complex cases AI did not resolve early, the ones requiring nurse case management, independent medical examinations, vocational rehabilitation, and litigation support, all of which are ULAE-intensive activities concentrated in exactly the tail the platform does not touch.

A ULAE reserve built from a historical ratio to paid losses will understate the cost of that remaining open inventory in two compounding ways. First, if the ratio was calibrated on a pre-AI period, it reflects a claims population that included more low-complexity claims still open at later maturities, diluting the average handling cost per open claim. Second, as AI closes the easy claims earlier, the open inventory at any given evaluation date becomes more concentrated in complex cases with above-average handling costs, so the same historical ratio applied to a smaller but costlier population produces a bigger understatement with each successive accident year. This is not a hypothetical: it follows directly from the mechanics of AI triage described above, and it means ULAE reserve adequacy deserves its own explicit review wherever a carrier's paid claim mix has shifted meaningfully since AI deployment, rather than being carried forward on the same ratio methodology used before the platform went live.

Recalibrating When the Anchor Period Predates AI

The practical fix starts with checking what period a carrier's loss development selections are actually anchored to. For most carriers, triangles built through accident year 2022 or 2023 reflect a claims-handling environment that predates widespread AI claims deployment; EvolutionIQ's own carrier count did not reach nine of the top 15 disability writers until early 2026, and CCC's workers comp and third-party casualty expansion, including an April 2026 multi-year agreement with Allstate, is newer still. A five-year blended selection made today that leans on 2019 through 2023 experience is anchoring the bulk of its weight on a pre-AI development pattern, even though the most recent one or two diagonals already show AI-driven compression. That mismatch is what produces the artificially fast early development this article describes: not because the actuary made an error, but because standard blending methodology assumes stationarity that AI deployment has broken.

Actuaries working through this recalibration have a few concrete options rather than a single correct answer. One is to separate the triangle by claims-handling regime, treating AI-managed and legacy-managed claims as distinct populations with distinct development patterns until enough AI-era experience has matured to support blended selections again. Another is to hold tail factors fixed at pre-AI levels while allowing only the early-period factors to reflect recent experience, on the theory that AI has not yet demonstrated it changes the ultimate cost of complex claims, only the speed at which simple claims close. A third, more conservative approach is to add an explicit margin to the central estimate that scales with the carrier's AI adoption penetration, disclosed as a distinct source of parameter uncertainty rather than absorbed silently into the point estimate. None of these approaches requires waiting for NCCI's countrywide factors to reflect the shift, because NCCI's own countrywide data is itself a blend across carriers with very different AI adoption timelines, and it will lag the same way an individual carrier's five-year blend lags.

What Rate Filings Should Reflect

The calibration question has a regulatory dimension that individual carriers cannot resolve on their own. State rate filings and actuarial opinions that rely on industry-average development factors, whether NCCI's advisory loss costs or a bureau state's promulgated factors, inherit whatever blend of AI-adopting and non-adopting carriers sits inside that countrywide data. A rate adequacy certification built on an industry LDF table that has not yet separated AI-driven closure speed from genuine severity improvement risks certifying a rate level that looks adequate on paper while understating the true cost trend for carriers still working through complex claims that were reported in 2023 or 2024 and have not yet reached the maturities where the AI-versus-legacy divergence becomes visible.

NCCI's annual loss cost filings extend loss costs, rates, and expected loss rates to three decimal places for every classification code in the 2026 cycle, a level of precision aimed at responsiveness at the classification level (NCCI 2026 loss cost filing documentation). That precision does not by itself resolve the AI-closure-speed question, because it operates on the same countrywide development data that blends AI-affected and unaffected experience. Actuaries submitting rate filings that lean on industry-average LDFs for workers comp should be prepared to document whether their own carrier's closure speed has diverged materially from the industry pattern, and if it has, whether that divergence has been tested against the diagnostic described above before the filed rate is treated as adequate. A regulator reviewing a filing that shows an unusually favorable loss trend alongside accelerating industry-wide severity, as NCCI's 2025 data shows, has grounds to ask specifically how AI-driven closure speed was separated from genuine cost improvement in the trend selection.

Why This Matters

The workers comp system NCCI described in May 2026 is profitable by every headline measure: a 91 combined ratio, a twelfth consecutive year in the black, and $14 billion of reserve redundancy sitting on the balance sheet. None of that headline data resolves the specific reserving risk this article describes, because the redundancy is concentrated in older accident years that predate widespread AI claims deployment, while the accident year 2025 combined ratio of 102 and the 4% severity growth on both medical and indemnity components describe the book that AI-accelerated closure is now actively shaping. A reserve review that credits the redundancy from 2018 through 2022 accident years while assuming 2024 and 2025 will develop the same way, simply faster, is applying a comfort level from one population to a fundamentally different one.

Actuaries who run the first-diagonal-versus-tail diagnostic on their own triangles now, before the next annual reserve review, gain a two-to-three-year lead on a problem that will otherwise surface as adverse development the way it typically does: quietly, in a single accident year, well after the pricing and reserving decisions that assumed the old development pattern have already been made.