I perform transcatheter aortic valve implantation (TAVI) procedures for a living. Nothing about my work in the cath lab has been taken over by artificial intelligence, and I don’t expect it to be anytime soon. But almost everything I do outside the procedure room (the documentation, the teaching preparation, the manuscript revisions, the patient handouts) has changed beyond recognition in the past two years.
I want to describe what that actually looks like, because most of what physicians hear about AI comes from two unhelpful extremes: breathless promises that AI will revolutionize medicine, and warnings that it will hallucinate us into malpractice. The day-to-day reality, at least in my practice, is quieter and more useful than either story.
Where the hours actually go
Like most academic physicians, my week contains two jobs. The visible one is clinical: procedures, outpatient clinic, rounds. The invisible one is text: discharge summaries, referral letters, letters of medical necessity, reviewer responses, lecture slides, exam questions, recommendation letters. Nobody went to medical school for the second job, yet it quietly consumes evenings and weekends.
That second job is where AI earns its place. A discharge summary drafted from my case notes takes minutes to verify instead of forty minutes to write. A response letter to a peer reviewer (the task that used to stall my manuscripts for weeks) now starts from a structured draft the same evening the review arrives. Patient education sheets can be rewritten at three reading levels, translated, and back-translated so I can check that no instruction shifted in translation.
Real physician voices, twice a week
Free, and one click to unsubscribe.
None of this involves AI making a single clinical decision. That distinction matters more than any other point in this essay.
The habit that makes it safe
Early on, I made two rules for myself, and they have survived two years of daily use. First, no patient-identifiable information ever goes into a general-purpose AI tool. Not names, not dates, not rare combinations of findings that could identify someone. This rule is absolute, and it shapes how I write every request: I work from de-identified notes and structured summaries.
Second (and this is the habit I wish more colleagues would adopt), I treat every output as a draft written by a bright student who has never met my patient and occasionally invents things with total confidence. The question is never “Is this good?” but “Where, specifically, does this kind of output go wrong?”
The answer differs by task, and learning those failure patterns is the real skill. Medication lists: AI finds interaction candidates well, but its missed-dose instructions are wrong often enough that I check them against the drug class every time. Guideline summaries: Always ask for the publication year, because the model may be summarizing the version before last. Statistics: A plausible-sounding analysis plan still needs a statistician’s sign-off, and I ask the AI to flag which decisions need one. Reviewer responses: The quoted manuscript text must match the actual revision word for word, so the quote gets updated after I finalize the edit, never before.
Verification sounds like it would erase the time savings. It doesn’t, for a simple reason: Checking a draft against ground truth is far faster than producing the draft. That asymmetry is the entire economic case for AI in clinical work, and it only holds if the physician actually does the checking.
What I tell my residents
When residents ask me how to start, I give them three pieces of advice. Start with the lowest-stakes text you write repeatedly. Not clinical reasoning: the referral letter, the journal club summary, the lecture outline. Learn the tool’s behavior where errors are cheap.
Write your requests the way you would brief a new rotating student: context, task, format, and what to flag for your review. The difference between a useless output and a useful one is almost always in the quality of the instruction, not the model.
And build the verification step into the prompt itself. I routinely end requests with instructions like “Flag any information that appears missing or inconsistent so I can verify it” or “If you are not certain, say so explicitly rather than guessing.” An AI that is asked to expose its own uncertainty is far safer than one asked to sound complete.
The stakes are generational
I teach medical students who will practice into the 2060s. They will not be competing against AI; they will be competing against (and collaborating with) physicians who use it well. If the physicians who care most about accuracy and patient safety refuse to engage with these tools, we cede the territory to those who care least. The safest possible version of medical AI is one shaped by clinicians who verify everything.
The technology will keep changing. The habit (verify, always, specifically) will not.
Yusuke Watanabe is an interventional cardiologist in Japan.
