On July 8 a study went around that most health system leaders scrolled past. A randomized trial, published in Nature Medicine, tested a generative AI tool built to help clinicians in primary care. The AI improved the documentation. It pushed clinicians toward more guideline-concordant care. And it did nothing measurable for the patients. Treatment failure ran about 2 percent in the control group and 2.2 percent in the group using the AI. No statistically significant difference. The senior author said it about as cleanly as you can: the technology looks safe and clearly improves the decision-making, but turning that into real patient benefit is much harder, especially in everyday primary care.
Here is the part nobody in a health system wants to say into a microphone. You spent the last two years buying AI to make your doctors' notes faster, and the best evidence we have says you bought a better note and the same patient.
Let me be fair to the study before I use it, because Seth-brand honesty matters more than a clean argument. This trial ran in 16 primary care clinics in Kenya, and most of the cases were infections and antibiotic decisions. That is not American chronic disease. A smart CMIO who read it pointed out, correctly, that U.S. primary care is mostly diabetes, hypertension, and cardiometabolic risk, the messy multi-system stuff where an AI's ability to synthesize a pile of data could plausibly help more. Fine. I will grant all of it. The trial does not prove AI is useless in the exam room. But it lands on exactly the thing the whole industry has been careful not to test, and the answer it gave should make you uncomfortable, because it matches every other honest data point we have.
Look at what you actually know about the tools you already bought. Ambient AI scribes are on track to be one of the fastest-adopted technologies in the history of American healthcare. The Peterson Health Technology Institute counted around 60 products on the market last year. As of last June, nearly two-thirds of hospitals running Epic had turned ambient AI on for documentation. And Peterson's read on all that spend was blunt: the scribes cut clinician burnout and cognitive load, and they have not yet shown a clear return on investment. Cost runs somewhere between $100 and $600 per provider per month depending on the vendor. Do that math across a medical group and tell me it is a rounding error.
Now, the burnout number is real and I am not going to pretend it isn't. A JAMA Network Open study out of Mass General Brigham and Emory found something like a 21 percent absolute drop in burnout at Mass General Brigham after 84 days, and Emory reported roughly a 31 percent jump in documentation-related well-being at 60 days. (Verify those exact figures before you quote them, but the direction is not in dispute.) That is worth money. Doctors who hate their charts less quit less, and turnover is its own line item. Keep the scribes for that reason if you want. I would.
But that is not the reason you told your board you were spending the money. You did not stand up in a strategy meeting and say we are going to invest seven figures to make our physicians feel better about paperwork. You said this was going to improve care. And the quiet finding underneath all the pilots is that faster, prettier documentation of the visit does not change what happens to the patient after the visit ends. It was never going to. The note is a record of the twelve minutes you already had. It is not the thirty days you didn't.
This is the whole argument, and it is the one I have been making since before it was fashionable. Outcomes are not decided in the exam room. They are decided in the gap after it. The patient's blood pressure does not move because the SOAP note was elegant. It moves because somebody called in week two, checked the reading, caught that he stopped the medication because it made him dizzy, and fixed it before it became a stroke. That work does not happen inside the visit, so no amount of AI aimed at the visit will ever touch it. You have pointed the most powerful new tool in a generation at the fifteen minutes that were already the least broken part of the system.
I am not anti-AI. We run AI agents at my company every day. I just have a rule about where they get to stand. AI never makes the clinical call. A human always does. What AI is genuinely good at is the enormous, boring, between-visit machine that no health system has ever had enough people to run. Calling the eligible patient who never gets called. Getting the reluctant one to consent. Reminding, scheduling, chasing the reading, flagging the one number that means a nurse needs to pick up the phone today. That is unglamorous work, and it is exactly the work that decides whether your quality scores and your readmission rate move. Point AI there and you get outcomes and margin. Point it at the note and you get a nicer note.
So here is Monday morning. Pull your AI budget and draw one line down the middle. On the left, everything aimed at the visit itself, the scribes, the note assist, the coding help. On the right, everything aimed at the thirty days after the visit, the outreach, the enrollment, the monitoring, the follow-up. If the left column is your whole budget and the right column is zero, you have automated the end of the problem that was already working and left the part that actually kills your patients running on a spreadsheet and three overwhelmed nurses. Then ask your quality lead one question. Name a single quality measure that moved because of the scribe. If nobody can, you have your answer about where the next dollar goes.
The vendors will keep selling you speed inside the visit, because it demos well and it is easy to buy. Let them. The note was never the patient. Stop paying like it is.
Sources: Nature Medicine randomized trial of an AI clinical decision support tool in primary care, reported by Healthgrades, July 8, 2026 (https://resources.healthgrades.com/pro/ai-improved-documentation-patient-outcomes-primary-care-trial; underlying study https://www.nature.com/articles/s41591-026-04503-6). Becker's Hospital Review, "Ambient AI scales across health systems as burnout data emerges," citing the Peterson Health Technology Institute report, the AJMC adoption study, and the JAMA Network Open Mass General Brigham / Emory burnout study (https://www.beckershospitalreview.com/healthcare-information-technology/ai/ambient-ai-scales-across-health-systems-as-burnout-data-emerges/). Burnout percentages should be re-verified against the primary JAMA Network Open article before external publication.
