Why Are Health Systems Building Their Own AI Instead Of Buying It?

Years ago, in my DTC days, I watched a Stripe dashboard cross a million dollars in a single day. I remember the feeling in my chest, that mix of pride and vertigo you get when a number gets bigger than your ability to understand it. I also remember, a few months later, realizing the product underneath that dashboard wasn't very good. Scale doesn't fix a bad product. It just hides it for a while, at higher volume, until it can't hide it anymore.

I think about that dashboard every time I read about a health system racing to launch its own AI chatbot.

Why are health systems building their own AI instead of buying it?

Because they're watching patients turn to AI with or without them, and they'd rather own that relationship than lose it. TechTarget's reporting on this trend lays it out plainly: hospitals are building proprietary AI products specifically so patients ask their questions inside a system the hospital controls, not inside a general chatbot with no connection to the patient's actual chart.

That instinct is correct. I just don't think most health systems need to build the whole stack from scratch to act on it.

What are patients actually using AI chatbots for?

Mostly the same three things they'd ask a nurse line if they could get through. Medication questions. Symptom triage. "Is this normal after my procedure." According to KFF's tracking poll, the share of adults using AI monthly for health information jumped from 17% two years ago to 29% in 2026. That's not a niche behavior anymore. That's close to a third of patients, going somewhere for health answers, and a real question about whether that somewhere is safe, accurate, or connected to their actual care team.

Should health systems build their own AI or use a vendor?

Honestly, it depends what "vendor" means to you. If it means a black box hosted by someone else, running on a model you don't control, with your patients' conversations sitting on servers you can't audit, then yes, build your own, or at least demand something that isn't that.

But if the choice is "build every layer ourselves" versus "own an open, self-hosted system that we can actually inspect," the second option gets you the control without the multi-year build. HANA runs open-source and self-hosted, with no dependency on OpenAI or any single external model provider. You're not locked into someone else's roadmap, someone else's pricing, or someone else's decision about what your patients' data is used for.

What happens when health systems get this wrong?

The same thing that happens to any bad product that gets more traffic. It just fails louder. A chatbot that gives a patient the wrong answer about their medication doesn't stay a small problem. It becomes a headline, a compliance review, and a very uncomfortable meeting with legal. I've sat in versions of that meeting, for less serious reasons than patient safety, and I can tell you the feeling doesn't improve with scale. It gets worse.

This is why the "build it ourselves" instinct needs to be paired with actual clinical rigor, not just engineering ambition. Across more than a million patient interactions, we've run at zero critical adverse events. That number matters more to me than almost anything else in the business, because it's the one number that, if it ever moves in the wrong direction, means we failed at the only job that actually counts.

How do you keep an AI health tool accountable to patients, not just to the roadmap?

You build it the way you'd want to be managed yourself. My daughter is ten. A while back she told me, matter-of-fact, "I work for you, not the other way around," which is a wildly good sentence for a ten year old and also, I think, the correct model for how any patient-facing system should relate to the people it serves. The system exists to serve the patient in front of it, not the quarterly roadmap, not the procurement cycle, not the vendor's growth targets.

Practically, that means the AI has to be auditable. It means case studies that show real outcomes, not just demo reels. It means clinicians can see exactly what the system said to a patient and why, and can correct it. Health systems building in-house get this instinct right. They just don't always have the years it takes to build the auditability layer from zero.

What should a health system actually look for in a patient AI platform?

Three things, in order. First, can you see inside it. Not a vendor's word for what it does, actual visibility into what it says to your patients and why. Second, can you run it without being permanently tied to one AI provider's pricing and policies. Third, does it have a real clinical track record, not just a demo. If a platform can't answer all three honestly, it's not ready for your patients regardless of how good the pitch deck looks.

Key Takeaways

Health systems are right to want to own their AI relationship with patients instead of ceding it to a general chatbot with no clinical context. Where it goes wrong is assuming ownership requires building every layer from scratch, on an internal timeline measured in years, with a black-box model underneath it anyway. An open-source, self-hosted platform gets you the control and the auditability without the multi-year build, and without trading one dependency for another. The instinct to own the relationship is correct. The execution just needs to match the stakes.

FAQ

Why are health systems building proprietary AI chatbots instead of using third-party tools?

They want to control the patient relationship and keep conversations connected to the patient's actual clinical record, rather than routing patients to a general AI tool with no access to their chart or care team.

Is patient use of AI chatbots for health questions actually growing?

Yes. KFF's polling shows the share of adults using AI monthly for health information nearly doubled, from 17% to 29%, between 2024 and 2026.

What should a health system prioritize when evaluating a patient-facing AI platform?

Transparency into what the system says and why, independence from a single external model provider, and a real clinical track record with measured outcomes. If you want to see what that looks like in practice, book a discovery call.