🩺 The Prediction Was the Easy Part July 19, 2026 · https://tavi-blog.github.io/the-prediction-was-the-easy-part/

The Centers for Medicare and Medicaid Services gave the insurance industry a deadline this year: seventy-two hours to decide on an urgent prior authorization request, seven calendar days for everything else, after the industry spent most of last year pledging to fix the process on its own. The CMS administrator put it to insurance executives about as plainly as a regulator can, telling them to fix it themselves or the government would write the rule for them. What both sides are now arguing about, in the trade press and in physician surveys published this year, is whether artificial intelligence is the thing that finally makes that deadline achievable, or the thing that makes the whole system worse while looking faster.

The optimistic case isn't a strawman. A prior authorization queue that used to take a reviewer twenty minutes per case can, in principle, be triaged by a model in seconds, freeing staff for the requests that actually need judgment. Requests have already dropped by double digits since the industry's reform pledge, and at least some of that is real process change rather than fewer people bothering to ask. If the bottleneck really is throughput, and a lot of it is, a faster first pass is a legitimate answer, and dismissing it just because AI is involved would be its own kind of denial.

What worries physicians, according to a survey published this spring, isn't throughput. Six in ten expect unregulated use of these models to increase how often care actually gets denied, and a majority already believe it's happening. A Senate committee report that circulated last year found denial rates on some AI-assisted reviews running many times higher than the historical baseline for comparable cases. Read one way, that's a model doing exactly what it was tuned to do for whoever owns it. Read another way, and it's the one I keep landing on, this isn't really a story about the model at all.

I've built one version of this problem, at a much smaller scale and with much lower stakes. A model I put together for work predicts how long a clinical study submission will take to clear initial authorization, based on attributes available on the day it's filed. It has nothing to do with denying anyone care. But the shape of the deployment problem was identical to what I keep reading about now: getting the prediction right was, comparatively, the easy part. Training the model on historical timing data and getting it to generalize took real effort, but it was bounded, solvable effort. The actual work was making that number land somewhere a coordinator with no data science background and no appetite for a new tool would trust and use, without asking them to install anything or take my word for how it was calculated.

That constraint shaped almost every decision I made, down to serializing the scoring logic into a spreadsheet script instead of running it anywhere that looked like a model. Not because the math needed to hide, but because the people relying on the number needed to open the file and watch it work in a format they already trusted. A prediction nobody can inspect isn't a tool, it's an oracle, and asking a research coordinator, or a patient, to just believe an oracle is a much bigger ask than asking them to trust arithmetic they can see happen.

The argument over prior authorization keeps circling back to whether the model is accurate enough, or whether it should be regulated the way a drug or a device would be. Both are fair questions. What they skip is whether the person on the receiving end of a denial gets to see anything about how it was reached, in a form that doesn't require a data science background to check. A state law now requiring insurers to report their AI-driven denial numbers to a regulator every quarter is a start, but a quarterly aggregate sent upward isn't the same thing as a patient seeing why their specific claim came back the way it did. One is oversight after the fact. The other is the harder, less fundable problem, the one that ate most of the time on my own much smaller version of this, and it's the one nobody in this debate seems to be building toward.