🗄️ The Back Office the Report Didn't Count August 20, 2026 · https://tavi-blog.github.io/the-back-office-the-report-didnt-count/

A report making the rounds this week says roughly a quarter of hospitals, and more than half of health plans, now use AI somewhere in their administrative workflows, with another twenty-one billion dollars in savings still sitting in processes nobody's gotten around to automating yet. The number comes from an index that's been tracking healthcare's administrative costs for years, and this year's read is that the rebuild is happening quietly, function by function: denial prevention, coding audits, revenue cycle, staffing. Not the dashboards and copilots that make headlines. The unglamorous stuff, finally getting touched.

I want to take that at face value, because it's probably true for what it's measuring. Cutting the time a claim sits in a denial queue, or the hours a coder spends cross-checking a chart against a billing code, is a real efficiency gain with a real dollar figure attached, and dollar figures are what get budget approved. If a hospital can point to a documented return on a prior authorization tool, that tool gets funded next quarter. That's not cynicism, that's just how large institutions decide what to build. The twenty-six percent adoption number is a genuine sign that some of this is finally sticking instead of dying in a pilot.

Here's what the number can't see, though, and it's not a small gap. Everything in that index is scoped to the part of a hospital that touches a payer. Claims, coding, prior authorization, denials, staffing to cover volume, all of it eventually resolves into a dollar amount that shows up on someone's balance sheet, which is exactly why it gets surveyed, funded, and counted. There's a whole other back office running in parallel that never touches an insurer at all, because its counterpart isn't a payer, it's a regulator. Every hospital that runs clinical research has a layer of administrative work behind that research: screening submissions before they reach an ethics board, tracking approval timelines, keeping consent and regulatory documentation current across a portfolio of studies that each move on their own clock. It's real institutional overhead, and it's every bit as manual and error-prone as a claims queue. It also produces no return-on-investment story anyone can hand to a finance committee, because nothing in that pipeline generates or saves revenue in a way that's easy to point to. So it doesn't get surveyed by an index built around payer-facing costs, and by extension it doesn't get counted as part of "the industry," even though it sits inside the same buildings, sometimes the same floor.

What automation exists there mostly isn't the AI-vendor kind the report is describing. It's closer to whatever the person who noticed the bottleneck could build with the tools already licensed to them, patched together after hours or squeezed in around other work, because there was never a line item for a vendor contract to fix it properly. A flow that catches a missing field before it becomes a rejected submission. A dashboard that finally puts three separate tracking systems on one page instead of three browser tabs. None of that shows up in an adoption survey, because the survey is asking whether the institution bought something, not whether someone inside it built something to cover a gap the institution never officially acknowledged.

The honest version of the twenty-one billion dollar figure, then, is that it's a lower bound on what's actually sitting in manual process across a hospital's administrative surface, not an upper one. It only counts the manual work that has a payer attached to make it visible as a cost in the first place. The research and regulatory side of the same institution is carrying its own version of that debt, unmeasured, because nobody's built an index around it, and nobody's going to, because there's no invoice at the end of it to make the case.

What strikes me isn't that this layer gets left out. It's that the report can be entirely accurate about what it measures and still leave the impression that the hospital back office is one thing with one adoption curve. It isn't. It's two systems running side by side under the same roof, one of them legible to a spreadsheet because money moves through it, the other invisible for the same reason it stays underfunded: nothing about it was ever built to be counted.