When a clinical AI product is evaluated, the conversation usually starts with model performance and ends with data protection. Both matter. Neither explains the failure mode I have seen most often, which is that the tool does not survive contact with the working day.
The demo happens at a quiet desk with a compliant example. Tuesday afternoon happens with a queue at reception, two calls waiting, a district nurse on the other line and a clinician who has four minutes. A product that adds thirty seconds to that moment will be abandoned, and the abandonment will not be recorded anywhere as a failure of the technology.
Three patterns worth recognising
Extra steps get skipped. If using the system properly takes longer than the shortcut, staff will take the shortcut. This is not resistance to change. It is rational behaviour under time pressure, and it is the single most reliable predictor of a tool quietly falling out of use.
Parallel systems create divergence. When a new tool sits alongside the old process rather than replacing it, two records drift apart. Within weeks nobody can say which one is authoritative, and the governance value of both disappears.
Silent failure is worse than loud failure. A system that stops working and says so can be managed. A system that keeps returning plausible but incomplete output teaches the team to trust something that is not trustworthy.
What good implementation looks like from the clinical side
Watch real sessions rather than reviewing screens. Count the clicks in the busiest hour, not the average one. Define who the human in the loop is by name, not by role, and check that they are actually available at the times the system will need them. Decide in advance what the system must never do without a human, and write that down where the operational team can find it.
What buyers should ask suppliers
Who uses this, at what point in their day, and what does it cost them in time. What happens at 6pm on a Friday. What does the failure path look like in practice, and can I see the log of a real failure. Which of your claims has been evaluated, by whom, and against what baseline.
The uncomfortable conclusion
Adoption is not a commercial metric that sits beside clinical safety. It is part of it. A tool that is not used has no safety benefit at all, and a tool that is used in a workaround has an unknown one. Any evaluation framework that stops at the model and never reaches the workflow is measuring the wrong thing.
If you are building or buying clinical software and want an independent clinical read on workflow fit, that is work I do.