93 million CT scans were performed in the United States in 2023. About a third covered the abdomen and pelvis. The colon is in the picture on every one of them, and almost nobody was looking at it. A model published in Annals of Oncology this spring reads the colon on the noncontrast CT that was ordered for something else. In 18,419 consecutive real-world scans, it ran at 86.6% sensitivity and 99.8% specificity, with a positive predictive value of 63.4%. Five cancers nobody had found. About 30 seconds a scan. Ten radiologists using it gained 20 points of sensitivity. This is not a colon story. Next month it will be a different finding on a different scan, and the month after that another one. Each arrives as a paper, then a vendor, then a question for a radiology department that answered the same question last month for a different tool. Departments deciding tool by tool cannot keep up, and most have stopped trying. Hospitals already have the structure for a stream of new capabilities that each need an evidence bar, local validation, a pathway before adoption, monitoring after, and a process for removal. It's called a formulary. New drugs arrive every month and no hospital relitigates whether to have a pharmacy. Radiology needs the same committee for algorithms. The platforms to plug them in already exist. The committee mostly does not. Then the part I care about. I order the scan for a kidney stone. At 2 am the read comes back: stone, and a flag. Probable colorectal mass, model-detected. What do I do with that? Today the honest answer is that I tell the patient, I write "follow up with GI" on the discharge paperwork, and the loop closes some of the time. The ED already has a process for results that arrive after the patient leaves. Positive cultures get a callback, a nurse, a list, a phone call until someone answers. A model-detected mass belongs in that process, with an owner and a tracking list. Referral is the easy part. Whether the scope happens is the system. One more thing radiology will have to contend with. A local validation takes months. The model's successor ships in three. Validating once and deploying forever cannot work at this cadence. The tool has to be checked the way a lab analyzer is checked, continuously, against known results, with an alarm when it drifts. A model can now find a cancer on a scan nobody ordered it for. Who owns that result?