Insurance claims automation in 2026: what's automated, what's still manual

Claims automation has genuinely arrived, but it is concentrated at the front of the pipeline and thin everywhere judgment is required. For a claims-transformation owner, the practical question in 2026 is no longer whether to automate - it is which stage is left to automate, because the easy wins at intake, triage, and estimating are mostly booked. The harder, higher-value stage is the one that has lagged the rest: investigation of flagged claims.

The adoption data is broad and the scaling data is narrow, and the gap between them is the story. A Sedgwick report, reported by Risk & Insurance, found 82% of carriers now use AI tools somewhere in their operations, yet only 7% of property carriers have achieved scalable AI success. A Bain & Company survey of 160 global insurers, also reported by Risk & Insurance, found 78% of P&C insurers have adopted generative AI but only 4% have scaled it meaningfully across claims operations, and just 27% are pursuing comprehensive claims transformation.

Read those numbers together and the pattern is clear. Most carriers have AI running on discrete, high-volume tasks - data extraction, documentation, low-severity claims - rather than re-engineering the harder stages. The loss-ratio impact has landed where the work is repetitive. The stages that demand judgment, including the investigation of suspicious claims, are where adoption stalls. This post maps the full claims-automation stack layer by layer, places the named vendors at each layer, and isolates the one layer that still has no software incumbent.

If you want the full definition of the stage at the end of that map, our pillar on autonomous AI claims investigation covers what it is and how it differs from detection.

Walk the claims lifecycle stage by stage and the automation footprint is easy to trace. Each stage has matured at a different rate, but the front three-quarters of the pipeline now run with meaningful software automation at most mid-market and national carriers.

FNOL and intake

First notice of loss and intake are largely automated. The Sedgwick report found intake automation cut average processing time from 10 days to 36 hours. Mobile-first FNOL and digital intake run on platforms like Snapsheet and the FNOL modules inside Guidewire ClaimCenter and Duck Creek Claims, which capture the loss, validate the policy, and open the case file without a human keying it in.

Triage and straight-through processing

Triage and straight-through processing now auto-approve low-complexity claims. Crawford & Company's Joel Raedeke, SVP of U.S. technology, told Claims Journal that "as AI drives more claims automation, we will see more straight-through processing of low complexity claims in 2026," with simpler claims passing through all decision gates to auto-approval. The Sedgwick data backs this up: low-severity claims now process up to 80% faster with AI in the loop.

Damage estimating

Damage estimating is automated through computer vision and structured estimation utilities. Tractable's visual damage assessment compressed cycle time from days to minutes on specific claim types in its Tokio Marine deployment, and Verisk's Xactimate remains the property-estimate standard. The Sedgwick report attributes up to 54% efficiency gains to AI photo analysis. An estimate, though, answers "how much" - not "is this claim legitimate."

Payments and routing

Payments and routing are increasingly touchless. Once a low-complexity claim clears triage and estimating, the payout and case routing complete without manual handling. Across these four stages, the through-line is the same: automation has scaled where the task is high-volume and rules-friendly. None of these stages, by design, investigate whether a flagged claim is fraudulent.

Fraud detection has its own mature automation layer, and it is easy to mistake a flag for a resolution. Detection scores or flags suspicious claims after FNOL, and the vendors here are well established. This layer scales. The question it leaves open is what happens to the flag.

Shift Technology's deployment at AXA Switzerland screened over 1M claims and helped stop over EUR12M in fraud, which shows detection works at scale. FRISS and Verisk ClaimSearch operate in the same layer with cross-carrier data and scoring. What none of them do is resolve the flag. They produce a signal that something is worth a closer look, and then a human SIU team takes it from there.

With roughly 10% of property-casualty losses involving fraud and an estimated $308 billion lost annually in the US, per the Coalition Against Insurance Fraud, the volume of flags a mature detection layer produces is large. Every one of those flags is an investigation that someone has to run, or quietly drop. That is where the automated pipeline ends and the manual one begins.

The one stage automation has not reached at scale is the deep investigation of flagged claims, and the industry's own data admits it. The Sedgwick report is explicit that AI works most effectively on "high-volume, repetitive tasks - data extraction, documentation, low-severity claims handling and intake validation," and that 75% of claims professionals believe AI requires human oversight. Investigation is the opposite of high-volume and repetitive. It is low-volume, high-judgment, and case-specific, which is exactly why only 4% of carriers have scaled genAI in claims and only 7% have achieved scalable AI success.

For an SIU director, the consequence shows up as a coverage gap, not a productivity statistic. A manual SIU investigation takes 14+ days per case, and a single investigator carries 200+ open cases. The arithmetic does not close: with that caseload and that cycle time, manual SIU teams across US P&C carriers investigate roughly 25% of flagged claims. The other 75% are paid without full work, denied without full work, or queued indefinitely. The detection layer keeps producing flags; the investigation layer keeps absorbing only a quarter of them.

This is not a tooling failure inside SIU - it is a structural one. We covered the mechanics of it in detail in why flagged insurance claims never get investigated. The short version: the bottleneck is human attention, and no amount of front-end automation relieves it, because front-end automation produces more clean throughput, not more investigated flags.

Investigation stayed manual because it is not a single repetitive task that an automation script can wrap. FNOL, triage, and estimating each reduce to a well-defined operation a rules engine or a vision model can own. A fraud investigation is 15+ distinct judgment-heavy phases per case - document forensics, OSINT, statement cross-referencing, timeline reconstruction, financial pattern analysis, network checks - each producing evidence the next phase depends on.

Manual SIU teams run those phases one case at a time because a human investigator's attention is the bottleneck. You cannot run document forensics and OSINT and timeline reconstruction simultaneously when there is one person doing all of them. That serialization is the reason a case takes 14+ days and the reason coverage caps near 25%. The constraint was never that the work could not be defined - it was that one human could only do one phase at a time.

Autonomous investigation agents change the mechanism. They run all 15+ phases in parallel on every flagged claim, because an agent's per-case attention is not a shared resource. We break down how this works in parallel processing of SIU investigation phases. The result is order-of-magnitude compression: 14+ days drops to 2-4 hours, coverage moves from roughly 25% toward 100% of flagged claims, and per-case cost falls from about $2,500 to about $150. The investigator's role shifts from execution to decision-making - reviewing, overriding, and signing the audit-ready output rather than assembling it by hand.

Put every layer on one map and the gap becomes impossible to miss. Each layer has a job, a set of named automation vendors that occupy it in 2026, and a maturity level. Across FNOL, triage, estimating, payments, and detection, there is a named software vendor at every layer. At the investigation layer, the only incumbent is the manual SIU team.

This is the point a detection vendor's byline cannot make. FRISS, Shift, and Verisk sit in the detection row, and their layer is mature - that is a fact in their favor, not a criticism. Tractable and Xactimate sit in estimating. Guidewire, Duck Creek, and Snapsheet own claims management and FNOL. Every one of those layers is occupied by software. The investigation layer - taking a flagged claim and resolving it end-to-end with a documented, audit-ready report - has no named software incumbent. Its incumbent is the SIU team at every carrier, working one case at a time. That is the layer Hesper is purpose-built for, and it is the only layer on the map where the answer to "which vendor automates this" was, until recently, "nobody."

If you have automated FNOL through payments and bought a detection platform, your next marginal loss-cost lever is the investigation layer, not more front-end automation. The front of the pipeline is where the easy efficiency lived, and the data shows it is largely booked: intake is down from 10 days to 36 hours, low-severity claims process 80% faster, photo analysis lifts handling efficiency by up to 54%. Squeezing those further yields diminishing returns. The un-automated stage is where the next basis points of loss ratio are sitting.

The coverage math is the lever. Only 27% of insurers are pursuing comprehensive claims transformation, and most of that 27% are still leaving investigation out of scope. With roughly 10% of P&C losses involving fraud and $308 billion lost annually, the gap between investigating 25% of flagged claims and investigating 100% of them is the single largest un-pulled lever in the fraud stack. No prior generation of fraud-tech moved it, because every prior generation operated upstream of investigation.

The unit economics make full coverage affordable rather than aspirational. At about $2,500 per manual investigation, investigating every flagged claim was never financially possible - which is exactly why coverage settled at 25%. At about $150 per AI-investigated case, full coverage stops being a budget problem and becomes a workflow decision. When you are ready to evaluate the vendors at each layer side by side, our AI fraud platforms compared breakdown maps detection and investigation tools against each other, and our top fraud detection platforms guide covers the detection layer in depth.