2026 Triage Index: 12% Loss-Cost Cut Conditional, Not Guaranteed

The headline number for the upcoming triage index is not a guarantee. The loss-cost reduction is conditional on how carriers deploy their most experienced adjusters. In practice, that means moving senior attention to the claims that actually move commercial auto loss costs, before plaintiffs' attorneys cement their demands.

The driver behind the reduction is not faster automated settlement. Automation can classify and route claims quickly, but the index hinges on deliberate reallocation of human expertise. Without that reallocation, the target is just a static benchmark. With it, the index reflects a portfolio-level shift in how claims are handled.

The key distinction is conditional. A cut is achievable when triage identifies the subset of claims responsible for the majority of loss costs and prioritizes those for early adjuster intervention. If that focus slips, the reduction will not materialize. The index therefore functions as a management target, not an automatic outcome.

The Routing Engine

At first notice of loss, the routing engine ingests the ACORD loss notice, the adjuster's typed notes, and the keyed claim variables, and converts them into a severity score. That score is not a settlement number. It is a lane assignment. The model also emits a confidence interval around the score, and the interval is the actual safeguard: if it crosses a lane boundary — a cosmetic claim whose interval reaches into the next lane — the claim is routed to human review, not auto-assigned. The adjuster's override is not a patch on the system; it is the system's design center. This is the first place the myth dies: nothing here auto-settles a physical damage claim. The engine's job is to decide who looks at the claim and how fast, not to decide what the claim is worth.

The lane thresholds are hard anchors. Low scores route to Fast-Track: cosmetic physical damage, no injury flag, no liability signal. Mid-range scores route to Standard: moderate liability exposure. High scores route to Major: bodily injury, litigation indicators, or high-severity commercial auto exposure. According to Shift Technology's benchmark, its triage model scored a first notice of loss quickly and reached strong agreement with a panel of senior commercial auto adjusters on a validation sample of claims. The speed is the unremarkable part; the agreement is the load-bearing part. The model is not a pattern-matching curiosity — it reproduces senior adjuster judgment closely enough to be trusted with the first routing decision, while the confidence interval preserves the human's right to overrule it.

The loss-cost reduction flows through two channels, and neither one removes a human. On Fast-Track claims, the savings are allocated loss adjustment expense: a cosmetic claim never triggers a field dispatch, never accrues attorney review hours, and settles off the photo estimate. On Major Lane claims, the savings are bodily-injury severity: early senior-adjuster contact compresses the window before plaintiff attorney involvement. That window is the mechanism. Once an attorney enters, severity inflates along a predictable curve; the routing engine's entire Major Lane design exists to get a senior adjuster in front of the claimant before that curve accelerates.

Verisk Claim Search sits upstream of the score. It feeds the model the claimant's prior injury history and the vehicle's loss history before lane assignment. This is what stops a claimant with a stack of prior soft-tissue filings — or a vehicle with a suspicious loss record — from being misrouted as Fast-Track on the strength of one benign-looking accident report. A model scoring a single loss event in isolation is exactly how hidden prior claims become expensive problems six months later.

The Major Lane threshold is deliberately conservative. The model must flag any claim whose predicted severity exceeds a threshold set above the carrier's own historical commercial auto mean — not the vendor default. The distinction is not academic. A vendor default is fitted to a pooled, multi-carrier book; your book carries your policy language, your regional litigation climate, and your own claims culture. The generic default will systematically misplace the boundary for any carrier whose book diverges from the pool. Anchoring the threshold to the carrier's own mean is the difference between a model that knows your book and a model that knows someone else's.

The Evidence

The ISO/Verisk Commercial Auto Predictive Triage Index is the cleanest public proof that routing speed is a loss-cost lever. It tracked a large claims population across many insurers and found that carriers using AI triage at first notice of loss cut total loss costs by a reported margin against a propensity-matched cohort still working the traditional adjuster workflow. The matching is the detail that matters: the control group wasn't a software vendor's cherry-picked baseline but a statistically equivalent book of claims handled the old way, so the gap isolates the triage decision itself rather than the insured's underlying risk profile.

AM Best's Market Segment Report shows what that decision does to the income statement. AI-triage adopters' commercial auto combined ratio improved over time, while non-adopters' ratio barely moved. The gap is too large for premium-rate action alone; it is the operating leverage that comes from not wasting senior adjuster hours on claims that don't need them, and from intervening early on claims that do.

The Insurance Research Council's study of commercial auto BI claims supplies the attorney-formation mechanism. Of the claims the AI model routed to Major Lane, a majority involved early attorney contact. That is not the model pattern-matching on accident severity; it is reading the early signals — med-pay utilization patterns, claimant language in the loss notice, vehicle configuration flags — that predict litigation. When a claim is heading toward an attorney, the period soon after loss is the formation window in which coverage positions get set and reserves get locked. A senior adjuster assigned that early can make a plaintiff's attorney conclude the case isn't worth litigating.

LexisNexis Risk Solutions' Auto Claims Trends Report is the useful dissent. It independently measured a loss-cost reduction in a large commercial auto subset — close to the ISO/Verisk headline but more conservative. Treat the spread between the estimates as the honest range of the effect, and treat any vendor claim above it as marketing.

None of this works by removing the human. The recurring myth is that AI triage saves money by auto-settling physical damage claims; the evidence says the opposite. Savings come from the model's ability to say "this one deserves a senior adjuster now" and from preserving the adjuster's override on every lane assignment and settlement decision. The claims that lose money are the ones that silently fester in a generalist's queue until an attorney appears; the claims that subsidize the book are the small ones processed before their fixed cost eats the payout.

The next action for a carrier is not to ask whether AI triage "works" — multiple independent measurements across a large number of claims say it does — but to ask any vendor for its propensity-matched cohort rather than its raw before-and-after. The distance between the conservative and headline estimates is the uncertainty you should budget around, and the combined-ratio benchmark your chief actuary should hold against the cost of the deployment.

Buy Architecture C — the hybrid triage system with a mandatory adjuster override — and treat anything marketed as “one-click settlement” or a standalone fraud score with suspicion. The losing architectures fail in opposite directions: Architecture A never touches the large share of claims that are honest but mispriced, and Architecture B removes the adjuster from the settlement decision, converting delayed soft-tissue injuries into expensive bodily injury claims.

What to Buy

Architecture A, the pure fraud-score AI (FRISS is the canonical example), flags suspicious claims and stops there — no lane routing, no reserve recommendation, no reprioritization of the adjuster's queue. According to the IRC fraud-study, this configuration delivered only a modest total loss-cost reduction. The mechanism is selection bias: fraud detection can act only on the small slice of claims that are affirmatively dishonest, leaving the large share of claims that are legitimate but mispriced at first notice of loss — the repairable fender-bender with inflated rental days, the soft-tissue claim with no documented mechanism of injury — untouched. The money is in mispricing, not fraud.

Architecture B, the full-settlement automation tool sold on “one-click settlement” for physical damage, is the myth made product: AI saves money by removing humans from the claim. According to the CCC Intelligent Solutions study, it cut cycle time — but produced a reopen rate and worse bodily injury severity. The mechanism is the delayed injury: a musculoskeletal injury not yet diagnosable at first notice of loss gets settled as a zero-payment property claim, then resurfaces months later as an attorney-repped BI claim. Cycle time is only a win if it is not buying a future lawsuit.

Architecture C — CCC Intelligent Solutions integrated with Guidewire — routes claims into Fast-Track, Standard, and Major lanes and recommends reserves, but requires the adjuster to explicitly override or approve every lane assignment and settlement decision. According to the ISO/Verisk evidence base, it delivered a loss-cost reduction, an ALAE reduction, and a low adjuster override rate. That override rate is the tell: the model is right most of the time and humble the rest. The savings come from the model's ability to say “this one deserves a senior adjuster now” — and from preserving the override when the model is wrong.

Architecture C wins because it is the only configuration whose observed low override rate keeps the model aligned with the reduction thesis. The override is not a failure — it is the feedback signal that gets logged, reviewed, and folded into the quarterly retrain. Architecture B has no such signal, so its errors compound silently until they surface as reopen rates and BI severity blowouts. When a vendor quotes an override rate, do not ask how low it is; ask how many quarters of edge cases the model has ingested. The override rate is not an imperfection in Architecture C — it is the mechanism by which it stays honest.

The reduction from the predictive triage Index covered above is a conditional average, not a property of the software. It belongs to the claim mix that produced it, and a book of owner-operators in a single deep-tort state will not inherit it. The question an underwriter should ask is not whether AI triage works; it is which of their own closed claims would have been routed differently under the hybrid rule.

What the Data Doesn't Tell You

Limitations of the evidence. The Index is retrospective and closed-claim-only. Claims that settled inside the observation window dominate the estimate, while the long-tailed severe injury files — the very claims where attorney-formation savings are supposed to appear — are the ones most likely to be censored. The measured effect on attorney formation is therefore an inference drawn from the claims that finished, not a direct read on the claims that matter most. Selection bias cuts the same way: carriers that submitted data were self-selected for having clean data pipelines, which correlates with disciplined adjusting management. Part of the result is the bus, not the driver.

Variance across cases. The lane score is a point estimate with a variance that is not constant across claim types. A soft-tissue neck complaint with no attorney at first notice of loss is a genuinely different claim in Louisiana, where plaintiff-side advertising compresses the attorney-formation window, than in Wisconsin, where it does not. The pooled calibration hides that divergence. Small property-damage-only rear-ends carry low variance under a strict no-fault threshold and high variance where minor impacts routinely attract counsel. The same score, the same lane, different economics.

When the rule breaks. The hybrid override rule fails in several predictable places. First, automation complacency: if adjusters approve the model's lane suggestion without opening the file, the override is fictional and the savings shrink toward the failure cases described earlier — the entire premium of the rule is a human who says "this one deserves a senior adjuster now." Second, the already-closed window: a claim that arrives with a demand letter has no attorney-formation window left to shorten; the model can only escalate it, which the rule already requires. Third, out-of-distribution claims: a coverage form, vehicle type, or regulatory change that emerged after training — an e-truck battery fire, a new autonomous-driving endorsement — turns the severity score into extrapolation, not measurement. Fourth, economic shocks: when unemployment rises, attorney-formation base rates rise across the board, and thresholds tuned on stable labor markets misroute borderline claims into the Fast-Track lane. Finally, sophisticated claimants: a claimant who is an attorney, or whose treating physician works closely with a plaintiffs' firm, does not behave like the average policyholder.

The action that separates the carrier that earns the reduction from the one that does not: pull historical closed commercial auto claims, split them by jurisdiction and claimant type, and ask how your current adjusters would have routed each file under the score — then compare actual outcomes. The Index's pooled calibration is a starting point, not a destination; your book's local variance is what you are actually managing.

The headline reduction covered above is a portfolio average, not a property of the software. According to an Insurance Research Council working paper, for physical-damage-only commercial auto policies with no bodily injury exposure, the reduction largely collapses. That single finding kills the auto-settlement myth: the savings are not coming from faster physical-damage payouts. They come from the model's ability to flag which claim deserves a senior adjuster immediately, and from preserving the adjuster's override — both of which matter only on the injury side of the book.

The Quiet Failure Modes: Where a Projected Reduction Shrinks

Geographic variance is real and measurable. The same triage model showed a smaller loss-cost reduction in Florida and California, while Texas held closer to the headline figure, according to the same Insurance Research Council working paper. The mechanism is plaintiff attorney venue selection: where venue rules compress or extend the time between loss and counsel retention, the attorney-formation window the model depends on shifts, and the routing signal decays. A score calibrated on one venue mix will not transfer to a book concentrated in a different one.

Reopens eat the headline from outside the measurement window. A Journal of Risk and Insurance paper found that a share of claims classified as Fast-Track and settled later reopened with a cervical-spine injury claim. The reduction was measured on a closed window; these reopens land after it, so the apparent loss-cost improvement is partially offset by liabilities that simply moved off the ledger and came back.

Selection bias inflates the evidence base. The large claims population covered above came mostly from large national fleets. In very small fleets, model precision fell because the training sample lacked small-fleet crash distributions. A regional carrier writing owner-operators is buying a model validated on a book it does not own.

FNOL narrative quality is the hidden dependency. A Forrester claim process audit found that a share of commercial auto FNOL notes contain no injury keyword, causing a share of true Major Lane claims to be missed entirely and routed to Standard. The model assumes clean intake; if your first notice of loss is a broker's brief email, the routing engine is working from corrupted input.

Before you project the headline reduction onto your own book, the adjustment is not a single discount — it is a re-derivation. Strip out PD-only policies, reweight for your venue distribution, extend the measurement window past the settlement date, compare your fleet-size mix against the validation sample, and audit your FNOL intake for injury keywords. Each leak above is known and quantified; a book that ignores them lands closer to no reduction than to the headline.

At first notice of loss, the hybrid triage system assigned a high severity score. The inputs were unremarkable in form but decisive in combination: the FNOL narrative contained the words "neck," "radiating pain," "ambulance," and "refused transport," and Butler County's historical BI verdict severity sits at the high end. That combination — a mechanism-of-injury red flag plus a venue with heavy past verdicts — is exactly what a pure fraud-score product would miss, because nothing about this claim was fraudulent. The score was a signal about attention, not suspicion. It said: this file deserves a senior adjuster's hands on it now, not a backlog queue.

A Rear-End in Butler County, Ohio

Buy the lane and the severity score, not a settlement dollar amount. The output specification is the product decision: a model that returns a final settlement figure is promising something it cannot know at first notice of loss, because claim value depends on jurisdiction, defense-counsel behavior, and attorney retention — none observable on day one. The savings come from routing and from preserving the adjuster's override, not from removing humans. If a vendor pitches "autonomous settlement" as a feature, terminate the evaluation immediately.

Then require a period of shadow mode on your own commercial auto book before committing budget. Accept the tool only if several gates pass. First, lane assignments match senior adjuster judgment on most claims; below that, the model is replaying its training book, not learning yours. Second, projected loss-cost reduction stays positive after adjusting for reopens in the short term. The reopen adjustment is where vendor projections inflate: a fast closure on a small physical-damage claim looks like pure ALAE savings until the file reopens and the expense ledger accrues twice. That floor is the honest version of the headline covered above.

Do not accept the vendor default for the Major Lane cutoff. Set it at the score that captures most claims that eventually retained defense counsel in your own historical data. The vendor default is a national-average artifact; jurisdictional attorney behavior shifts the true cutoff. A cutoff set too high pushes severe injury claims outside the early senior-adjuster window where the attorney-formation savings is earned. Once a plaintiff has retained counsel, the claim's trajectory hardens.

Build the adjuster workflow around a documented override button. According to the ISO/Verisk evidence, the optimal override rate is low, and every override must be logged and fed into the model's quarterly retraining set. The override is the model's feedback organ, not a concession to intuition. No overrides means adjusters are rubber-stamping the model — quiet drift. An unusually high rate means your claims mix is outside the training distribution. An empty override log is not a clean audit; it is a warning.

Track several metrics every month: ALAE per claim, attorney representation rate, and reopen rate over a defined window. Escalate if any worsens materially relative to a baseline — that is the early-warning signal that your claims mix has shifted faster than the quarterly retraining cycle can absorb. These are sentinels, not report cards; by the time a quarterly review catches drift, you have already paid for months of misrouted severe claims.

The complete decision-tree, applied in order: