A large commercial payer can produce a clinically worded, guideline-cited denial on a claim no human ever opened. It costs them almost nothing, and they can do it again on the next claim and the one after that.
On the provider side of that exchange, a nurse auditor pulls the chart, reads the documentation, figures out what the payer is really arguing, and writes something defensible. Forty minutes if the account is clean. Three hundred more are behind it.
That asymmetry is the actual problem in denials right now. I run product for a company built on this (@sifthealthcare), so I’m in it every day, and I’ve come to think most of the industry conversation is aimed at the wrong target.
Health systems already see their denials. That’s not the gap.
Every revenue cycle leader knows their administrative (CARC/RARC) initial denial rate, top five payers, and worst reason codes. Every health system has this visibility, with dashboards that show trends by month and sliced by payer. And most revenue cycle leaders are now tracking clinical denials and takebacks at deeper levels.
What hasn’t been solved is the actionability. The gap is between surfacing a denial and the specific action that prevents it.
A payer downgrades your sepsis DRG. You know it happened, and you can probably guess why. But that is a long way from knowing which secondary diagnosis they’ll challenge on this account, what in the chart actually supports it, what a coder should verify pre-bill, and whether the diagnosis added from a query your CDIS writes will hold up. It’s also almost entirely manual, which makes it a staffing problem, which is exactly why the payers win. They generate denials faster than you can work them (payers are years ahead in deploying AI).
Where and how AI helps, and doesn’t.
The easy logic is, if the payer automated their side, automate yours. Half the RCM vendors in the market are selling that story. We went down that road early, and it’s not that simple.
If you hand a general-purpose model a real case, the output is pretty good, until it isn’t. A model will write you a fluent, confident recommendation on that sepsis account. But, buried in the middle of it, it will also invent an ICD-10 code that doesn’t exist, cite a length-of-stay benchmark it made up instead of the one CMS publishes, and draft a non-compliant physician query. Not every time, but often enough that no one could put it in front of a CDI team without a human checking every line, meaning you’re not saving any time/touches.
A DRG’s relative weight is not something you want a model to guess at. It has exactly one right answer, that’s in a reference table, republished every year. Same with the LoS benchmarks and compliant-query standards. Those are facts you look up, and that your team knows, not judgment calls you improvise.
At Sift, we balance making our models “smart” with making them accountable.
The AI models that actually predict revenue risk and prevent adverse payment outcomes work like very disciplined analysts who never guess.
When it needs a code, it retrieves the real one. When it makes a recommendation on your downgraded sepsis claim, that recommendation reflects how that specific payer adjudicates that specific DRG, built from what we’ve watched happen across hundreds of them. That’s why Sift maintains 329 MS-DRG playbooks covering 772 DRGs, each built and kept current by revenue cycle and clinical SMEs. It’s unglamorous work, and it doesn’t demo well. It’s also the only way a recommendation gets specific enough to be worth workflow alerts and action.
Our AI model still does real, innovative work. It reads the chart, summarizes the documentation, drafts the language, and all very quickly. It just never gets to invent a fact that has a real source one lookup away.
Our goal is to deliver Payments Intelligence that revenue cycle teams can really use. A nurse auditor on account four hundred at four in the afternoon extends trust exactly once. If the tool is confidently wrong once, she stops opening it, and you’re back to appealing by hand.
Defensibility + Trust
You are not going to out-hire the payers. They removed people from the cost of denying you, and no appeals budget competes with that.
What changes the math is taking out the manual bottleneck without lowering the bar on what’s defensible, which is hard to build. When a vendor tells you they use AI on your denials, the question isn’t how advanced the model is, but how defensible and trustworthy its output is. You should understand where a specific recommendation came from, down to the source and its version.
The payers did the slow, unglamorous work of industrializing denials years ago. Beating that back takes the same kind of work on defense. Not a smarter guess, just a faster version of being right.
