🧮 An Agent That Never Acts Still Needs a Number October 21, 2026 · https://tavi-blog.github.io/an-agent-that-never-acts-still-needs-a-number/

The new scorecard on healthcare AI returns sorts every deployment the surveyed hospital systems reported into one of two drawers: clinical, or everything that isn't. Revenue cycle work, the software that chases down a denied claim or codes a bill, lands almost entirely in the second drawer, and it's also the one with nearly all of the autonomy in it. About two-thirds of those tools now run with little or no human in the loop. Clinical deployments sit at the opposite end, barely a handful of percentage points autonomous, still mostly advisory: a program that drafts a note or flags a result for someone else to read and decide on. The report frames this as a story about autonomy driving returns, and on its own numbers it isn't wrong. Letting software finish a task on its own, instead of stopping to ask a person, is where most of this year's realized return actually sits.

That framing deserves its due before I push on it, because it isn't a lazy read. If a system can both find the problem and fix it without anyone standing in the middle, you've collapsed a task that used to cost a person's attention twice into one that costs it never, and that's a real, countable saving. A claim gets rebilled correctly on the first pass instead of the third. Nobody has to notice, let alone measure the noticing. The honest version of the report's argument is that autonomy pays precisely because it removes the step where a human would otherwise have had to look, and looking is expensive even when it's quick.

Where it loses me is the drawer itself, because I've spent real time trying to put a number on the kind of system that never sorted cleanly into either one. For a while, part of my job was turning the usage logs of an advisory agent inside a hospital research operation into something leadership could trust well enough to keep funding it. The agent didn't act on anything. It answered a researcher's question about a submission timeline or a regulatory requirement, the kind of thing that used to mean an email to a colleague who had better things to do, and then it got out of the way. No claim got rebilled. No record got updated. The only trace it left behind was a transcript and a session count sitting in a table nobody outside the team would ever open.

That trace is real information, but it isn't a dollar figure, and building a model that turned it into one meant making judgment calls nobody could fully defend. How much is a researcher's few minutes worth when the alternative wasn't zero minutes, it was someone else's afternoon instead? Do you credit the agent for every question it answered, or only the ones that would otherwise have become a request sitting in a queue somewhere, which you'd have to guess at, because most of them never would have been logged as a request in the first place? I didn't invent the resulting numbers out of nothing, but I chose the assumptions that produced them, and a different reasonable person inside the same program could have chosen assumptions that produced a noticeably different figure. Revenue cycle software runs into that problem far less often. A rebilled claim carries a dollar amount attached to it by the payer, not by the team trying to justify the tool.

None of that makes clinical and advisory tools worth less than the autonomous back-office ones. It just means the back-office tools get to inherit a number that already existed before the software ever showed up, while the advisory ones have to manufacture one from scratch, in full view of the same leadership deciding whether to keep paying for them. A report built from a survey of executives is going to reflect whichever return was easiest for those executives to state with confidence, and confidence tracks the availability of a pre-existing figure far more than it tracks whether the thing actually helped anyone.

I don't think the model I built undercounted what that agent was worth on purpose, and I also don't think it got the number right. It was the most defensible figure I could construct out of a kind of use that was never going to show up on anyone's claims ledger, filed under a line item that existed mostly because a person needed one to point to. I'd guess most of the advisory tools sitting in that quiet second drawer are being measured, right now, by someone doing exactly the same thing: making a defensible call, writing it down as if it were a fact, and never once getting asked which assumptions got them there.