In short. Detection is the part the market has solved. Every alert it produces opens two pieces of work: recovering the claims that already denied, and stopping the next one being created the same way. The first leaves a receipt in the form of dollars recovered. The second shows up as nothing happening, so it keeps getting deferred, and the denial rate stays where it was. Appealing a denied claim costs around $57 in administrative time and succeeds a little over half the time. Preventing it costs a fraction of that, once.
An alert creates work. It does not create capacity.
A detection layer that finds forty issues a month, handed to a team with the hours to work three, has not improved the position. It has described it more precisely.
Finding the leak converts an unknown loss into a known one. The money is identical, and now somebody is accountable for having been told.
Which is why a second dashboard rarely helps, and why the reflex to buy better analytics after the first analytics underdelivered tends to produce the same outcome at a higher cost. The constraint moved downstream the moment detection started working.
Every alert opens two pieces of work
Take a real alert. Fourteen claims, one reason code, one payer, first occurrence three weeks ago. What has to happen next splits in two, and the two halves behave nothing alike.
Recovering the fourteen is visible work with a visible result. Someone appeals, some of the money comes back, and the recovery appears in a report with a number against it.
Stopping the fifteenth produces no such artifact. A claim that goes out correctly and gets paid looks exactly like every other claim that goes out correctly and gets paid. The work succeeded and left no trace that it happened.
Prevention has no receipt
That asymmetry decides how the budget gets spent, and it does so without anyone arguing for it.
Recovery can be measured in the unit finance already uses. Dollars recovered this quarter, against hours spent. It defends itself in a budget meeting. A revenue integrity team that recovered $400,000 has a case for next year.
Prevention is measured in denials that did not occur, which is a counterfactual. Nobody can invoice for it, nobody gets praised for it in a month-end review, and when capacity is short it is the half that gets deferred to next quarter. It is deferred every quarter for the same reason.
Which produces an operation that gets progressively better at recovering from a problem it never stops creating.
What the arithmetic says about that choice
Premier's survey work puts numbers on both sides of this, and the comparison is not close.
Around 15 percent of claims submitted to private payers are initially denied. The average administrative cost of fighting one has risen to roughly $57 per claim, up from about $44 the year before. Across US hospitals that added up to an estimated $19.7 billion spent contesting denials, of which Premier judged a substantial share avoidable.
Then the part worth sitting with. A little over half of contested denials are ultimately paid. Read that from the provider side and it says the appeal was usually right, which is encouraging. Read it from the operating side and it says you paid $57 a claim to establish something that was true before you started.
Winning an appeal means recovering money you would have kept for nothing had the claim gone out right the first time.
The same logic scales down to a five-provider clinic, where the absolute numbers are smaller and the hours are scarcer, so the deferred half is deferred harder.
What an agent does on each side
Both halves are assembly work with a decision at the end, which is the shape agents handle well and the shape that makes governance straightforward.
On the recovery side the agent pulls the clinical documentation the reason code calls for, drafts the appeal against the payer's own published policy, and ranks the fourteen by recoverable value against days remaining in the appeal window. A person reviews and signs each one.
On the prevention side it traces the fourteen back to the step where the claim was assembled, names the element that was missing, and then checks the next claim in that cohort against the same requirement before submission. The rule itself is approved once, in writing, by a named person, and that approval is what makes the check defensible later.
Then it re-measures the cohort and reports whether the leak closed, which is the only way the invisible half ever becomes visible.
How to measure the half you cannot see
Counterfactuals are hard to prove and easy to argue with, so do not try to count denials avoided. Measure the cohort instead.
Fix the cohort definition before anything changes: this payer, this procedure group, this reason code. Record its denial rate, its rework rate and its days to detection for a month while the team works as usual. Apply the rule. Then read the same three numbers for the following month on the same cohort.
A cohort rate that falls while volume holds is the closest thing to proof available in healthcare revenue cycle, and it is defensible at a budget review in a way that an avoided-denial estimate never is.
Four things to do this week
Count the alerts your current reporting produced last month, and count how many produced a change to how a claim gets assembled. The ratio is the finding.
Take the single alert with the largest recoverable value and split it on paper into the two halves. Name who owns each half. In most organizations the second name is missing.
Baseline one cohort properly, before anything is built, because the comparison is worthless afterwards if nobody wrote down the starting point.
Then ask what it would take for a rule to be approved once and applied to every claim in that cohort from then on. If the answer involves a committee, that is the real constraint, and it is a governance conversation rather than an analytics one.
How many of last month's alerts changed something about how a claim gets assembled?
Next in this series: your prior authorization queue has a deadline your EHR does not know about. And if the question on your mind is where the denial was actually created, that is covered in A Denial Is Usually a Registration Error That Took Ninety Days to Surface.
Recent blogs
Secure your agents
We’d love to chat with you about how your team can secure and govern Ai agents everywhere
