You don't unleash AI. You decide what it decides.

The future of AI in healthcare is not AI doing everything. It is AI doing what only AI can do, inside a system deliberately built to make every patient's experience predictable.

That is the whole vision, and it is worth stating plainly because so much of the industry is chasing a different one. The prevailing story says that as models get smarter, structure becomes obsolete: hand the AI a goal, get out of its way, and let intelligence handle the rest. We see it differently. The smarter AI becomes, the more valuable structure becomes, because structure is what turns raw intelligence into a product that thousands of patients can trust in exactly the same way. The organizations that win with AI will not be the ones that use it the most. They will be the ones that know precisely where it belongs.

What only AI can do

Anyone who has spent time around patient access knows that patients rarely call with tidy requests. A patient calls about a refill, mentions halfway through that she is leaving town for three weeks, and wonders out loud whether she should just come in instead. There is a real need in there, but it arrives tangled, and for decades our industry asked the patient to do the untangling. Software could only accept requests in the shape it expected, so we handed people phone trees and portals and made them translate their lives into our menus.

AI is the first technology that does the translating for us. It can take that winding request and find what the patient actually needs. It can look at the situation in front of it and recognize which of the clinic's rules apply, the way an experienced front desk coordinator would. And it can do all of this in a conversation that feels natural, moving at the patient's pace and in the patient's words. None of this was possible before, no matter how much software we wrote, because these are acts of judgment, and judgment could never be programmed. This is the work AI was made for, and in our products it is exactly the work AI is given.

What should never be left to chance

Judgment is only half of a great experience, though. The other half is made of decisions that should never vary from one patient to the next: which options a patient is offered and in what order, which steps a workflow follows, when a conversation moves to a member of the staff, and what the system will simply never do. If those decisions get made in the moment, conversation by conversation, the experience starts to drift, and slowly no two patients are walking through the same front door.

So we make those decisions the old-fashioned way, ahead of time, and we enter them into the system discretely as settings, defined lists, and explicit rules. If a particular insurance plan needs to be scheduled by a specialized team, the agent does not work that out from conversation. The plan comes straight from the medical record, and a simple rule decides what happens next, the same way every time. Decisions like these are deliberate rather than emergent. Someone chose them, anyone can review them, and they will produce the identical result for the thousandth patient that they produced for the first.

It would be easy to read this structure as a limitation on the AI, but it is closer to the opposite. The structure is what frees the intelligence to be excellent, in the same way a great clinician is freed by protocols she never has to second-guess. The AI spends its judgment where judgment is needed, and it leans on the frame everywhere else.

Two strengths, one experience

Medicine figured out this balance long before software did. Checklists exist because certain steps have to happen identically for every patient, without exception, and clinical judgment exists because no checklist has ever met the patient actually sitting in the room. Nobody in healthcare would ever ask which one great care comes from, because the answer is so obviously both.

We hold our products to that same standard. The AI takes the reins wherever the situation is ambiguous, wherever rules need weighing, wherever a person deserves to be heard like a person. The structure holds everything that has to be repeatable, from the paths a conversation can take to the promises the system makes along the way. The patient never sees the line between the two. She just has a conversation that feels human and lands exactly the way it should, whether she is the first patient to call or the ten thousandth. When people ask what makes an AI product great, that seam is the answer, far more than the strength of any model or the depth of any configuration.

Loosening the reins

The balance we strike today is not where this story ends. It is how the story gets to the good part.

What we are really building is a transition, and it works one setting at a time. A decision that lives in a discrete rule today becomes a candidate for AI judgment tomorrow, once the performance is proven and the trust is there. When it graduates, we loosen the reins by exactly that much, and no more. Then the next setting, and the next. Each step is small enough that patients and health systems can see it, feel the benefit of it, and grow comfortable with it together, at the same pace. Trust is not something we can ask either of them to extend in advance. It has to be earned in increments, and the discrete structure is what makes the increments possible.

This is why I am more optimistic about AI in healthcare than I have ever been. The predictable systems we are building right now are not a ceiling on what AI will become. They are the staircase it climbs, one earned step after another, toward a future where patients reach care through conversations that feel effortless and clinics know with certainty that every patient gets the experience they designed. The technology for that future already exists. What will separate the products that reach it from the products that never do is not how capable the models become.

Intelligence is no longer scarce. Knowing where it belongs is.

‍