The sharpest thing written about claims automation limitations this year never mentions automation. In September a claims trainer argued that carriers are producing processors instead of adjusters, and defined the difference as a single act: a processor files the police report, an adjuster reads it and notices the witness account contradicts the insured's statement. He was making a training argument. He had also just described fraud detection.
Comparing two records and noticing they disagree is the step no claims tool owns. Every point tool answers its own narrow question correctly, and the file still closes wrong, because nothing checked output one against output three. So the limitation that costs money is not model accuracy. It is the unowned space between the models.
Three sources measure that space from different directions: a public dataset on fraud referral acceptance, a state regulator's AI bulletin, and the training editorial that never mentions fraud at all. For the stage-by-stage view of what automates today, read this alongside our claims automation pillar.
What claims automation actually does well
Answer
What is claims automation and which parts of a claim does it handle well?
Claims automation is software that completes a defined claims task without a person: intake extraction, document classification, triage scoring, estimate review, payment checks. It works. Inside a single stage with a clear right answer, a well-built model beats a queue of people doing the same task by hand, and carriers have deployed it accordingly.
Adoption is not in doubt. In the NAIC's Big Data and Artificial Intelligence Working Group surveys, 88% of 193 responding private passenger auto insurers said they use, plan to use or plan to explore AI/ML models (December 2022), and 70% of 194 homeowners insurers said the same (August 2023).
The NAIC also records how carriers use those models. Its private passenger auto survey team report puts claims at the top of the operational areas, at 70%, and records the models as deployed "mostly as an informational resource for adjusters" by 96 companies. That scope is the tell: the model informs, and a person still assembles the file and makes the call.
Academic work scopes itself the same way. Claim Automation using Large Language Model (Mo, Quan, O'Donohue and Zhong, February 2026) builds an initial decision module inside the claim pipeline. Non-fine-tuned models reached 56.0-64.4% accuracy on valid outputs; the fine-tuned model reached 81.5%, and 92.0% on a high-quality subset.
Front-door vendors are good at their stage too. Liberate, which sells voice, SMS and email intake agents that run inside core claims systems, reported more than $100 billion in premiums processed in March 2026. None of what follows is an argument that stage automation does not work.
Claims automation limitations are between the tools, not inside them
Answer
What is the biggest limitation of claims automation?
Every published list of claims automation limitations is a list of per-tool limitations: data quality, legacy integration, ambiguous coverage, change management. The expensive failure sits one level up. Each tool returns a correct answer on its own sub-task, and nothing compares output one against output three, so the file closes clean and wrong.
The per-tool version of that list is set out in practitioner terms by Travis Sommerfeld, SVP in J.S. Held's building consulting practice, writing in Claims Journal on May 8, 2026. His argument is that technology is most valuable on "structured, repeatable and process-driven tasks" and that problems start when it is "positioned not as an enabler of expertise, but as a substitute for it". He is not arguing against automation.
Read that list as a set rather than six warnings. Every entry is a tool being asked a narrow question and answering it correctly inside its own frame. Not one is a model defect, and no amount of per-tool accuracy closes the gap, because the gap is a comparison that no tool was asked to make.
Point automation hands off. An investigation compares.
Each stage tool answers its own question and passes the file on. The comparison step is not any tool's job.
- Intake extractionAnswerFields pulled from the notice and its attachments, then handed to the next tool
- Estimate reviewAnswerEvery line item inside tolerance, then handed to the next tool
- Weather checkAnswerHail reported in the area on the date of loss, then handed on
- The comparisonUnownedNothing compared answer one against answer three, so the file closes clean
- 01One claim recordEvery document extracted into one sourced record at intake
- 02Phases in parallel15+ investigation phases run on the same file at the same time
- 03Outputs comparedEach phase output checked against the others, not just filed
- Coverage decision
- Reserve
- Settlement
Point tools each return an answer. One investigation returns a comparison.
One hail claim, five correct answers, one wrong file
Answer
How can automated claims tools all be right and the claim still be wrong?
Here is the failure mode in one file. A commercial hail claim runs through five automated stations: imagery classification, weather verification, rules-based estimate review, an AI repair estimate and an automated code report. All five return correct answers. The file is still wrong, because no station compared its answer against the others.
Walk it through. The imagery classifier marks a surface anomaly as storm-consistent, which is correct inside its classifier and, on Sommerfeld's reading, does not separate storm damage from an installation defect. The weather service confirms hail in the area on the date of loss, correct as a modeled output and not a forensic conclusion.
Five tools were right. The file was still wrong.
Each station answers its own sub-task correctly, which is why no per-tool accuracy score catches this file.
The imagery classification against the weather date, and the estimate-review pass against the recorded statement in which the insured described the repair as already finished and paid for.
Five green lights. The file is clean at every station and wrong as a whole.
Nothing in that chain compared the classification against the weather date, or the estimate-review pass against the recorded statement. Five green lights, one wrong file, and no tool failed. The comparison was never anybody's job, which is why it is missing from every per-tool limitation list.
The processor-adjuster gap, in an adjuster's own words
Answer
What can a claims adjuster do that claims automation cannot?
Compare two records and notice they disagree. Chris Casaleggio, Northeast regional director at MC Consultants, drew the line in Claims Journal on September 3, 2026: a processor files the police report, an adjuster reads it and notices the witness account does not match the insured's statement. That is the scarce skill.
His article is about training. It never mentions fraud, SIU or any vendor, which is exactly what makes the sentence below worth borrowing: a description of what an adjuster is for, written by someone with no software to sell.
That one act is fraud detection, leakage control and coverage defence at once. His own anecdote shows the cost of skipping it: during an early recorded statement the insured mentioned travelling to the airport, he ran his checklist instead of following up, and later found "the insured was operating a livery business while the vehicle was insured under a personal auto policy".
His section header is "Technology Should Provide More Time to Adjust", and he attributes the 88% and 70% adoption figures to NAIC surveys. He is asking for better training, not for less automation. Our argument sits next to his rather than against it.
One thing to add. Contradiction-spotting is treated as a human skill built over years, and it is. It is also the only task on this page that parallelises cleanly, because it is nothing but many sources checked against each other. Hesper AI runs 15+ investigation phases on one claim at once and compares their outputs, so a workup lands in minutes, not weeks, against the 14+ days a manual investigation takes.
The contradiction was never hidden. It sat in two documents nobody put side by side.
What the referral data shows about machine-made volume
Answer
Do automated fraud detection tools produce better referrals than adjusters?
They produce more of them. In the Coalition Against Insurance Fraud and Aon 2024 SIU benchmarking study of 35 P&C carriers, automated tools generated 45% of referrals at a 36% acceptance ratio, while adjusters generated 49% of referrals at 70%. Automation supplies the volume; the measured hit rate stays with people.
One public dataset measures that handoff directly. The 2024 Insurer SIU Benchmarking Study, the sixth since 2012, found automated tools generating 45% of referrals, up six points since the 2022 edition, with the adjuster share down to 49%. The acceptance ratios moved the other way: automated referrals up to 36% from 32%, adjuster referrals down to 70% from 76%.
The arithmetic is worth doing once. Per 1,000 referrals reaching an SIU, machines send 450 and 162 are accepted, while adjusters send 490 and 343 are accepted. The study's second figure reconciles with that: it reports accepted referrals by source as adjusters 60%, automated tools 29% and other sources 10%, and 29% of the implied 572 accepted referrals is 166, within rounding of 162.
Not accepted is not the same as wrong
A referral can be declined for scope, jurisdiction or workload rather than because the signal was false. The acceptance ratio measures what an SIU took on, not what was true. An automated referral being accepted 36% of the time is a statement about the evidence that arrived with it, not a verdict on the model that sent it.
The capacity side did not move to meet the volume. SIU staffing grew 0.5% in 2024 against 1.4% in the previous study, coverage fell to 0.67 full-time equivalents per $100 million of gross written premium from 0.9, accepted referrals per investigator rose to 174 from 162, and budgets slipped to about 0.12% of premium from 0.13%.
Machines send 45% of referrals and 29% of the accepted ones
Acceptance ratio is the share of a source's referrals that an SIU takes on for investigation.
More referral volume, a lower hit rate, and the same people to absorb it.
A scoring model finds an anomaly. It does not assemble the evidence that justifies opening a case, which is the gap behind the 60-85% false-positive rates reported for rules-based detection. The machine sends a signal; the adjuster sends a signal plus a reason, and the reason is what an SIU accepts.
The bench expected to absorb it is shrinking
Every automation business case assumes a person is free to handle the exceptions. The US Bureau of Labor Statistics projects employment of claims adjusters, appraisers, examiners and investigators to decline 6% between 2025 and 2035, from 389,700 jobs, a net loss of 21,800.
Experience leaves faster than headcount. Sedgwick's 2026 Loss Adjusting Insights Report, reported by Claims Journal on October 2, 2026, expects a quarter of claims adjusters to retire by the end of 2027. The report carries a talent survey by The Institutes in which 73% named loss of industry knowledge as the greatest impact of that retirement wave, and 53% named recruiting a new generation.
Where straight-through processing stops
Answer
Why does straight-through processing stop at a certain share of claims?
Straight-through processing is a routing decision, not a judgment. It removes the files that need no judgment, so the residue it leaves behind is harder than the average claim it took away. A higher STP rate does not make the remaining files easier to investigate; it makes them denser.
The ceiling is visible in the NAIC data itself. Carriers describe AI/ML claims models as an informational resource for adjusters, and in our read of those same surveys, not one of the 193 large auto insurers surveyed reported a model that denies a claim. Machines may pay; only people deny.
That split is the right one, and it has a consequence carriers rarely price in. If automation resolves only the clean files, the investigation work per remaining file goes up, which is exactly the shape of the referral data: more flagged volume, flatter capacity, a lower acceptance ratio.
Treating those as two queues is what produces the gap. Clean claims resolve straight through; suspicious claims get an investigation-grade workup, out of the same engine and into the same evidence file. Manual investigation reaches about 25% of flagged claims at 200+ cases per investigator, and running the phases in parallel is how that becomes 100%, rather than a prettier score.
What regulators already expect automation to prove
Answer
What do insurance regulators require when AI is used in claims decisions?
A state regulator has already written it down. South Carolina's Department of Insurance Bulletin 2026-07, issued September 24, 2026 on the NAIC model bulletin language, makes insurers responsible for "any automated or AI-related conclusions, decisions or determinations" they rely on, whether the system was built in-house or by a vendor.
The bulletin states that insurers "are responsible for decisions made and actions taken by Insurers, their producers, adjusters and other authorized representatives based on information derived from or generated by the AI Systems". It applies whether the system was developed by the insurer or a third-party vendor.
It also names the decay. Model Drift is defined as "the decay of a model's performance over time arising from underlying changes such as the definitions, distributions, and/or statistical properties between the data used to train the model and the data on which it is deployed".
Controls are then scaled to, among other things, "the extent to which humans are involved in the final decision-making process" and "the extent and scope of the insurer's use or reliance on data, Predictive Models, and AI Systems from third parties".
Section 1.7 expects the program to cover an AI system's whole life cycle: design, development, validation, implementation, use, ongoing monitoring, updating and retirement. The bulletin then ties all of it to the state's unfair claims practices standards "for the investigation and disposition of claims".
This is where single-stage automation becomes difficult to defend, and model quality has nothing to do with it. If four tools each contributed a finding and no artefact records how they were reconciled, there is no reconstructable chain to hand a market conduct examiner. Evidence behind every decision is an audit requirement before it is a product claim.
Which means the decision rights get written down rather than inherited from any vendor's brochure, ours included. Ours are in human-in-the-loop AI investigation: the agents do the legwork across every phase and log each action with its source, reasoning and timestamp, and the licensed adjuster keeps the coverage decision, the reserve and the settlement. Adjusters review evidence-backed files instead of building them.
Key takeaways
- Claims automation limitations now sit between tools rather than inside them: each point tool answers its own sub-task correctly and nothing compares one output against another.
- The Coalition Against Insurance Fraud and Aon 2024 SIU benchmarking study found automated tools generating 45% of referrals at a 36% acceptance ratio against 49% of referrals at 70% for adjusters, while SIU staffing grew 0.5%.
- The act that separates an adjuster from a processor is noticing that a police report does not match the insured's statement, and that act is contradiction-spotting across sources, which parallelises better than any other task on the list.
- The people expected to catch what automation misses are leaving: claims adjuster employment is projected to fall 6% between 2025 and 2035 per the Bureau of Labor Statistics, and a quarter of adjusters are expected to retire by the end of 2027.
- South Carolina's 2026 AI bulletin makes insurers responsible for the automated conclusions they rely on, including a vendor's, so an unreconciled set of tool outputs is a documentation problem as well as a leakage problem.