Article Summary: Steve Palma, President and General Manager at Penstock, explains how AI is making provider appeals faster and easier to produce, and what that means for the post-pay findings health plans count on. Plans that continue using last year’s concept library risk more overturned findings and a wider gap between identified and collected dollars. He offers three questions to help determine whether a program’s findings will hold up.

‍

AI-assisted provider claim submissions are changing your historical claims payment trend.

‍

Somewhere in your appeals queue right now is a letter nobody wrote.

It's well organized. It cites your own policy back to you. It arrived faster than a person could have produced it, and it looks a great deal like three others that came in the same week. Plans around the country are seeing them, and the fingerprints of a language model aren't subtle.

Hold onto that, because it changes what every finding your program produces is actually worth. I'll come back to it.

‍

How AI Is Changing the Way Providers Appeal Payment Integrity Findings

First, the conversation everyone is having instead. Health plans say AI is pushing costs higher. Hospitals say payers have automated denials at a scale no provider can match. By late September, the headlines were calling it a bot war, and the chief executive of one large health system said his nightmare was a data center occupied entirely by his bots arguing with an insurer's bots.  

The Blue Cross Blue Shield Association put a number on it: close to a billion dollars in excess costs across its system over two years, with the share of inpatient stays billed as complex climbing from roughly 37% to 40%.

‍

Why AI Coding Follows the Same Path as Electronic Health Records

We've seen a version of this before. When electronic health records arrived, hospitals used them to document more completely, and billing followed. David Brailer, who ran the federal push to adopt them, made that comparison recently. He doesn't think there's anything about AI that escapes the same market forces, and I don't either.

The dollars are real and plans are the ones absorbing them. But this isn't a fight you win once and close out. Fee-for-service pays for more complete documentation, the tools that produce it are widely available now.

‍

Why Rising Coding Severity Doesn't Surface Before Payment  

The research is worth a closer look, though not for the dollar figure. Researchers compared what hospitals coded against what was actually delivered, across years of paid claims. The coding moved. The care didn't move with it. Hospitals dispute that reading, and their objection is a fair one: a diagnosis doesn't require a particular treatment to be valid. That argument will run for years and I don't expect to settle it here.

What isn't in dispute is when anyone noticed. Nobody caught this at the point of payment. Every one of those claims was plausible, documented, and adjudicated correctly against the rules in place. There's no adjudication failure in that story. It took looking back across a whole population of paid claims to see anything at all.

‍

Why Auditing More Claims Doesn't Always Improve Payment Integrity Results

The reflex is to conclude that plans should audit more. I'd be suspicious of anyone in my seat who said so. The constraint isn't how much you audit. It's what you're looking for. Most post-pay payment integrity programs run the same concept library they ran last year. Apply the rules, report the findings, repeat next month. Push harder on that and you don't surface new problems. You surface the same ones again, and a program can look busy for years that way without getting any better.

‍

How AI Is Lowering the Cost of Provider Claim Appeals

Which brings me back to the letter nobody wrote.

Recycled findings are the ones providers have seen before, and they know exactly how to argue with them. That was always true. What’s changed is how much work it takes to argue. Appeal rights are bounded, by contract and by the levels available before a case goes to independent review. The effort it takes to use them isn’t bounded anymore. A provider who would have let a marginal finding go two years ago can now push it through every level the contract allows. The letter takes minutes to produce. It’s competent. It cites your policy, attaches the record, and arrives inside the window.

The cost of that doesn't land where people expect. It lands in your shop. Somebody has to read the appeal, pull the file, re-review the claim, and write back. Then do it again when the next one arrives, because appeal volume is no longer limited by how many people the other side can staff.  

Which is where this stops being an operations problem and starts being a finance one. Gross identified and net collected were never the same number. What's changing is the gap between them, and it widens fastest on the findings that were thinnest to begin with. A recovery nobody can defend doesn't simply fail. It gets discounted in the forecast, and then finance learns to discount the next one too.

‍

What Makes a Payment Integrity Finding Defensible

What holds up hasn't changed: contract language, plan policy, and the record. What's changed is the margin for error. A finding that would have survived on plausibility two years ago now has to survive a bot that never gets tired and costs nothing to run.

AI-assisted coding and documentation are changing what arrives in plan claims data. PwC projects costs up 9% next year, and close to 70% of the plans it surveyed rank providers’ AI documentation and coding tools among their top three inflators. That number is going to be in every planning deck between now and January, and it should be. Next to it, I'd want answers to three questions.

1. What's your overturn rate by concept, not in aggregate? Aggregate hides the problem. You're looking for the concepts that get appealed away nearly every time, because those are costing you more than they return.

2. How far apart are identified and collected, and how long does it take someone to tell you? If that number is hard to produce, you already have your answer.

3. And when did your post-pay concept library last change? If what you’re looking for this year is what you were looking for last year, you’re auditing a billing environment that no longer exists.

Providers can now contest any finding you send, and do it quickly and well. Before you count a finding as money, ask whether it would survive that letter. If it wouldn’t, don’t count on collecting it.

‍