Denial appeal automation delivers speed, but not change.
With the proliferation of LLMs, denial appeal automation has become part of every revenue cycle vendor’s pitch. Feed the model the denial, the clinical record, the payer’s own policy language, and it writes an appeal in seconds. The output is good, maybe better than what an overloaded analyst would produce.
But LLM-generated appeal letters are close to the least consequential place a health system can add AI. Payers adjudicate with an algorithm. You’re now answering with an algorithm. Two models trading documents about a rule neither of them wrote. The letters get faster and nothing upstream moves.
The speed difference is the actual problem. A payer can change a payment policy in a bulletin and have it live the following month. Your side of that exchange is a contract amendment cycle, an IT queue, and a CDI training calendar. Most of these changes don’t even arrive looking financial. They land in a provider manual update or a network bulletin, and nobody’s revenue cycle team is reading a 200-page manual refresh hunting for the paragraph that reprices a third of your inpatient volume. Appeal automation makes you faster inside the lane the payer built. It does nothing about who gets to build the lane.
What denial appeal automation changes
Automated appeal letters reduce your cost per appeal. Administrative cost per denied claim was $57.23 in 2023, up from $43.84 the year before. Across a few thousand accounts, that saving is meaningful. It’s also a floor rather than a representative number. A DRG downgrade that moves through clinical validation review, a second-level appeal, and then contractual dispute resolution costs a multiple of it, and those are the categories where the dollars actually sit.
Automation doesn’t touch prevention. Faster appeals don’t change whether the same denials happen again next month. An appeal works one claim at a time. Win at 60%, win at 20%, the edit that produced the denial is still live on the next claim. Your cost of playing the back-and-forth game went down (but so did the payers).
And, there’s the population an appeal can’t reach at all…
The fastest-growing adverse payment outcomes never generate a denial
Aetna’s level-of-severity policy is the prime example. An urgent MA stay of one to four midnights gets paid at an observation-level rate unless it clears Aetna’s internal severity criteria. That isn’t a denial. It’s a paid claim at a lower tier, and the remit returns a generic adjustment code with no transaction-specific explanation of why this particular stay failed the criteria. The code is technically present, but operationally useless. Your workqueue has nothing to route, and your analyst has nothing to answer.
The AHA made this exact point in its letter to Aetna. Without an official denial, hospitals may not know an underpayment happened, and the dispute leaves the appeals process for whatever resolution mechanism the contract specifies. Usually arbitration, which the AHA called more costly and burdensome than appeals, with outcomes that stay private.
Post-payment takebacks behave the same way, and they hand you the reconciliation bill. Money leaves through a provider-level adjustment weeks or months after payment, detached from the claims it came from. Reassembling which accounts a takeback actually hit is its own project, so root cause rarely gets established, so the same recoupment happens again next quarter. The burden of proof and the burden of accounting both sit on your side.
Denial appeal automation has nothing to reach in either population.
The loss you don’t see for three years
Every downgrade you accept is also a data point. Downgrades pull your case mix index down, and CMI is an input to the next rate negotiation. If you absorb enough of them without contesting the pattern, you build the payer’s argument for it. Your documented acuity looks lower, so your base rate gets set lower, and the loss stops being a claim problem and becomes a contract term.
That’s the part appeal volume can never recover. You can win the individual accounts and still lose the rate they get priced against.
Beyond denial appeal automation, places where payment outcomes actually change
Contract language. Whatever your agreement doesn’t say is where unilateral policy changes live. Requiring CMS criteria for MA inpatient status determinations. Defining what counts as a reviewable denial. Prohibiting mid-term payment policy changes. And payers run their operations against regulatory timelines the contract itself never names, so your team gets held to deadlines that don’t appear in the document they’re working from. (Put those in writing too.)
The joint operating committee. The only standing forum where you put evidence in front of the payer and get an answer, on the record.
Dispute resolution and arbitration. Where severity-review underpayments are designed to land. It runs on patterns, not claim narratives or appeal letters.
Regulatory and public escalation. CMS complaints when an MA plan is more restrictive than traditional Medicare, state DOI filings, and now litigation. Jefferson Health sued Aetna in April over the severity policy, citing the two-midnight rule and breach of contract.
These are bigger undertakings than appeal letters, and the outcomes are durable. All four need the same input from you. Claim-level, payer-level evidence that a payer treats you differently than its peers do, and differently than your contract defines. The pattern, with a dollar figure attached.
The evidence, or Payments Intelligence, that you need
Same-DRG comparison across payers. Your denial and downgrade rate on a given DRG at one payer, set against the same DRG across the rest of your book. Differential treatment on identical clinical facts is the most useful exhibit you can build, and the hardest one for a payer to talk around.
Overturn rate paired with days to resolution, by payer and denial type. A payer that overturns 70% of a denial type after 94 days isn’t making clinical determinations. It’s financing itself with your cash. That’s a contract conversation, not an appeal.
Paid claims below contracted rate, grouped by adjustment code and by the policy behind it. Generic codes across thousands of claims become a specific pattern once you sort them by payer and service line. This is the population appeals can’t reach, and most systems can’t see.
Rolling 12-month PLB activity by payer, traced back to the originating accounts. Recoupment dollars, reason codes, and whether each takeback ever tied to something reviewable.
Case mix index movement set against downgrade volume by payer. If your CMI is drifting while your documentation scores hold, that’s a payer behavior finding, and it belongs in your next rate conversation, not your denial dashboard.
Dollars and days, not counts. Claim volume is one thing; “$4.1M in delayed payment over 11 months, all traceable to one policy” can actually get a response.
All of it lives in your own 835 and 837 history. Sift’s Denials Insights Report assembles it into payer-specific, claim-level proof, and it’s the same data RevProtect uses to flag underpayments and recoupments that never surface as denials.
Denial appeal automation is a cost play, but it’s not necessarily leverage. You build leverage out of your own remit data, and it’s how you change what the next thousand claims look like, along with the rate they get paid at.
If you want to see your organization’s payer-specific patterns, Sift can run a deep dive for you.
