This is the final piece in a three-part series. Part 1 looked at where revenue actually breaks down, which is usually much further upstream than the billing queue. Part 2 covered the signals that show you a problem forming while there is still time to act. This one takes on who, or what, should read those signals.
Automating a broken process doesn’t fix what’s broken. It just allows the same problems to happen faster and at greater scale. A missed hand-off becomes more missed claims. Gaps in data visibility become harder to untangle as more activity flows through a system that still can’t explain what’s happening. Whether an AI initiative improves net collection rate or only speeds up the work comes down to what is already in place underneath it.
The order I come back to is people, process, then technology. Technology applied to a process that already works makes that process faster and more consistent. Technology applied to chaos gets you to chaos faster. If a handoff between intake and billing is unclear about who owns it, automating that handoff produces the same unclear handoff at higher volume. The gap does not get fixed. It gets harder to see, because now it is running underneath a system nobody has to touch.
That is the part the market conversation tends to skip. Most of the discussion is pitched at whether to adopt AI. The more useful question for anyone running an operation is which parts of the work it should touch, and in what order.
Test readiness before adding AI to your revenue cycle
Before adding technology to any part of the revenue cycle, I ask three things, and I ask for evidence rather than an answer.
Without those three answers, there is no way to prove later that the technology worked.
What AI does well in revenue cycle management
Once that groundwork holds, there is real work the technology handles better than a person can.
Reconstructing a claim’s full timeline across systems. The claim history lives in the practice management system. The payer response lives in the remittance. The authorization lives somewhere else again, and the documentation lives with the clinical team. A biller answering one question about one claim used to open four systems to do it. Pulling that timeline together is exactly the kind of retrieval work that should not consume an experienced biller’s day.
Sorting denials at volume and tracking which categories are moving. This is what makes the leading indicators from Part 2 practical at scale. Watching denial categories as a share of total, month over month, is useful when you can do it across your full volume rather than a sample. At Joyful Health, we've processed more than $1.4 billion in claims. At that scale, technology makes it possible to identify patterns across a volume of work that would be difficult to evaluate claim by claim.
Working the steps where the right action is clear and the result is checkable. Plenty of claims have an obvious next step and a quick confirmation that it worked. That work can move without a person in the middle of every touch, with people picking up what gets stuck.
Spotting patterns nobody catches claim by claim. A payer slowing down. A category climbing in one region. Exposure building ahead of a new state launch. No individual claim shows you that. The pattern across thousands of them does.
Where RCM expertise still matters
The more of the revenue cycle that runs on its own, the more it matters where experienced people are pointed.
Some situations require context that reaches past an individual claim: payer relationships, organizational priorities, and decisions about how the business should operate. Three come up constantly.
How to balance recoverability against cost to collect. A claim 90 days outside the timely filing window may still be appealed, but the likelihood of payment is low and the effort required may outweigh the potential recovery. AI can assess claims against established recoverability criteria and act accordingly. The organization still needs to decide where those thresholds sit, when exceptions are justified, and how aggressively to pursue marginal balances. Those decisions reflect financial priorities that extend beyond any single claim.
How to position an appeal, and what a payer is signaling behind a response. Payer behavior carries context. A response that technically says one thing often means something more specific about how that payer is handling a category right now. Reading that and deciding how to come back at it is expertise.
What to change upstream once a pattern is identified, and who owns the change. The system can show that authorization denials are climbing in one region. Deciding whether that is a staffing question, a workflow question, or a payer question, and then putting it on a specific person’s desk, is a human call. It is also the one that determines whether the pattern stops.
That combination of technology and experienced people is what we build around at Joyful Health. Our model has achieved a 95% claim resolution rate across specialties, with AI handling work at scale and RCM experts applying judgment where the situation calls for it. The goal isn't to remove people from the process. It's to make sure their expertise is focused where it can have the greatest impact.
How to decide what to automate in your revenue cycle
There is a test you can run against your own task list this week.
First: can the right action be determined from reliable information and clear decision criteria? Consider what information is available, how consistently the work can be evaluated, and whether the system has enough context to act. The more clearly those conditions are defined, the more work AI can take through resolution independently.
Second: what happens when the system encounters something unexpected? A well-designed automated workflow needs a way to recognize uncertainty, document what happened, and escalate when necessary. That becomes especially important when decisions affect recoverability, timely filing deadlines, or financial reporting.
The goal is to expand what AI can resolve independently while making sure the exceptions reach the right people with the context they need.
Why AI alone won’t prevent denials
That last part is where prevention becomes measurable, and it is worth holding the organization to it. Three numbers show whether the loop is closing:
- Repeat denial rate by root cause. If the same driver keeps producing denials after a fix, the fix did not hold.
- Time from root-cause finding to implemented change. How long an identified problem sits before anyone acts on it.
- Share of root-cause findings with a named owner. A finding without an owner is an observation.
That brings us back to people, process, then technology. Clear ownership and well-designed workflows give AI the foundation to operate effectively. AI then creates capacity, accelerates resolution, and surfaces patterns that teams can use to make lasting improvements.
That is the whole series: understand where revenue breaks down, recognize the signals earlier, and build a revenue cycle where technology and experienced people work together to resolve problems and prevent them from recurring.
From Denial Resolution to Lasting Prevention
Eric Worthan, CEO of Ally Psychiatry, rebuilt most of his revenue cycle internally across 71 locations and seven states before concluding that one piece was not solvable in-house: the extra five to ten percent of revenue disappearing through what he calls “daily paper cuts,” where every lost dollar comes straight out of contribution. Finding that revenue took claim-level root-cause work, and keeping it took getting those findings back to the people who owned the upstream process.
In our Denial Prevention Virtual Event, Eric and I walked through how his team is putting denial intelligence into action, addressing upstream issues, and preventing the same denials from recurring.
