Every health system wants clinical AI. Far fewer want to talk about how to pay for it.
The economics are hardest exactly where the need is greatest. In safety-net and public systems, margins are thin and the payor mix is unforgiving. Clinical AI — the kind that touches diagnosis and treatment — is expensive to build, slow to validate, and high-stakes to deploy. Pitch it to a board on faith and you'll wait years for a yes.
The usual response is to pitch AI as a cost-cutter instead. I think that's the wrong frame entirely. The point of AI in healthcare isn't to cut cost — it's to expand capacity. Staff, beds, and clinic hours are finite; demand is not. The question that matters isn't "how many dollars did we save?" but "how many more patients did we serve with what we already have?"
But conviction doesn't fund a platform. So here's the operating model I've come to believe in after fifteen-plus years building data and AI products in healthcare: start in the revenue cycle, and let the back office pay for the bedside.
Where AI earns its first keep
Denials, charge capture, payment variance — this is where AI proves itself fastest. The outcomes are measured in dollars and visible within months. The stakes are lower than anything touching a treatment decision, so you can iterate quickly. And the data — claims, remittances, clinical documentation — already exists. Done well, those early wins do two things at once: they fund the data platform every later use case needs, and they buy credibility with the clinicians and executives who will decide whether clinical AI ever ships.
Utilization management is the example I've seen up close. Deciding whether an inpatient admission is medically necessary used to mean a clinician spending hours combing documentation. An AI tool that reads the notes, labs, and encounter data in real time and lines the case up against clinical criteria turns those hours into minutes — surfacing the vitals, results, and presenting symptoms that matter, with suggested diagnosis codes and the evidence behind each one. Reviewers confirm rather than excavate. The tool also catches supportable codes that human review missed, which physicians — reasonably skeptical at first — noticed and appreciated. The result is fewer denials, protected reimbursement, and faster patient flow. And a line I hold firmly: the AI is decision support, not the decision-maker. A human makes the call; the machine does the reading.
The data layer underneath
None of that works without the data foundation, and three lessons there have held up everywhere I've applied them.
Work backwards from a use case. "Get the data ready for AI" is an ocean-boiling project with no finish line. "Get these four feeds ready for utilization review" is a quarter of work.
Build a canonical model early. Patients, encounters, claims, providers — abstracted into consistent objects with clear lineage rules. Every use case after the first gets cheaper.
Don't wait for modernization. Plenty of the feeds that matter still arrive as HL7 messages and CSVs over SFTP. That's fine. Waiting for a pristine platform is how AI programs die in year one.
Champions, not committees
The people side matters as much as the pipes. The single biggest success factor I've seen is a champion inside the domain — someone with real subject-matter expertise and the appetite to sit through iterative build cycles. Some clinicians operate like startup founders; find them. Then run the early work like an internal consulting group: small, fast, validating the concept with the people who'll actually use it — before you formalize anything or buy infrastructure.
The most common mistake is the mirror image: building the thing without a rollout plan, an adoption target, or monitoring. An AI model isn't a report you ship and forget. It drifts, the world changes underneath it, and adoption — not accuracy in a lab — is the only metric that ultimately matters.
Equity is an accuracy requirement
One more thing belongs on the first-wins list, and it isn't financial. Systems that serve everyone can't treat equity as a compliance checkbox. Language models perform worse in languages other than English; models trained on unrepresentative data lose accuracy for exactly the patients who can least afford it. If your patients speak sixty languages, a tool that only works well in English doesn't work.
The engine, not the point
The revenue cycle isn't the point. It's the engine. Every recovered dollar and returned hour compounds into the thing that is the point: more patients seen, sooner, with the staff and space you already have. Cost-cutting shrinks a health system to fit its budget. Capacity-building grows what the budget can do.
Start where the money leaks. Spend what you recover at the bedside.