Ask two dermatology practices how to book a new patient and you’ll get two different answers. Ask two providers inside the same practice and you might get two more.

We maintain 147 custom scheduling rule sets in production. When I mention that number, the polite question underneath is usually: why? Scheduling is scheduling. Find a slot, book the patient. Shouldn’t one good system cover everyone?

It’s a fair question, and the answer is the most important thing I’ve learned building AI for healthcare operations.

What’s actually in a rule set

Start with visit types. One EHR we integrated with had 553 distinct visit types configured. The AI’s first job on any call is mapping what a patient says (“I have a weird mole I want someone to look at”) to the one visit type that books correctly, at the right slot length, with the right prep instructions. Get that mapping wrong and everything downstream is wrong.

Then provider panels. Not every provider sees new patients. Not every provider takes every insurance plan, and credentialing changes quarter to quarter. Some providers only perform certain procedures at certain locations. A cosmetic consult might be 30 minutes with one provider and 45 with another, because that’s how they practice.

Then template logic. A Mohs surgeon’s schedule protects surgical blocks that a routine skin check must never touch. A provider might hold Friday afternoons for follow-ups only. These rules exist for clinical and economic reasons, and violating them produces an appointment the practice has to unwind by hand.

And then the layer that took us longest to respect: the rules nobody wrote down. Every practice has a coordinator who just knows that a particular laser treatment goes on one provider’s template and not another’s. That knowledge lives in sticky notes and fifteen years of pattern recognition. It’s load-bearing, and it isn’t in the EHR.

Why you can’t standardize your way out

The tempting engineering move is to treat all this variation as mess to be cleaned up. Build one canonical scheduling model, ask practices to conform to it.

Some cleanup genuinely helps. One operations leader we work with cut those 553 visit types down to 40 before go-live, and that consolidation made everything after it faster. Practices carry configuration debt like anyone else, and go-live is a good forcing function to pay some down.

But past a point, the variation isn’t debt. It’s the practice. The rules encode how a specific group of providers has decided to run their clinical and financial operation. A rule that looks arbitrary from the outside usually turns out to protect a surgical block or a payer contract. Flatten the rules and you haven’t simplified the practice. You’ve just made the AI wrong about it.

This is where one-size-fits-all voice AI breaks in production. The demo never tests the rules. A demo books “an appointment,” and it looks great. Production is nothing but the rules. An AI that books the wrong visit type, the wrong slot length, or a provider who doesn’t take that plan generates work instead of removing it. Each individual error looks small. At thousands of calls a month, the front office spends its day cleaning up after the system, which is worse than not having the system.

Customization as a competency, not a chore

So we made a decision early to engineer rule set configuration with the same seriousness most companies reserve for their flagship features.

That means discovery built around extracting rules, including the undocumented ones, from the people who hold them. It means encoding rule sets in a form we can test against a practice’s real historical bookings before go-live, so we find the conflicts on our side of the launch. And it means treating every rule set as a living artifact, because practices don’t hold still. Providers join, panels shift, plans get dropped. A rule set that was right in March is quietly wrong by June unless someone owns keeping it right.

We’ve spent more engineering time on this layer than on the conversational AI itself. That ratio surprises people. It shouldn’t. Two weeks ago I wrote about the last mile, the work of turning a good conversation into a confirmed appointment in the EHR. The rule sets are what make that last mile worth driving. The write-back can execute perfectly and still fail the practice if it books the wrong thing correctly.

One hundred forty-seven rule sets was never a target. It’s just what the problem turned out to require once we stopped pretending practices were interchangeable. The AI that books the right appointment for a specific provider at a specific practice will always beat the one that books an appointment anywhere. The difference between those two is the rules.