A study made the rounds again this month, and it should stop every health system that is penciling AI into next year's clinical workflow. Researchers watched endoscopists do colonoscopies before and after their practices switched on AI polyp detection. When those same doctors later worked without the AI, their adenoma detection rate dropped from 28.4% to 22.4%. The skill that catches early cancer eroded after a few months of leaning on the machine. That study is observational, so nobody can prove the AI caused the slide, and I would check the exact figures before you put them in a board deck. But a wave of newer work landed this fall that points the same way, and a November 5 roundup pulled it into one place. Give a clinician AI and the numbers look better while it is running. Turn it off and they are worse than before you installed it.
Here is the part the vendors selling you "AI efficiency" will not print on the slide. You are not buying a tool. You are renting a crutch, and the leg underneath it gets weaker every month you pay for it.
I want to be fair to the technology, because I am not an AI skeptic. We build AI at Welby and we use it every day. But read the actual evidence instead of the press release. A randomized study of 70 physicians this year had AI give either the first read or a second read on written diagnostic cases. First read scored 85%, second read 82%. No meaningful difference, and the study never bothered to test whether the doctor could still do the work once the AI went away. A separate trial put misleading AI explanations in front of medical students and watched their accuracy fall. Correct explanations did not lift them above working alone. So the AI helps when it happens to be right and hurts when it happens to be wrong, and the clinician cannot tell which one they are looking at in the moment. That is not a productivity gain. That is a coin flip wearing a lab coat.
And this is not new for documentation either. Back in April, researchers at the VA and the University of Washington ran 11 AI scribe tools and 18 human clinicians, including nurse care managers, against the same recorded visits. Blinded raters scored the notes. The humans won every case, some of them by a lot. The authors' own conclusion was that AI scribes should be treated as a rough draft a person has to fix, not a replacement for the person. That was the honest read in April. It is still the honest read now.
So let me say the thing most health system leaders are thinking and will not put in a town hall. The pitch for clinical AI was always that you could do more care with fewer clinicians. The newest evidence says you get a workforce that looks faster on a dashboard and is thinner underneath, and not one of these studies has shown a single patient lives longer because of it. You are being sold deskilling and told it is progress. The efficiency is real for about a quarter. The skill loss is the part that compounds.
None of this means you rip the technology out. It means you stop pointing it at the one thing it keeps failing, which is clinical judgment, and start pointing it at the mountain of work that is actually burning your people out. Prior authorization. Documentation a human reviews and signs. Scheduling. Eligibility checks. Outreach to the patient who missed a refill. That is where AI gives a nurse her Tuesday afternoon back without putting a patient at risk. At Welby, AI never makes a clinical decision. A person does. The software clears the paperwork so the clinician has the time and the standing to make the call herself, and to stay sharp enough to make it well.
Because the call is the whole game, and it does not happen in the exam room. Your quality scores and your readmission penalties are decided between visits. The diabetic who stops filling his metformin in week three. The heart failure patient gaining four pounds of water at home who waits for the ER to mention it. Patients forget somewhere between 40% and 80% of what you told them before they reach the parking lot, and I would verify that range before you quote it, though you already know from your own panel that it is directionally true. An algorithm does not fix any of that. A human reaching the patient on the right day does. We run that work at roughly one nurse to two hundred patients, and when it is done right we see A1c come down 20 to 30% and blood pressure brought under control inside 90 days. Those are our internal numbers, not a published trial, so pressure test them against your own claims data before you believe me. The point holds regardless of my figures. The thing that moves outcomes is a skilled person doing skilled work, and that is exactly the asset these AI deployments are eating.
So here is Monday morning. Pull a list of every place you have put AI or plan to this year. Sort it into two piles. One pile is the work where the software is making or shading a clinical call, triage, a diagnosis, a "is this patient okay" decision. The other pile is paperwork and logistics a human signs off on. Everything in the first pile is a skill you are trading for short term speed, and you will own the outcome when that skill is gone and the vendor has moved on. Everything in the second pile is fair game, and you should move faster there, not slower. If you cannot tell which pile something belongs in, that is your answer. You have not decided whether you are augmenting your clinicians or replacing them, and the technology will decide it for you if you do not.
Buy the tool that makes your nurse stronger. Fire the one that makes her forget how to do her job. Your patients can tell the difference even when your dashboard cannot.
Sources (verify before publishing):
Internal Welby figures (1 nurse:200 patients, 20 to 30% A1c reduction, BP controlled within 90 days) are internal results, not a published trial. Do not present as peer-reviewed.
