The Layer Beneath the Model
A 3-Arc Framework
Every few weeks, I sit on a call where someone presents an AI pilot in a health system. The slides are polished. The metrics look promising, and somewhere around the thirty-minute mark, the conversation drifts — quietly, the way important things often do — toward what isn’t on the slide.
The data didn’t move from System A to System B.
The model ran, but the workflow didn’t catch the output.
The clinician dismissed the EHR alert because it appeared eleven seconds too late.
The pilot ended.
The patients didn’t.
This is the conversation that’s been happening in every health system conference room I’ve sat in for over a decade. It’s the conversation that doesn’t make it into the trade press, the conference keynotes, or the strategy decks. And it’s the conversation that determines whether AI in healthcare ever actually reaches a patient.
I want to name it publicly.
The model isn’t the problem
The healthcare AI conversation has a model problem only in the sense that we keep talking about models. Every week, a new foundation model. A new clinical reasoning benchmark. A new partnership announcement.
Meanwhile, on the floor of an inpatient unit in Greater LA — or an ambulatory clinic in the Inland Empire, or a research cohort at an infectious disease clinic in San Diego — the bottleneck isn’t intelligence. It’s whether the right data reaches the right point of decision in time to change what happens next.
The layer beneath the model is infrastructure. And in most health systems, it is the most underdiscussed, undercapitalized, and underdesigned layer in the stack.
This is the upstream ripple effect that fewer people in this space are willing to trace: a missing data integration in 2022 becomes a delayed prevention screening in 2024, becomes a late-stage diagnosis in 2026, becomes a patient outcome we then try to fix with another pilot. Are we really practicing preventive medicine? Is it delivering optimal care, reducing costs, addressing clinician administrative burden? The model never had a chance, because the layer beneath it was never built.
We do not have a problem with the model. We have a healthcare infrastructure connectivity issue we keep trying to solve with ever-hyped fragmented “AI models” — each tailored for a specific need, none designed for the ecosystem-wide solution healthcare actually requires.
The 3-Arc Framework
After a decade implementing across the largest integrated health system and a top academic medical center, I have found the same architectural pattern in every organization I’ve worked with — including the ones that don’t yet know they have it.
I call it the 3-Arc Framework.
Arc 1 — Interoperability. Can the data move? Not in theory. Not after a nine-month integration project layered with process improvements, spread programs, Agile, Lean Six Sigma, and whatever methodology was in fashion that fiscal year. In real time, between the systems that need to see each other. If the answer is no, no model on the planet will save the workflow downstream.
Arc 2 — Analytics. Once the data moves, can it be made legible — to clinicians, to operational leaders, to revenue cycle, to the C-suite? Reporting is not the same as insight. Scaling Tableau inpatient reporting and watching it organically spread nationally at the largest integrated health system taught me that the distance between raw data and a decision is a design problem, not a tools problem.
Arc 3 — Intelligent Action. Only here does AI earn its keep. Action is what happens when a prevention screening flag fires inside every patient encounter automatically — not when a model produces a recommendation no one acts on. Action is the test of whether Arcs 1 and 2 were built right.
You cannot skip arcs.
AI applied to a broken Arc 1 amplifies the breakage. AI applied to a missing Arc 2 produces sophisticated outputs nobody trusts. The order matters because patients are downstream.
What I am not saying
I am not saying AI doesn’t belong in healthcare. I am saying it belongs in the right arc of a system that was built to receive it.
I am not saying interoperability is a new conversation. I am saying we have spent twenty years calling it a technical problem when it is, more accurately, a design and governance problem that technical tools cannot resolve alone.
I am not saying any of this is original to me. I am saying it has been said quietly, in conference rooms, after the slides close — and that the people saying it operationally are not always the ones writing publicly. That is the gap I am stepping into.
What this opens up
I am developing the 3-Arc Framework into a methodology paper, in conversation with collaborators at the intersection of implementation science, health policy, and population health. The thesis is field-tested; peer review is the next step, and I welcome it. I am interested in talking with operational leaders, researchers, providers, and policy practitioners who have seen the same pattern — or who see it differently and can sharpen the argument.
If you have been part of the offline conversation about why healthcare AI stalls — the one that happens after the slides close — I want to hear from you. Especially if you have watched a pilot end and asked, quietly, what would have actually moved the needle.
A healthcare infrastructure and operations leader with over a decade across UC San Diego Medical Center and Kaiser Permanente. She holds a Master of Applied Science in Population Health Management from the Johns Hopkins Bloomberg School of Public Health. She writes Perspectives biweekly on the operational layer of healthcare, emerging science and technology, and the public conversations that haven’t yet happened.
