I let an AI draft my notes. Here's the line I won't let it cross.

I build tools. I'm also the person legally and morally on the hook for what's in the chart. Those two facts have to coexist, and the way they coexist is a line I drew on purpose.

What I happily hand over

• Transcription and structure. Turning a messy conversation into a clean, formatted draft. This is typing, and typing is not medicine.
• Surfacing what I might've missed. "You didn't document the medication reconciliation." Great. That's a checklist with better manners.
• Drafting the boring 80%. The review of systems, the templated counseling language, the formatting Veradigm wants.

What stays mine, always

• The assessment. The synthesis — this is what I think is going on and why — is the one thing that is irreducibly the clinician's. An AI can suggest. It cannot be accountable.
• Anything that asserts a fact about the patient I didn't verify. A draft that hallucinates a normal exam I didn't perform isn't a time-saver; it's a liability with good grammar.
• The decision to trust it. Every generated line gets read like a medical student's note: assume it's wrong until I've confirmed it's right.

A clinical AI should reduce the cost of being thorough. The moment it reduces the cost of being wrong, you've built the wrong thing.

Why I'm telling you this

Because the scaled version of this — the patient-facing check-in system I keep hinting at — lives or dies on exactly this discipline. It's going on its own domain, with its own security posture, precisely because PHI and "move fast" are a sentence you should never finish. The landing page can be playful. The thing that touches patients cannot.