I build AI systems for a living, so opening with a list of nevers might seem like bad salesmanship. It’s the opposite. The practices that get burned by AI aren’t the ones that used too little of it — they’re the ones that let it wander into rooms it had no business entering. Knowing where the line sits is most of what “using AI well” means.
Here’s where I draw it, in my own builds and in any tool I’d recommend.
Clinical judgment. Ever.
No AI system should diagnose, triage, or suggest treatment to a patient — not in an email, not in a chat widget, not “just as a first pass.” The liability is obvious; the deeper problem is that clinical judgment is the entire reason your patients chose a licensed human. A patient FAQ written in plain language is a fine job for AI drafting. “Should I be worried about this pain?” is not.
Bad news.
A declined claim, a missed refund, a schedule change that ruins someone’s week, a condolence. Hard things deserve a human hand, every time. Patients forgive imperfect wording from a person; they do not forgive polished wording from a machine when it mattered. If a message carries weight, a person writes it — or at minimum rewrites it until it’s genuinely theirs.
Anything a human hasn’t read.
This is the approval-first rule, and it’s the spine of everything I build: AI drafts, your staff reviews, a person approves the send. Every message, no exceptions. The draft saves the time; the review keeps the judgment, catches the error, and keeps a human accountable for what leaves the building with your name on it. A system without an approval step isn’t automation — it’s abdication.
AI does the drafting. The judgment about what’s good enough stays human.
Patient data in consumer tools.
The free chatbot in your browser is not a HIPAA-covered environment, and pasting a patient’s name, condition, or complaint into it can be a reportable problem — even with good intentions, even “just this once” to draft a reply. Business-grade tools with a business associate agreement exist for exactly this reason. The five-question version of this vetting is in its own guide.
The case for cutting your staff.
This one is my own line, and I state it on the front page: I don’t build systems designed to replace people. Not because it’s impossible — because it’s bad medicine for a practice. Your front desk carries relationships no system can hold. The right use of AI is taking the tedious drafting off your people so they have more attention for the humans in the room. If a system can’t make someone’s day lighter, it shouldn’t get built.
What’s left is plenty
Review replies, recall drafts, reactivation emails, intake summaries, FAQs, welcome sequences — the unglamorous, repetitive writing that eats your evenings. That’s the territory where AI genuinely earns its keep, precisely because every draft still passes through human hands on its way out.
Think I’ve drawn a line in the wrong place? Tell me — I’d rather be corrected than comfortable. Or head back to the Classroom.