This is Practical AI for Physicians. You are here for a rule you can run before a tool touches a patient or a chart. Local models first. You keep the judgment.
I’m Travis Mendel. Radiation oncologist. Medical Director at Rio Grande Cancer Specialists and Rio Grande Urology in El Paso. I write this because I use the tools, and because most of the writing I see comes from people who do not sign the note.
This is not clinical decision support. It is not a vendor roundup, and it is not a case that AI will replace you.
You do not let a tool near a patient or a chart until you can answer four questions.
1. Who can see the data, and where does it go?
If audio, a note, or a chart leaves the office, you need a name for who receives it and a place for where it sits. A public chatbot is out. That is a third party you did not contract, on servers you do not control, with a retention story you cannot check. Local models are the default because the data stays on hardware you own. If a vendor has to see PHI, you want a BAA that covers this workflow: this product tier, these connectors, this retention. A BAA is a contract, not a privacy sticker. If you cannot name the destination, the tool stays off.
2. Has it been validated for the specific clinical task?
A demo is not validation. “Works for documentation” is not the same as validated for this note, in this specialty, in this EHR, against this template. If they claim ambient capture of a follow-up visit, you want evidence on that task. Not a staged new-patient video. Not a different specialty. Not an Epic integration you do not have. Until you have that, treat the output as untested draft language. If you cannot get even that, do not use it.
3. Can its outputs be audited and traced?
If you cannot show where a sentence came from, do not file it. The model drafts. You see the source: the transcript, the report, the policy line, the lab. A clean paragraph with no trail is a guess. When something is wrong in six months, you need to reconstruct what the tool saw and what you signed. No trail, no tool.
4. Does it reduce cognitive load rather than add to it?
A tool that drafts a note and then asks you to click through three extra screens has not helped. A tool that writes a message you still have to rewrite, then drops the rewrite back in your inbox, has added work. The test is whether the person who already owns the step is lighter after the draft. If it adds load, turn it off. Time back to the physician is the point. A second inbox is not.
Treat these tools like a good resident. Fluent, fast, sometimes confidently wrong, always supervised. Draft work you can inspect and correct is fine. Diagnosis, treatment, final consent, dose, and plan sign-off are not. Fluency does not transfer authority.
More to come!