Claims triage automation: the 2026 playbook

Claims triage automation has solved routing and severity scoring, and it has not solved what happens to the claims it routes to your SIU. Software now classifies each incoming claim at first notice of loss, scores its complexity and fraud risk, and sends it to the right handler in minutes. That part works. The problem is what sits downstream of the flag: a manual investigation queue that still moves at 14+ days per case and still covers only about a quarter of the claims triage refers to it.

This is the core argument of the post. Automating the top of the funnel without automating the investigation at the bottom does not close the fraud-loss gap - it widens it, because better triage flags more claims into a queue that was already over capacity. Below, we define what triage automation actually does, where it stops, how the rules-based and AI approaches compare, who occupies each layer of the 2026 stack, and which numbers a Claims VP should measure before signing.

If you want the broader map of what is already automated across the claims lifecycle, the companion piece on what is automated in insurance claims in 2026 establishes that FNOL, estimating, and payments are handled and that investigation is the last manual step. This post zooms into the triage layer specifically.

Claims triage automation is software that classifies each incoming claim - usually at or just after first notice of loss - by severity, complexity, line of business, and fraud risk, then routes it without manual sorting. A clean low-value claim goes to straight-through payment, a routine one to a fast-track adjuster, a complex one to a senior specialist, and a claim with fraud signals to an SIU referral.

It replaces the manual sorting that used to depend on an adjuster's judgment and queue position. Adoption is now mainstream: Insurance Journal's 2026 reporting found that 82% of carriers use AI for routine claims tasks such as data extraction and automated routing (Insurance Journal, March 2026), and the Insurance Information Institute reports that 80% of carriers use predictive modeling to detect fraud, up from 55% in 2018 (Insurance Information Institute). Triage scoring is no longer a differentiator; it is table stakes.

The distinction that matters for the rest of this post: triage decides where a claim goes. It does not, on its own, investigate what the claim contains. Detection is upstream; investigation is downstream. Triage and detection scoring widen the top of the funnel - they surface more suspicious claims and route them correctly. The investigation layer is a separate job that runs on the claims triage hands off, and in 2026 it is the layer most carriers still staff manually.

Manual triage is slow, inconsistent, and leaks fraud signal at the front door. When sorting depends on an individual adjuster's read of a claim and how deep it sits in a queue, similar claims get routed differently, severity gets misjudged, and fraud indicators that a model would surface get missed because no one had time to look. The cost of that miss is large at the line level.

Insurance fraud costs the US an estimated $308.6 billion a year (Coalition Against Insurance Fraud, via the Insurance Information Institute), and roughly 10% of P&C claims involve fraud. The loss concentrates by line, which is exactly where routing accuracy matters most. The Coalition's category breakdown, as reported by the III, puts the largest P&C-relevant pools where a misrouted claim costs the most.

The deeper cost is what manual triage feeds. A claim that an adjuster routes correctly to SIU still lands in a queue most carriers cannot clear. That is the structural leak: the loss does not come only from claims that are mis-triaged, but from correctly-flagged claims that are never fully worked. We mapped this in detail in why flagged insurance claims never get investigated, and it connects directly to the leakage math in the claims-fraud leakage pillar, where uninvestigated and under-investigated claims account for $30B+ in US leakage.

Automated triage runs in four stages on every incoming claim: data capture at FNOL, severity and complexity scoring, fraud-risk scoring, and routing - including straight-through processing for the cleanest claims. Each stage narrows what a human has to touch, and the largest cycle-time gains land at the simple end of the funnel rather than the complex end.

Severity and complexity scoring

The first job is sizing the claim. The system reads the structured FNOL data plus any extracted document and image data, then scores severity (likely cost and reserve) and complexity (how many parties, coverages, and investigation steps are involved). Low-severity, low-complexity claims become candidates for fast-track or straight-through handling. High-complexity claims are escalated to senior specialists before any payment logic runs. This is where carriers see the headline speed numbers: Insurance Journal reports some carriers see 80% faster processing on low-severity claims, because those claims can be handled without an adjuster touch.

Fraud risk scoring and the SIU referral

Running alongside severity is a fraud-risk score, often supplied by a detection vendor sitting on cross-carrier data. The score determines which claims become SIU referrals. The accuracy of this stage is where naive automation gets expensive: rules-based fraud flagging produces a 60-85% false-positive rate, meaning most flags are not fraud. A high-recall, low-precision flag still has to be worked by someone, so a noisy triage layer does not just waste model cycles - it manufactures investigation work the SIU cannot absorb. The trade-off between rules and ML at this stage is the subject of legacy rules versus autonomous AI fraud detection.

Straight-through processing for clean claims

Straight-through processing (STP) is the payoff of good triage: intake, validation, severity scoring, and payment all run by software with no manual intervention, for simple low-risk claims where the facts are clean and the payout sits within set tolerances. STP scales the simple end of the book and is what produces the 80% processing-speed figure. The deliberate constraint is that any claim that trips a fraud flag is pulled out of STP and routed to SIU. STP does nothing for those claims - and those are the claims that drive most fraud loss.

Automated triage stops at the SIU referral. It flags the claim, scores its fraud risk, and drops it into the investigation queue - and then the manual workflow takes over. Better triage produces more referrals; it does not produce more investigation capacity. That is the wedge: triage automation widens the mouth of the funnel while the investigation throat stays the same width, so the backlog grows.

The numbers make the gap concrete. A manual SIU investigation takes 14+ days per case and costs roughly $2,500. A single investigator clears around 10 investigations a month and carries 200+ cases. Because of that throughput ceiling, most US P&C carriers fully investigate only about 25% of flagged claims - the rest are paid, denied without full work, or queued indefinitely. Feeding that queue more accurately-flagged claims does not raise the 25%. The queue, not the flag, is the limit.

This is also where the industry's own data shows the gap. Insurance Journal reports that nearly two-thirds of carriers see a gap between their AI vision and reality, only 12% report fully mature AI capabilities, and only 7% have achieved scalable AI success (Insurance Journal, March 2026). The automation that scaled is the routine-task, triage-adjacent kind. The part that did not scale is the judgment-heavy investigation work downstream of the flag.

Closing this is the investigation layer, and it is the layer no triage or detection vendor occupies. Hesper AI takes a flagged claim and runs the full SIU playbook - 15+ investigation phases including document forensics, OSINT, statement cross-reference, timeline reconstruction, and financial pattern analysis - in parallel, in 2-4 hours instead of 14+ days. That lifts flagged-claim coverage from about 25% to 100% at roughly $150 per case instead of $2,500. Make every flagged claim investigable: that is the structural fix that turns better triage into lower loss rather than a longer queue. From fraud detection to fraud resolution is two jobs, and triage automation only solves the first.

The comparison a buyer actually runs is across three operating models: manual triage, rules-based automated triage, and AI/ML triage paired with autonomous investigation. The first two differ on speed and consistency at the front of the funnel. The third differs on what happens after a claim is flagged - which is the dimension that moves loss-ratio.

Read the table down the last three rows and the point is clear: the FNOL-to-routing speedup that rules-based triage buys is real but front-loaded, and it leaves the coverage and cost-per-flag lines untouched. Rules-based triage often makes the coverage problem worse by producing more flags for the same queue. Only the third column changes what happens after the flag.

That throughput delta is the mechanism behind the coverage shift. It is not that AI triage flags differently; it is that the investigation layer behind it can actually work every flag the triage produces.

The 2026 triage stack has three layers, and most carriers run vendors at each: claims-management systems where routing rules live, detection vendors that score fraud risk, and the investigation layer that resolves the flags. Mapping who does what avoids the common mistake of treating these as competitors when they are sequential.

At the routing layer, claims-management platforms - Guidewire ClaimCenter, Duck Creek Claims, and Snapsheet - hold the FNOL intake and the triage routing logic. They decide which queue a claim enters. At the detection layer, FRISS, Verisk ClaimDirector (a 0-999 score on ISO ClaimSearch data), and Shift Technology generate the fraud score that triage routing acts on. They route and prioritize; they do not investigate the flag.

Shift has extended into handler-assist agentic AI for adjusters with Shift Claims, which speeds the handler and the claim - per Shift's announcement, with figures around faster handling and high automation on routine claims. That is upstream of investigation: it assists the human handling the claim, not the autonomous resolution of a flagged one. The detection and routing layers are well covered in 2026. The investigation layer is where carriers still rely on manual SIU.

The gap every one of these layers leaves open is the same: triage automation determines where a claim goes, and detection scoring determines how suspicious it is, but neither determines what the investigation finds on the claims routed to SIU. That is the investigation layer. It has no incumbent vendor except manual SIU teams, and it is the layer Hesper is purpose-built for.

For the triage layer itself, the ROI metrics are straight-through-processing rate, FNOL-to-routing time, routing accuracy (how often a claim reaches the right handler first time), and adjuster time saved on sorting. Those are real and worth tracking. But the larger ROI question sits one layer down: of the claims triage routes to SIU, what share gets fully investigated.

For a Claims VP measuring loss-ratio outcomes, that ratio is the number that moves basis points. Most carriers can only investigate about 25% of flagged claims at roughly $2,500 each over 14+ days. Improving triage without improving that ratio grows the SIU backlog rather than reducing loss. The honest cycle-time number to ask a vendor for is end-to-end - FNOL to resolved finding - not FNOL-to-routing, because a fast triage feeding a 14-day investigation queue has compressed the cheap front of the process and left the expensive back of it untouched.

The human-oversight point matters for the SIU Director and the compliance reviewer: Insurance Journal reports 75% of claims professionals believe AI needs human oversight. That is consistent with the investigation model, not against it - the investigator's role shifts from execution to decision-making. The agent runs the 15+ phases and produces an audit-ready report; the human SIU lead reviews, overrides where needed, and signs. The coverage goes from 25% to 100% without removing the human from the defensible decision.

So the buyer scorecard has two tiers. Tier one, the triage metrics, tells you the front of the funnel is working. Tier two - investigated-flag coverage and cost per investigated case - tells you whether the automation actually reaches the loss. Carriers that measure only tier one report the vision-reality gap Insurance Journal documented; carriers that measure tier two are looking at the lever that closes it.