Claims triage framework: the five lanes, and why you only measure four

A claims triage framework is a prediction made minutes after first notice of loss, on a file nobody has read, about how a claim will behave over the next 40 to 400 days. Get it right and the claim lands with the person and the process that can close it at the right indemnity. Get it wrong and you pay for the error twice: once in handling cost, once in loss cost.

Most carriers now make that prediction well. Severity models, complexity scores, coverage validation and straight-through rules have made the fast-track and desk lanes fast and accurate, and the tooling to run them is mature. But the failure that costs money is not in the model and not in the threshold. A triage framework is only as good as its worst-served lane, and carriers only measure the lanes they can staff.

What follows is the decision architecture rather than the tooling: the five destination lanes and what each one optimizes for, the two error modes and why only one of them is visible, the lane where capacity rather than merit decides what actually gets worked, the regulatory clock that runs regardless, and a five-metric scorecard whose fifth metric most carriers cannot produce. For the category view of how the whole lifecycle automates around this decision, start with the complete guide to claims automation.

One clarification before the lanes. This is not a fraud post. Investigation appears here as one lane among five, and it gets the most attention for a structural reason rather than a thematic one: it is the only lane whose throughput does not scale with software.

Claims triage is the use of what is known about a claim at a single point in time to forecast how that claim will behave, and to route it accordingly. That makes it a classification problem, not a sorting rule. Classification problems produce false positives and false negatives, so a triage framework carries an error rate in both directions and can be scored on both.

The cleanest definition in the literature comes from Hindawi and Modlin in Predictive Modeling Applications in Actuarial Science, Volume 2: "The broad objective of claims triage is to use the characteristics of each individual claim at a specific point in time to predict some future outcome, which then dictates how the claim will be handled." Two clauses, and the second one is the one carriers underweight. Predicting the outcome is a modeling problem. Dictating how the claim will be handled is a capacity problem.

Milliman frames the same split operationally. In its March 2022 guide to claims triage, Michael Paczolt describes the objective as segmenting high-cost from low-cost claims "and then allocate the appropriate resources to optimally manage them." Segment, then allocate. Most carriers have done serious work on the segmentation half. The allocation half assumes the resources are there to be allocated.

The volume explains why this has to be a policy rather than a judgment call per file. Verisk's Annual Insurance Claims Trends Report, reported by Insurance Journal in May 2026, counted 31.6 million personal auto claims in 2025, 5.3 million homeowners claims (the lowest in five years, down 19% from 6.5 million in 2024), 1.8 million commercial auto claims (down 5%), and 710,000 commercial property claims (down from 910,000 in 2023). Sum those four lines and you get roughly 39.4 million routing decisions in a single year. That figure is a sum of four Verisk-reported lines, not total US property and casualty claim volume, but it sets the order of magnitude: routing at that scale is an automated policy with a measurable error rate, whether or not anyone measures it.

The four questions behind one route

A single routing decision is the collapse of several independent predictions into one destination. On most new claims a carrier is running some version of the following, in parallel:

• Severity: what will this claim end up costing, in indemnity and expense?
• Complexity: how many parties, coverages and handling steps will it take to close?
• Litigation propensity: is this heading toward representation or suit, and is the reserve adequate if it does?
• Fraud suspicion: is any part of this claim not what it says it is?
• Subrogation potential: is somebody else's money in this file?

Five questions, one destination. In practice only three of them move a file between lanes. Litigation propensity and subrogation are usually scores layered on top of a lane rather than lanes of their own: a subrogation flag does not pull a claim out of the desk lane, it adds a work item and a diary date to it. Severity, complexity and fraud suspicion are the three that change the destination, and they answer different questions on different data, which is the first structural reason one score cannot drive a whole framework. Triage is stage two of eight in the claims process from FNOL to settlement, and every stage after it inherits whatever this stage got wrong.

Most US carriers run about five destination lanes: straight-through processing, fast-track or express desk, standard desk adjuster, complex and large-loss with litigation, and investigation. Each lane has a different objective function. That is why one score cannot drive the whole framework, and why lane definitions matter more than the threshold values people spend meetings arguing about.

The routing rules themselves physically live in the claims management system - Guidewire ClaimCenter, Duck Creek Claims or Snapsheet - and the scores that feed them come from a mix of internal models and vendor inputs such as Verisk ClaimDirector, FRISS or Shift Technology. That division of labor is stable and it is not the subject here. If you want the vendor and ROI view of the tooling, claims triage automation: the 2026 playbook covers the stack, the build-versus-buy question and the payback math. This post sits one layer above that: what the lanes are, what each one is trying to be good at, and what happens to the file after it lands.

Straight-through is not fast-track

The two get used interchangeably and they are different lanes with different owners. Straight-through processing means no human touches the claim: intake, validation, scoring and payment all execute in software. Fast-track means a human adjuster handles it on a compressed workflow with fewer mandatory steps. A carrier that reports a combined number is reporting two operating models as one, and cannot tell you which of them is producing its cycle-time gain.

The ceiling on true straight-through is narrower than the marketing suggests. Lemonade, a digital-native carrier writing renters, homeowners and pet, disclosed in its FY2025 annual report that "96% of the time, it is AI Jim that will take the first notice of loss from a Lemonade customer without human intervention" and that "roughly 55% of our claims were automated, resulting in instant or near-instant processing from start to finish." That is the digital-native ceiling on a book built specifically for it, in low-severity categories, with no bodily injury in the mix. It is not an industry average and it is not a target for a carrier writing commercial auto.

The more typical gain is in the fast-track lane. Sedgwick reported that using AI to handle low-severity claims "has led to 80% faster processing times for some carriers," per Insurance Journal in March 2026. The same piece carries Joel Raedeke, Senior Vice President for US Technology at Crawford & Company: "As AI drives more claims automation, we will see more straight-through processing of low complexity claims in 2026." Both statements describe the light lanes getting lighter. Neither describes anything happening in the heavy ones. For a stage-by-stage map of what is genuinely automated across the lifecycle today, see what is automated and what is still manual in 2026.

The two heavy lanes, and why they are worth getting right

The large-loss lane is a rounding error in volume and a large share of the money. NCCI's 2020 analysis of workers compensation large losses found that claims above $1 million "account for 10% of losses, but less than 0.2% of all lost-time claim counts," per NCCI Insights. Two tenths of one percent of files carry a tenth of the losses. The expected value of routing one of those correctly dwarfs the expected value of routing a thousand fast-track files correctly, which is why severity models get the modeling budget.

The investigation lane works differently from all four others. It is the only lane a claim enters on a suspicion rather than on a measurement, and the only one where the routing rule is deliberately tuned for recall rather than precision. Rules-based fraud flagging runs a 60-85% false positive rate in Hesper's benchmarks of deployed detection rule sets, which is a design decision rather than a defect: a detection layer that misses fraud is worse than one that over-refers, and a fraud flag correctly pulls a claim out of straight-through processing regardless of how small the loss is. The consequence is that this lane receives far more volume than any of the others were designed to absorb, relative to what sits behind it.

A triage framework fails in two directions. Over-triage sends a routine claim into a heavier lane than it needs. Under-triage sends a claim that needed attention into a light lane and closes it there. Both cost money. Only one of them generates a signal that ever reaches the person who owns the routing rules.

The under-triage number that should be on every claims operations wall comes from the same NCCI analysis: "One study estimated that two thirds of claims over $1M started out as 'fairly routine.'" Two thirds of the files that eventually became the most expensive claims in the book presented, at first notice, as ordinary. Whatever the severity model saw at FNOL, it was not enough to distinguish them from the population they came from.

That is a hard problem to model and an impossible one to learn from your own feedback, because the feedback never arrives. A fast-tracked claim that should have been escalated does not raise its hand. It closes, on time, inside SLA, and by the time the reserve moves the routing decision is several quarters old and the deterioration gets attributed to medical inflation or a bad venue or an aggressive plaintiff bar. The causal link back to the lane assignment is not in anybody's report.

So carriers tune against the error they can see. Every threshold conversation in a claims operations meeting is about the noise the heavy lanes are receiving, because the heavy lanes are where the complaints originate. The tuning is locally rational and directionally one-sided: it pushes routing toward under-triage, which is the expensive error. The money that walks out of the file as a result is the same money tracked in claims leakage from uninvestigated claims, arriving through a different door.

Reopened files are a symptom, not a benchmark

There is no authoritative published reopen rate for US property and casualty claims, and the numbers that circulate in vendor material do not carry a study behind them. Treat reopening as a qualitative signal about your own book rather than something to benchmark against an industry figure. The useful exercise is retrospective and internal: pull the files that reopened or developed adversely over the last eight quarters, look at what they had in common at FNOL, and check whether the routing rules would send those same characteristics to a light lane today. That is a slow, unglamorous audit, and it is the only direct read on under-triage a carrier can get.

Four of the five lanes scale with software. Straight-through, fast-track, desk and large-loss all absorb more volume when a carrier adds compute, better models or a cleaner workflow. The investigation lane scales with investigators. When referral volume rises and headcount does not, the routing decision stops determining the outcome and the queue starts.

The 2024 Insurer SIU Benchmarking Study from the Coalition Against Insurance Fraud and Aon, published in September 2024 across 35 property and casualty carriers, is the sixth iteration of that study and the only public dataset that measures what happens to a routing decision into this lane after it is made. Three of its numbers, read together, describe the whole problem.

First, automated detection tools now generate 45% of all referrals into the investigation lane, up six percentage points from the prior study, against 49% from adjusters. Second, of referrals that are accepted, automated tools account for only 29%, with adjusters at 60% and all other sources at 10%. Third, the acceptance ratio for automated-tool referrals is 36%, up from 32%, while the ratio for adjuster referrals is 70%, down from 76%. Divide those two and an automated routing decision into the investigation lane is about half as likely to be accepted as an adjuster's referral. That ratio is derived from the study's reported figures rather than stated in it.

There are two readings of that gap and both are partly right. The first is that automated referrals are noisier than adjuster referrals, which is true by construction, since detection is tuned for recall and an adjuster arrives with context a rule set does not have. The second reading is the one the capacity data supports: acceptance is not purely a quality judgment, it is a commitment to work a file. When the lane is full, acceptance criteria tighten, and the tightening lands hardest on the source with the most volume and the least narrative attached to each item.

The capacity side of the same study is unambiguous. Anti-fraud headcount grew 0.5% between 2023 and 2024, against 1.4% in the prior study. SIU staffing fell to 0.67 full-time equivalents per $100 million of gross written premium, down from 0.9 two years earlier. SIU budget sat at 0.12% of overall premium, down from 0.13%. Per investigator, accepted referrals rose to 174 a year from 162, with desk investigators carrying 17.9 accepted referrals a month and field investigators 11.6. Referral volume from automated sources is up, the capacity line is flat, and the load per investigator is rising. None of that is a model problem.

Set that against the baseline for the work itself. A manual investigation runs 14 or more days per case, one investigator carries 200 or more open cases, and the arithmetic lands at roughly 10 completed investigations per investigator per month. Full investigation therefore reaches about 25% of flagged claims. The other three quarters are paid, denied without full work, or aged in a queue until somebody closes them to make room.

This is the point where a triage framework quietly stops being a framework. The router says "investigate." The queue says "close it." Both statements live in the same claim file, and only one of them is in the metrics.

The same funnel shape shows up one level downstream, at the state agency that receives what carriers refer out. In fiscal year 2023-24 the California Department of Insurance Fraud Division reported that on the automobile side it received 12,559 suspected fraudulent claims, assigned 602 new cases, made 272 arrests and referred 354 cases to prosecuting authorities, with potential loss of $207,629,944. On the property, life and casualty side it identified and reported 4,580 suspected fraud claims, assigned 52 new cases, made 16 arrests and referred 21 submissions, with potential loss of $362,155,588.

Derived from those figures, 4.8% of auto referrals and 1.1% of property, life and casualty referrals became assigned cases. Two caveats matter. These are the state's criminal case assignments, not carrier SIU investigations, and the Fraud Division's mandate is prosecution rather than claim resolution, so a low assignment rate is expected and is not a failure. The reason to look at it is the shape: another investigation lane whose narrow end is set by the number of people available to work it, not by the merits of what arrives. Carriers face the same geometry with more volume and a tighter clock. The operational side of living inside it is covered in clearing a high-volume investigation backlog without adding headcount.

Automated routing sits under three kinds of rule: AI governance obligations covering how the model is built, documented and monitored; unfair-claims-practice standards covering whether an investigation happened at all; and statutory deadlines covering when the answer is due. None of them pause because a queue is full, and the floor is highest in the lane with the least capacity behind it.

The AI governance layer

The NAIC adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers on December 4, 2023. As of April 1, 2026, 25 jurisdictions had adopted it, with four more - California, Colorado, New York and Texas - running their own insurance-specific AI regulation or guidance instead. Alaska was first, on February 1, 2024; Hawaii is the most recent on the NAIC list, at December 10, 2025. A triage model is squarely inside the scope of these bulletins, because it is an AI system used in a decision that affects a consumer.

Colorado went further and put dates on it. The amended Regulation 10-1-1 (3 CCR 702-10) took effect October 15, 2025, with an interim narrative report due December 1, 2025 and full compliance required by July 1, 2026, followed by annual reporting. As Faegre Drinker summarized in September 2025, the amended rule covers individual life, private passenger auto and health benefit plans using external consumer data and information sources, algorithms and predictive models, and it requires board oversight, a cross-functional governance group, written policies for designing, testing and monitoring, an inventory of the data sources in use, ongoing algorithm performance monitoring and a documented third-party vendor selection process.

The phrase to sit with is ongoing algorithm performance monitoring, because it is not satisfied by a model validation report written at implementation. A routing model's performance is what happened to the claims it routed. If the destination lane does not record an outcome, the monitoring obligation has nothing to monitor, and the honest answer to a regulator asking how you monitor your triage model is that you monitor four of its five outputs.

The unfair-claims-practice layer

The NAIC Unfair Claims Settlement Practices Act, Model #900, lists the acts that constitute an unfair claims practice. Two of them describe a starved investigation lane precisely. Section 4.C: "Failing to adopt and implement reasonable standards for the prompt investigation and settlement of claims arising under its policies." Section 4.F: "Refusing to pay claims without conducting a reasonable investigation." A third, Section 4.K, opens on "Unreasonably delaying the investigation or payment of claims," although its full text is aimed at a narrower practice, requiring both a formal proof of loss form and a subsequent verification that duplicates it.

Section 3 supplies the test that turns any of those from an incident into an exposure. An act is an improper claims practice when "It has been committed with such frequency to indicate a general business practice to engage in that type of conduct." One file that sat in a queue is a file. A routing rule that sends a fixed share of claims into a lane that documents an investigation on a quarter of them is a pattern, and the pattern is recorded in the carrier's own claims system with timestamps. Frequency is exactly what an automated router guarantees.

The clock

California is the clearest published clock. Under 10 CCR 2695.7, subsection (b) requires that "Upon receiving proof of claim, every insurer...shall immediately, but in no event more than forty (40) calendar days later, accept or deny the claim, in whole or in part." Subsection (c)(1) requires written notice "every thirty (30) calendar days until a determination is made or notice of legal action is served." Subsection (d) requires that "Every insurer shall conduct and diligently pursue a thorough, fair and objective investigation."

Put the three together with the capacity numbers from the previous section and the conflict is arithmetic rather than legal. Forty calendar days to a decision. A thorough investigation required behind the decision. An investigation lane running 14 or more days per case at 200 or more cases per investigator. The clock does not pause because the queue is full, so what actually happens under time pressure is that the file receives a decision without the investigation the routing rule called for. The state-specific requirements around documenting that work are covered in the California 10 CCR 2698 SIU compliance guide.

A triage framework is scored on five things: straight-through rate within eligible claim types, first-time-right routing, lane-change latency, cycle time measured separately by lane, and worked-referral rate. The first four are standard claims operations reporting and most carriers can produce them with effort. The fifth is the one that closes the loop, and most carriers do not have it.

Metrics one through four all measure the router. Metric five measures the lane. A framework can score well on the first four and still be sending its highest-exposure files into a destination that closes most of them unworked, which is why the fifth metric is not a nice-to-have addition to the set. It is the only one that tells you whether the routing decision was a decision.

Compute it honestly or do not compute it. A documented investigation means an outcome record on the file: what was checked, what was found, what the disposition was and on what basis. A case-opened timestamp is not an investigation, and a referral accepted is not an investigation performed. The CAIF acceptance ratio measures the first step in that chain; worked-referral rate measures the last one. If your number comes back high on the first pass, check the definition before you celebrate it. The wider metric set for the lane itself is in the 12 SIU KPIs worth tracking in 2026.

One accounting note, because these two figures get set against each other and they should not be. The CAIF and Aon study puts budget expense per investigation just above $1,200, up around 11% from 2022. That number is SIU budget divided by investigations. A fully loaded per-case figure that includes investigator time, systems, vendor data and supervisory review runs closer to $2,500 in Hesper's internal benchmarks. Different numerators, different questions. If you are building a business case, say which one you are using and hold it constant.

When the investigation lane is full, the instinct is to move the threshold. Raise it and referral volume falls without raising the share of genuine cases that get worked. Lower it and the queue gets worse. Threshold tuning redistributes the two error modes between each other; it does not add capacity, and capacity is the binding constraint.

The arithmetic is worth stating plainly because it is what makes threshold conversations circular. Suppose the lane can complete N investigations in a period and routing sends it 3N files. Moving the threshold so that it sends 2N does not change N. All that changes is which files are in the 2N, and they are being selected by a score built for recall, on the assumption that a human will sort out the rest. The rest is not being sorted out. It is being closed to make room.

So the fix has to happen in the destination lane, and that is a different kind of purchase from anything on the detection side. Detection is upstream; investigation is downstream. FRISS, Shift Technology and Verisk score the claim and effectively set where the threshold can sit. Guidewire, Duck Creek and Snapsheet execute the route and hold the queue. All of those layers work. The layer that has had no software incumbent, only a human queue, is what happens between a file landing in the investigation lane and a documented outcome coming out of it.

That is the layer Hesper AI is built for. Autonomous investigation agents run 15 or more investigation phases in parallel on every routed file - document forensics, OSINT, statement cross-reference, timeline reconstruction, financial pattern analysis - and return an audit-ready record in hours rather than weeks. The effect on the framework is coverage rather than speed for its own sake: investigation of flagged claims moves from roughly 25% to 100%, throughput per investigator moves from about 10 investigations a month to 800 or more, and fully loaded cost per investigated case moves from roughly $2,500 to roughly $150. Make every flagged claim investigable, and the routing decision starts producing an outcome you can measure.

This is not a headcount argument and it should not be sold as one. The investigator's role shifts from execution to decision-making: the human opens a file that already has the timeline, the conflicts, the source list and the exhibit index assembled, and spends their hours on the judgment call rather than on building the record. The volume that used to be triaged out of the lane by attrition is the volume that now has a documented answer attached to it.

There is a governance dividend on top of the operational one. When every phase writes a cited, timestamped finding as it runs, the worked-referral rate becomes computable for the first time, the reasonable-investigation standard in Model #900 has evidence behind it, and the ongoing algorithm performance monitoring that Colorado now requires has something to monitor. A routing decision with a recorded outcome is a routing decision you can score, retune and defend. Where this sits relative to every other stage of the lifecycle is laid out in the claims automation pillar.

A triage framework is a set of promises about where claims go. Four of those promises are kept by software that already exists and works. The fifth is kept by whoever happens to be available that week. Until the investigation lane has capacity behind it, every threshold conversation is really a conversation about which files to fail quietly.