Two years ago, Inova Health System reviewed about one AI feature a month. Now it reviews more than twelve, has 92 in production, and is building oversight for as many as 500. That growth says less about one health system's ambition than about where AI in healthcare actually comes from. Most of it doesn't get built. It shows up. Keeping a health system AI inventory is turning into basic operating work, the same way tracking devices and software licenses already is.
Becker's reported this month that Inova's proposals come from three places: new capabilities from technology vendors it already uses, ideas from employees, and partnerships with model companies. Governance and post-deployment monitoring are centralized. Implementation is spread across the organization so review doesn't become a bottleneck. Chief data and AI officer Jon McManus put the tooling point bluntly: health systems need to "graduate from spreadsheets."
Take, shape or make
The more useful idea is how Inova sorts that work. In an August interview with HealthTech, McManus described three categories. "Take" is a vendor product, often an FDA-cleared device or decision support tool, that you implement as is and oversee. "Shape" is a vendor feature you configure to fit your workflow, usually through low-code settings. "Make" is something you build yourself. By his count about 10% of Inova's inventory is take, about 80% is shape, and about 10% is make.
That 80% matters. A feature that arrives in an EHR update or a vendor release still changes how people work and still needs someone to own it. It's easy to miss because nobody filed a project request for it.
Why a health system AI inventory needs people, not just a list
A list only helps if the tools on it get used well. UCHealth told Healthcare IT News that nurse adoption of its predictive fall model has been slow, two and a half years in. The work has been teaching staff how the model is trained, what its outputs mean, where it gets outliers wrong, and where a person stays in the loop.
For any care organization, large or small, the practical version is short. Write down every AI feature that's switched on, including the ones a vendor added. Note whether you took it, shaped it or made it, who owns it, and how you'll know if it stops working. Then budget time for training, because that is usually where adoption stalls. It's not glamorous, but it's what keeps human oversight real once the tools start adding up.