Before discussing an AI-assisted feature, define the task it is meant to support. Drafting an administrative summary, preparing a record for review and making a clinical decision are different activities with different consequences.
For an assistive workflow, specify what information the reviewer can inspect, how an error is corrected and what prevents an unreviewed draft from becoming an accepted record. Include incomplete inputs, conflicting information and uncertain outputs in the design conversation.
The organization should agree evaluation criteria, privacy requirements and acceptable use with appropriate clinical, operational and technical specialists. A pilot needs a defined scope and a way to stop or revise the feature if it does not meet those criteria.
Salt’s problem-first approach treats the model as one possible component. The question is whether the complete workflow is useful and accountable. AI availability alone is not a reason to build.
This is an engineering perspective. It does not establish clinical effectiveness or regulatory compliance for any particular tool.
Return to the library