Most claim management automation solutions are sold on throughput. And throughput is the easiest number in the business to move without improving anything. Here's what that looks like at scale.
In August 2026, KFF published the first year of prior authorization metrics that insurers are now required to report to the Centers for Medicare and Medicaid Services. Medicare Advantage insurers denied roughly 12% of standard prior authorization requests in 2025. Of the denials that were appealed, 67% were overturned. And among individual insurers, the overturn rate ran as high as 93%. KFF also notes that denials are rarely appealed at all.
Read those two facts together. A decision process reverses itself two thirds of the time when anyone checks. And almost nobody checks. Every one of those denials was a completed transaction. Every one counted as throughput.
Here's why that should worry a property and casualty carrier that has nothing to do with Medicare Advantage. Those numbers aren't public because those insurers are careless. They're public because a federal reporting mandate forced a reversal rate into the open. And the moment it appeared, it contradicted the throughput story the same systems had been telling internally for years.
Every automated claims operation has that number sitting inside it. And most have never calculated it. Which means most are running on a metric that can't fail. Keep in mind that none of this is an argument against automating claims. A high straight-through rate is worth having. And carriers that reach one properly take real cost out of the operation.
My argument is against treating the rate as evidence on its own. What follows is how that number gets moved without anything improving, and the five measures that tell you whether yours is real. Plus I’ll discuss how to raise it segment by segment, so the gain survives contact with a loss ratio.
How do you raise a straight-through rate without improving anything?
By changing what enters the numerator or the denominator. Neither of which requires better decisions. The threshold move is the most common and least visible. Automated completion sits behind a confidence cutoff. Lower the cutoff and more claims finish without a human immediately.
These additional claims are, by definition, the ones the system was least sure about. So error rises at the same time. It surfaces in the loss ratio 6-9 months later. Where it gets attributed to pricing or weather rather than to a routing change nobody logged as material.
The mix move requires no decision at all. Automation succeeds first on the simplest losses, which is correct sequencing. But a quarter with a favorable mix produces a higher blended rate with no change in capability. Unless the rate is split by claim type and complexity, a carrier can't tell a better system from a milder storm season.
The definitional move shows up in vendor reporting. What counts as straight through? A claim a human glanced at and approved in 8 seconds? A claim that was completed automatically and reopened three weeks later? Carriers rarely pin this down in a contract, and the definition drifts toward whatever makes the number look best.
What happens when nobody measures the reversal?
The Senate Permanent Subcommittee on Investigations answered that question in October 2024, after reviewing more than 280,000 pages of internal documents from the three largest Medicare Advantage insurers. Its report found that UnitedHealthcare's denial rate for post-acute care rose from 8.7% in 2019 to 22.7% in 2022. And that its skilled nursing facility denial rate increased ninefold over the same period.
In 2022 Humana's post-acute denial rate was more than 16 times its overall denial rate, while UnitedHealthcare and CVS denied post-acute requests at roughly three times their own overall rates. Together those three insurers cover close to 60% of Medicare Advantage enrollees. The subcommittee tied the increases to the adoption of automated review. And the broader investigation has continued into 2026.
Every one of those denials completed. Throughput went up. Cycle time went down. On any dashboard reporting the straight-through rate, that period looks like a success. And it took a Senate investigation and a federal reporting mandate to establish otherwise.
Yes, I know that health insurance isn't property and casualty. And the incentives differ. But the measurement failure transfers exactly. An automated decision that nobody revisits looks identical to a correct one.
Which numbers can't be gamed the same way?
There are 5 metrics that can’t be manipulated, and their value is that they move in the wrong direction when you pull the levers above. So let’s uncover what they are:
- Decision reversal rate is the one the Medicare Advantage data makes unavoidable. How often does an automated decision get changed later by a human through complaint, escalation or internal review? If the answer is 67%, the system isn't deciding. It's deferring.
- Reopen rate inside 90 days is the natural counterweight to throughput. Loosen thresholds to push volume through and reopens climb directly.
- Sampled leakage is third. Pull a stratified sample of automatically completed claims, have experienced adjusters review them against what the outcome should have been, and keep those adjusters independent of the team tuning the system. Grade your own homework and the grade drifts.
- Cycle time at the 90th percentile is next. The mean improves almost mechanically as automation absorbs simple claims, which makes it useless as a health measure. The tail is where customer harm concentrates.
- Lastly, complaint rate split by automated and manual path. If the automated path generates complaints at a materially higher rate, the carrier has a conduct problem regardless of the efficiency numbers. Carriers with audit ready reporting built into the workflow produce all five from one set of instrumentation rather than five reporting exercises.
What does a regulator expect a carrier to be able to show?
A written program with a name attached, not a configuration screen owned by operations.
The National Association of Insurance Commissioners adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers in December 2023. It sets expectations that insurers govern AI through a documented program with senior accountability, risk controls, testing and oversight of third party systems. And it tells insurers what a department may request during an investigation or examination.
The bulletin is guidance rather than model law, so it applies in a state only once that state's department adopts it. And thus far, adoption varies. The NAIC also maintains an implementation map of adopting jurisdictions alongside an overview of its artificial intelligence work. Check both against your actual footprint rather than an assumed one.
There are three elements that matter most for claims:
- Accountability needs a person for each automated decision type, not a committee.
- Testing has to cover outcomes rather than only performance, which means checking whether comparable claims produce different results across customer segments.
- And third party oversight extends to any model or data source you didn't build, because the expectation is that the carrier answers for behavior it didn't author.
How should a carrier expand without betting the loss ratio?
By segment, with an evidence gate at each step, and never by nudging a global threshold. Claim management automation solutions fail this test more often than they fail a technical one.
Start where the loss type is simple, the data is structured, and the value is bounded. Run it in shadow first with the system producing a decision and a human making it. Make sure to record every divergence.
Then go live on that segment with a defined review sample and a reopen threshold that triggers rollback. Hold it for a full claims cycle. Which means a quarter for most property lines and considerably longer for liability. Expand to the next segment only after that, and expand by segment rather than by threshold. That’s because a threshold change lands everywhere at once and can't be attributed when something moves.
Write down why you expanded each time, with the numbers that justified it. That record turns a run of operational choices into a governed program. And it's the first thing an examiner asks for. It's also what makes the return legible to a finance function that will otherwise book the savings and never see the reversals.
Carriers evaluating AI agents and workflow automation should treat the evidence gate as a platform capability rather than a project deliverable. Which is generally where a turnkey approach proves itself or doesn't. I know that was a lot to digest, especially for someone whose day job wasn’t to become an AI data scientist. If you’d like help with any of this investigative work, or are even interested in hearing about running a pilot with our managed platform, let’s chat.
What else do claims leaders ask about AI claims automation?
What straight-through rate should a carrier target?
No universal target is defensible, because the achievable rate depends almost entirely on line of business and claim mix. Aim instead for a rate that holds or rises while reversal rate, reopen rate, sampled leakage and complaint rate on the automated path stay flat or fall.
How do you measure reversal rate when almost nobody appeals?
Sample proactively rather than waiting for challenges. The Medicare Advantage figures show why. An appeal rate near 12% means the observed reversal rate describes a small, self selected slice, and the true error rate across the population that never appealed is unknown.
How large should the leakage review sample be?
Large enough to detect a material shift inside one reporting cycle, stratified by claim type and value band rather than drawn at random. Random sampling piles up in high volume, low value claims and misses the segments where leakage costs most.
Who should own automated claims decisions?
A named claims executive rather than the technology or data function, with the model owner answerable for performance and the claims owner answerable for outcomes. In adopting states the regulatory expectation places responsibility with the insurer whether the model was built in house or bought.
Explore more blog posts from Unframe.
