Every healthcare AI conversation I sit in eventually reaches the same place. Someone presents impressive model performance. Someone else asks how it will actually be used on a Tuesday afternoon in a stretched service, and the room goes quiet. That silence is the real state of healthcare AI: the models are increasingly good enough, and the operating questions are increasingly the constraint.
I say this as someone who wants AI in healthcare to succeed, and who works on the delivery side of national-scale digital health. The uncomfortable truth is that a benchmark result answers almost none of the questions that decide whether a tool survives contact with a real service.
The questions that actually decide adoption
Whose workflow does this live in, and what do they stop doing to accommodate it? Who is accountable when it is wrong, and do they know they are accountable? What happens to the output: does it create work, or remove it? Who maintains it when the vendor’s attention moves on, and who notices when its performance drifts? How does it behave for the patients who were underrepresented in its training data? And who pays, from which budget, against which of their existing pressures?
These are operating questions. They belong to service managers, clinical safety officers, information governance leads, procurement teams and the clinicians whose names sit next to decisions. A model can be excellent and still fail every one of them.
Why this is good news
Framing adoption as an operating problem is more hopeful than it sounds, because operating problems are solvable with known disciplines: workflow design, clinical safety casework, evidence generation matched to the buyer’s actual decision, honest total-cost economics, and deployment that starts narrow and earns its expansion. Teams that treat these as first-class product work, rather than friction to be endured after the science, are the ones whose tools end up quietly embedded in everyday care.
The practical implication for builders: your competitive advantage is probably not two points of benchmark performance. It is being the team that turns up already speaking the operating language, with a safety case, a workflow story, a realistic evidence plan and a deployment model a stretched service can actually absorb. The model gets you into the conversation. The operating work is what gets you adopted.