You've seen the patient. Written the H&P. Worked through the assessment and plan. And yet — there's still that nagging feeling. Did I miss something? Is there a diagnosis on the differential I didn't think to consider? A piece of the workup that could change the picture?
Most of us have been using generative AI for the easy stuff — drafting prior auth letters, writing discharge summaries, cleaning up patient notes. Useful, but indirect. I've been doing something different over the past couple of months, and it's the first time I've felt like AI is actually strengthening my clinical reasoning in real time, at the bedside, on complex patients.
I’ve worked through a complex patient. The H&P is done. The assessment and plan is written. But there's still a nagging feeling that I might be missing something like a diagnosis on the differential I didn't think to consider, a piece of the workup that could change the picture.
Have you had this feeling before? This is exactly where I've found generative AI most useful (honestly, only up until a month or two ago!).
For some context, I like to classify patients into bread-and-butter presentations versus complex zebra cases. For those outside medicine, there's a saying: when you hear hooves, think horses, not zebras, because base rates make horses more likely. But in a dense city like New York City with 8 million people, you see plenty of zebras, especially at quaternary academic centers.
For bread-and-butter cases, I'm confident. AKI after a viral illness and poor PO intake? Give fluids. Heart failure exacerbation? Lasix. Community-acquired pneumonia? Antibiotics. When a zebra shows up, you usually have a sense of how to navigate it. The question is whether you're missing something (like a unicorn, or as my daughter calls them, “uni!”)
Here's what I've been doing. You can watch the recording below for a real-time example.
How I use AI to fact check my own clinical reasoning (real patient example)
When I have a complex patient and I'm not fully confident in my assessment and plan, I take the full H&P (HPI, labs, imaging, physical exam) and paste it into Doximity or OpenEvidence. Then I use this prompt:
You are a seasoned attending reviewing my assessment and plan based on the H&P. Please evaluate the assessment and plan. Critique how evidence-based it is, and critique the differential reasoning. Please note anything missing that would otherwise improve patient care or clarify the diagnosis. Please grade me at the end.
I've already completed the full H&P. I've already thought through the reasoning. I'm asking: what am I missing?
That distinction is important, since I strongly believe we physicians should reason first, and the AI should react second—not the other way around. This is especially important for trainees who are still building that reasoning muscle. These tools should be stress-testing our thinking, not providing shortcuts.
In the example, the AI surfaces giant cell arteritis—something I hadn't considered at all. If the patient has it, that diagnosis needs prompt treatment. I use this regularly in the inpatient setting, especially with patients with liver disease where the complexity can be overwhelming. It catches gaps, and it teaches.
Up to now, most of my AI use has been workflow-related — drafting prior auth letters, patient notes, discharge summaries. Those indirectly improve care. This is different. This directly strengthens clinical reasoning at the point of care.
Like pilots who rely on AI-assisted systems for navigation, weather prediction, and landing—tools that have likely reduced in-flight errors—we're at a point where it's hard to justify not using tools that help us think more clearly at the bedside. Atul Gawande made this case for checklists in The Checklist Manifesto. Generative AI is the next iteration of that same principle: use structured support to catch what the human mind, under pressure and fatigue, might miss.
