Not Yet — Summary

Renata approved the rollout in eleven weeks at the retailer and has spent fourteen months blocking it at the hospital, and it is the same tool. At the retailer she tested the model against six months of campaign copy, found it better than the agency they were paying, and had it in production before the renewal. At the hospital system she runs clinical documentation improvement, and the same vendor came in with the same architecture and a demonstration more accurate than her staff. She has said no four times.

The standard account of why good technologies take decades is the cost of the change around them. Electric motors were available from the 1890s and the gain did not appear until the 1920s, because the factory was built around a central driveshaft with machines placed by available torque, and getting the benefit meant a different building. That is a capital program. None of it transfers. The model arrives as a subscription, and her hospital could run it against live documentation inside a month. The thing that explains the forty-year lag is absent and the lag is here anyway.

What she is doing in both jobs is comparing a distribution against a consequence: not the average but the bad tail, multiplied by what a bad output costs. At the retailer the tail was cheap and she could name the number. At the hospital the mean is also better and she does not dispute it, and the tail is a systematic coding pattern running undetected across forty thousand encounters and surfacing as a federal audit, where the question is whether the hospital can produce a person who reviewed it, and the answer decides whether this is a repayment or a fraud allegation. The bar is set by the tail, and capability improvements move the mean.

Which is why the demonstration never lands. A vendor’s evaluation shows the mean, because that is what a pilot surfaces in six weeks. A pilot long enough to characterize the tail would run until the rare failure occurred, and if it occurs the pilot has failed. The evidence she needs is the evidence nobody has an incentive to produce, and both parties leave the room believing the other is unreasonable.

Her arithmetic has a large error in it and she knows. The current process also fails, continuously, and those failures fall on patients and belong to no one. An error made by a system somebody chose to install has a name on it. Correcting the arithmetic would not change her answer, because she is not the party bearing the diffuse cost and very much the one who would bear the attributable cost. The gradient is consequence rather than regulation: marketing inside the same hospital adopted the same class of tool eight months ago with no friction at all.

Three things move the bar and none is a better model. A way to bound the worst case rather than reduce its probability, which is a different kind of object. Attribution, making an error traceable to a party who could have decided otherwise, which her staff supply by existing and the vendor cannot, because its liability position depends on not supplying it. And precedent, which is why adoption in high-consequence settings looks flat for years and then goes vertical in a shape unrelated to the capability curve. There is a fourth and it is the ugly one: consequence can be reassigned by indemnity, safe harbor or insurance, and most of what will look like a wave driven by improving models will have been a wave of contracts.

I wonder whether the language of readiness has been backwards this whole time, and organizations described as not ready are more often organizations that have correctly identified something the people selling to them have no way to supply.

The hospital will adopt this, and Renata expects to be the person who approves it, on materially the same evidence she has been declining for fourteen months. The printed test is still in her desk, under a stapler she does not use. It is the only piece of paper she owns that records her being right about something before anyone else in the building agreed.