Every model was right. The bill was still wrong.

When I think of AI in RCM these days, I picture a little bit of a maze. It’s not supposed to be a maze. Putting tech in place to streamline processes should make things easier to navigate, after all.

You've been in the physical version of it. You follow the signs to Radiology, the signs stop, and you're standing at double doors that open onto a loading dock. You have to go down to two to get up to four, because the 1974 tower and the 1998 tower don't share an elevator bank. There's a volunteer parked at the junction with a laminated map who's been pointing people toward Imaging since the Clinton administration.

Every one of those wings was a good decision. The tower was needed, the addition was funded, the connector was the cheapest way to join two buildings that were never designed to touch. Committees and boards made the decisions that led to it.

And here we are, doing the same thing again, in software.

Somebody in my comments named it

Last week I wrote about Mayo's AI chief saying maintenance costs are far higher than anyone expected, and about the fact that almost no AI business case in healthcare models year three (Becker's Hospital Review, Naomi Diaz). It got more response than anything I've published.

Down in the comments, a reader named Ra Puriri-Ahmu made a point that reframed the whole thing for me. His argument was that once a health system has 128 AI solutions in production, it's not managing models anymore. It's managing a living system where every model can be locally accurate while the portfolio goes collectively contradictory. He called it coherence debt.

He's building a product in that space, so read his interest the way you'd read any vendor's. The concept still holds, and it isn't his alone to sell. It's a real category of liability that behaves almost exactly like technical debt did for thirty years before finance people started taking technical debt seriously.

Mayo described it without naming it

The week after that maintenance story, Mayo's chief AI implementation officer characterized its 128 clinical tools this way: they're point solutions, each one aimed at a particular disease or specialty or a different part of you, without ever looking at the whole of you (Becker's Hospital Review).

That's a clinical way of saying something financial. Swap "a different part of you" for "a different part of the account" and you've got the revenue cycle version, which nobody's writing about because revenue cycle never gets the interesting article.

Walk one account with me

Take a patient. I'll call her Denise, because everybody's got a Denise.

Denise gets an estimate at scheduling from a tool trained on last year's payer behavior, and she's quoted $400. A propensity-to-pay model scores her; she's paid two small balances on time, so it routes her to standard follow-up instead of financial counseling. A charity screening tool runs in the background, but it only fires above a dollar threshold her estimate didn't cross. The visit happens. A coding tool applies the new CPT 2027 logic correctly, because somebody updated it in December like they were supposed to. The claim goes out. The denial prediction model, still carrying 2025 payer patterns in its weights, doesn't flag it. It denies. It gets reworked.

Her final bill is $1,900.

Now go find the error.

There isn't one. Every model did the job it was bought to do at the accuracy the vendor promised, and any one of them would pass an audit on its own. The estimate was right for the rules it had. The coding was right for January. The propensity model was right about her payment history. The charity tool followed its threshold.

Denise got a bill for almost five times what she was told, from a hospital that had every single piece of information it needed to prevent that, sitting in five systems that never spoke to each other.

She pays it or she doesn't. Either way she tells eleven people, and next time she has a choice she drives the extra twenty minutes. That's not a claim you lost; it's a household. Households are where lifetime value actually lives, and no AI ROI deck I've ever seen has a line for them.

Agentic turns the slow leak into a burst pipe

Most of these tools still suggest. A human sits at the seam and catches some of it, inconsistently and on a bad day not at all, but they catch some.

That's changing fast. Dr. Tomi Kolade at UTHealth Houston described the shift as moving from AI that summarizes to AI that prepares orders, works the inbox, coordinates care, and runs multi-step workflows inside the EHR. His line about governance is the one I'd tape to a wall: frameworks built for generative AI aren't adequate for agentic AI (Becker's Hospital Review).

Stanford's CTO, Christian Lindmark, said his biggest fear is agentic AI outrunning the ability to monitor it. Today he can handle it; relatively few agents are running and humans still review the consequential calls. What worries him is the version with hundreds or thousands of agents spread across clinical and administrative work, and he's blunt that retrofitting oversight after they're embedded would be far riskier than building it in now (Becker's Hospital Review).

Another reader, Gail Galusha, raised the version of this that ought to worry HR as much as IT. She's heard people predict a future where a single employee has hundreds of small task-specific models working for them, and she asked the right question: what happens when we lose that employee?

Nothing breaks. Everything keeps running exactly as configured, and not one person left in the building knows why it was configured that way.

Somebody's already funding a referee

ARPA-H just put $62.7 million behind the first FDA-authorized agentic AI system for cardiovascular care. Three companies build the patient-facing agents; Duke and Kaiser Permanente handle real-world deployment across dozens of sites (Becker's Hospital Review).

Stanford's role in that program is to build a separate supervisory AI whose entire job is watching the other agents for unsafe recommendations.

The federal government's flagship autonomous care program budgeted a referee before it built the players. Not a committee, not a quarterly review. A funded technical layer, because somebody in that room understood that a pile of individually correct agents isn't the same animal as a correct system.

If that's table stakes for a research program, what is it for the health system running nineteen vendors nobody's ever mapped?

What I'd actually do about it

None of this needs a consultant or a new steering committee. It needs about two weeks and a willingness to look.

Start by rebuilding your AI inventory around decisions instead of vendors. The list you have now is organized by tool and department, which is exactly the arrangement that hides the problem. Redraw it around the account: which models touch the same patient, the same claim, the same balance? Most people have never drawn that map, and drawing it is ninety percent of the work.

Then trace one account all the way through. Pick a real denied account that came with a patient complaint attached, and list every model, rule, and bot that touched it, in order, along with what each one assumed. You'll find a contradiction on your first try. I've never seen this exercise come back clean.

Name somebody who owns the portfolio, not just the tools. You've got owners for every application. You almost certainly don't have one human being accountable for whether those applications agree with each other. Until that job exists, coherence belongs to nobody, which means it belongs to the CFO in about eighteen months.

Test the seams instead of the tools. Vendor accuracy claims are per-model; end-to-end failure isn't. Your validation should follow a workflow across systems, not a model inside one.

Put it in the contract now, while you still have leverage. Lindmark's sticking points make a ready-made list: rights to use your data to train products for other customers, vague data-use language, resistance to audit rights, refusal to disclose underlying models or subprocessors, and liability capped at a fraction of contract value while you carry the real exposure. He treats walking away as routine rather than a last resort. That posture's available to you too, and it's cheapest before you're dependent.

And get comfortable saying no. CommonSpirit generated more than $100 million in AI value last fiscal year, and the number I keep chewing on is that its governance process has rejected fifteen proposed use cases (Becker's Hospital Review). Fifteen tools that'll never drift, never need monitoring, never contradict anything. That's coherence debt they simply declined to borrow.

The invoice never comes

Dr. Enitza George at SUNY Downstate said something this month that belongs in every AI business case in the country: technology scales whatever system you hand it, including the inefficiencies and the bad assumptions. Automate a broken process and you don't just preserve the break. You make it faster and much harder to see.

Coherence debt is what you get when you do that nineteen times and never check whether the nineteen agree.

Nobody's going to bill you for it. It shows up as a denial rate that won't improve no matter what you buy, and a cost-to-collect that flattens out at a number you can't explain, and a conference room full of smart people who each have a dashboard proving their tool is working.

What I'd like to know: if you traced one account through every model that touched it, what would you find? Tell me, and I'll publish the best answers anonymized in a future issue.