A behavioral health organization recently came to Joyful Health looking for guidance on how to strengthen its revenue cycle program. Partway through the conversation, the ask narrowed to something specific: rolling out an AI note-taker for their providers. Becky Carlson, Head of RCM at Joyful Health, had one question before going further. What did their documentation-related denials look like? Turns out, there weren't any.
It's a small moment, but it captures a bigger problem playing out across revenue cycle teams right now. Everyone is talking about AI. Budgets are being allocated, vendors are knocking, and leadership wants to know the AI strategy. But underneath the enthusiasm is a simple but critical question that most teams haven't actually answered: where does AI add real value in our revenue cycle, versus where are we just buying a shiny fix for a problem it can't solve?
That question is exactly what Joyful Health set out to answer last week. Carlson led revenue cycle, finance, and operations leaders through a three-session AI x RCM Masterclass, hosted by Joyful Health and Out-Of-Pocket, designed to cut through the noise. Her starting premise: before any team evaluates an AI vendor, they need to know whether they're solving a technology problem or a people and process problem.
Here's what came out of the three sessions.
Start with the problem, not the tool
The cohort brought together revenue cycle, finance, and operations leaders from across the industry — provider groups, health systems, digital health companies, even a payer or two. The idea that carried the most weight across all three sessions: start with the problem, not the tool. Before evaluating any AI platform, the group worked through a simple but often skipped question. Is the bottleneck a people problem, a process problem, or a technology problem? Applying AI on top of a broken process does not fix the process, it just gets you to chaos faster.
"Ask a vendor how they'll show you what it looks like to work with them," Carlson said. "Is this following your normal process? Is it scalable? If you give them 500 claims and they can get you an answer, what does it look like when you give them 5,000?"
The group also spent real time on evaluating and de-risking vendor relationships, including where payer data limitations will cap even the best built tool. A strong vendor can build a beautiful product and still be limited by what a payer is willing to send back, so understanding that boundary early prevents a lot of wasted implementation time.
Choosing the right team model
Carlson mapped the spectrum of team models — fully in-house, hybrid, fully outsourced — and the tradeoffs that come with each. In-house gives the most control but rarely gets prioritized for engineering resources. Fully outsourced offers speed and expertise but means giving up a meaningful amount of operational control. Hybrid models offer strategic control with more flexibility, but only work well when responsibilities and handoffs are clearly defined between teams. The most common failure point isn't the model itself, it's communication that breaks down between the people responsible for each piece.
None of these models are inherently right or wrong. The group spent time mapping which one fits based on organization size, growth stage, and how much internal capability already exists to support the function.
Where AI creates real leverage today
Across front-end and mid-cycle workflows, a few applications stood out as genuinely useful right now. In prior authorization, AI can auto-populate requests using clinical documentation and payer criteria, predict approval likelihood, and automate status tracking across payer portals so nothing sits unworked. In coding, coding copilots, HCC capture tools, and documentation improvement platforms are helping catch gaps before a claim goes out the door.
The group was equally clear-eyed on where AI still falls short. Novel diagnoses and off-label treatment requests still carry real uncertainty because there is not enough historical data to train against. Peer to peer escalations still require a clinician talking to a clinician. And AI code suggestions still require a qualified coder's review, since the coding decision on the claim remains the clinician's responsibility.
The organizations getting the most out of AI in RCM are the ones treating it as a co-pilot for their team, not a replacement for judgment.
Denials are a symptom of what happened upstream
One idea worth repeating on its own: a denial is rarely the root problem, it's a signal pointing to something that broke down earlier in the cycle. Back-end RCM, meaning denials, A/R follow-up, and underpayment identification, is also the hardest part of the revenue cycle to automate.
Carlson shared an example from one customer with more than $500,000 in CO16 denials, the code payers use for "missing information." When the team dug into the claims, the code meant four different things depending on the payer: a mismatched provider taxonomy code, a provider who was enrolled while their group was not, a missing payer-specific modifier, and several other distinct issues. One denial code, four different root causes, four different owners.
"It's six denials wearing a trench coat," Carlson said.
Beyond investigation, AI can help score open A/R accounts by dollar value, payer responsiveness, and likelihood of collection, so staff know where to spend their time first. But even the best scoring model depends entirely on the accuracy of an organization's own contract and fee schedule data, which shifts more often than most teams expect.
The group also talked candidly about a familiar risk in RCM teams: institutional knowledge that lives in one person's head.
"I like to call these folks the golden billers," Carlson said. "They bring so much value in making sure providers get paid, but they also introduce a layer of risk if that knowledge never gets disseminated across the team."
The data fragmentation problem
A thread that ran through every session, regardless of the specific topic, was data fragmentation. Remittances, EOBs, payer portals, and yes, still faxes, are scattered across systems that were never built to talk to each other. That fragmentation is exactly why back-end RCM is so hard to automate cleanly, and why so many organizations can see that revenue is missing without being able to explain where it went or why.
Attendees who mapped their own data flow across systems during breakout sessions kept landing on the same finding: the gaps between platforms are usually where denials are quietly born.
ROI is bigger than dollars, and vendor fatigue is real
The final stretch of discussion turned toward implementation. Groups worked through what ROI actually means beyond a straight dollar comparison, including the time and effort it takes to stand up a new workflow, and who inside the organization has the bandwidth to own that.
"If you can attribute a minute of time spent to an activity, you can attribute a dollar value to it, whether that's the person taking the action or the thing they aren't able to do because they're focused on this instead," Carlson said.
Several attendees raised a familiar tension: the risk of accumulating five or six point solutions that solve individual problems well but do not talk to each other. The people doing this work every day already know where the problems are. The barrier is rarely a lack of expertise. It is bandwidth, ownership, and the infrastructure to connect what everyone already knows into something actionable.
Didn't get to attend?
This recap only scratches the surface of what came out of three sessions of live discussion, breakout conversations, and peer problem solving. If any of this resonates with what your team is navigating around denials, aged A/R, or underpayments, Joyful Health would love to hear from you. Get in touch today.
