Your board just asked you the question. "What's your AI strategy?"

Four words, and suddenly you're expected to have a point of view on a category you haven't had time to evaluate, delivered with the same confidence as your answer on payer mix or provider recruiting.

I never got that exact question, because AI wasn't the category twenty-five years ago. But I got versions of it constantly. What's your EMR strategy. What's your telehealth strategy. What's your patient portal strategy. Different technology, same position: a board wants a confident answer, and you're the one who has to sit across the table from vendors afterward and figure out if their pitch deck matches reality. That part hasn't changed. So this is the letter I wish someone had sent me back then, translated to the question you're actually facing now.

You're not behind. You're being careful.

If you've put off evaluating AI for your practice, that's not a failure of vision. It's usually the sign of someone who has been burned before. You've sat through the demo that looked flawless and then watched the actual rollout require six months of your team's time. You've seen "this will make your front desk's job easier" translate into more training, more tickets, and less trust in the system.

Skepticism, in this context, is a rational response to experience. The goal of this letter isn't to talk you out of it. It's to help you evaluate it the way an operator would, not the way a vendor wants you to.

The three worries that are actually legitimate

Patient safety. If a system is going to touch scheduling, referrals, or any patient-facing workflow, the real question isn't whether it can perform the task. It's what happens when it can't. Every system has an edge case: an unusual name, an ambiguous request, a data mismatch between what the patient says and what's in the EHR. The question worth asking is not "does this work," but "what happens the moment it doesn't."

Staff anxiety. Your front desk team has heard "this will free you up for higher value work" before, and it often meant something closer to "we're changing your job without asking you." Staff resistance to AI usually isn't resistance to technology. It's resistance to being handed a system nobody explained to them, built by people who never worked a phone queue during flu season.

Implementation burden. You don't have a spare project manager and six unstructured months. If a vendor's rollout plan depends on either, that alone tells you something about how much thought went into the product.

A framework for evaluating vendors, not pitch decks

When you're in the room with a vendor, here's what I'd ask, drawn from twenty five years of being the one buying these systems, not selling them.

Ask for the failure rate, not just the success rate. Every AI system has one. A vendor who can tell you precisely what percentage of interactions require human intervention, and what happens in that percentage, understands their own product. A vendor who only talks about the success cases hasn't looked closely enough at their own data, or doesn't want you to.

Ask how many staff hours go-live actually requires, in hours, not slides. "Minimal disruption" is marketing language until it's backed by a number. Ask what the practice's team actually did during onboarding, week by week.

Ask how the vendor gets paid. This tells you more than almost anything else in the conversation. A vendor paid for activity, calls placed, messages sent, has no structural reason to make sure the outcome actually happens. A vendor paid when the appointment is booked has the same incentive you do. That alignment shows up in how the product behaves, not just how it's priced.

Ask what breaks, specifically. Names that don't match phonetically. Multi-location scheduling rules that conflict. Insurance verification edge cases. If a vendor answers in generalities, that's the tell. Specificity is what separates a team that has actually built and deployed the system from a team that has built a demo.

The real red flag

It isn't confidence. Every vendor in this space will be confident, because confidence closes deals. The red flag is the absence of specifics underneath the confidence. If someone can't tell you, in detail, what fails and how they catch it, they haven't actually operated the system at scale. They've built something that works in a demo, which is a different achievement than something that works at two in the morning when the phone rings and nobody's watching.

You don't need to become an AI expert

You need someone who has sat in your chair asking better questions than the pitch deck answers. That's the whole point of this letter. Not to convince you AI is ready for your practice, and not to convince you it isn't. Just to make sure that whoever you evaluate next has to earn your trust with specifics, the same way you'd expect from any operator you'd actually want running part of your business.