AI & Digital Health · 4 min read
From Data to Decision: Building AI Tools That Clinicians Actually Use
The graveyard of clinical AI is full of technically excellent tools that clinicians never used. Not because the tools failed — but because the design process failed to produce something that fit the way clinicians actually work, think, and make decisions. Building AI that clinicians use is primarily a design challenge, not a technical one.
The clinical AI adoption problem is not primarily a technology problem. The algorithms are increasingly capable. The datasets are increasingly available. The computing infrastructure is increasingly accessible. What continues to fail — repeatedly, predictably, across institutions and specialties — is the translation of technical capability into clinical utility. The tool that performs impressively in a validation study and disappears from clinical use within three months of deployment has not failed because the AI was wrong. It has failed because the implementation was wrong — because the design process that should have produced a tool that fits clinical practice produced one that fit a researcher's model of clinical practice instead.
The design principles that produce tools clinicians use
Workflow integration is the primary predictor of clinical AI adoption. A tool that requires clinicians to interrupt their workflow to interact with it — to log into a separate system, to enter data that already exists elsewhere, to perform an action that is not part of their existing process — will be used when the motivation to use it is high and abandoned when it is not. The design principle is uncompromising: the tool must fit the workflow, or the workflow cannot be changed until the tool has demonstrated sufficient value to justify the change. Design for the workflow as it is. Demonstrate value. Then negotiate workflow change from a position of demonstrated utility.
Transparency at the point of use — presenting the AI's output in a way that makes its basis comprehensible to the clinician without requiring them to understand the underlying algorithm — is the second design imperative. "This patient's risk score is 0.78" is less useful than "This patient's elevated risk is primarily driven by their age, pre-operative pain score, and BMI — factors consistent with the clinical picture you have already identified." The first is a number. The second is a clinically contextualised insight that a clinician can evaluate, discuss with a patient, and act on with confidence.
Co-design as the implementation strategy that works
The clinical AI tools with the highest sustained adoption rates share a design origin: they were built with clinicians, not for them. Co-design — involving clinical end-users in requirement definition, prototype testing, workflow mapping, and iterative refinement — produces tools that fit clinical reality because they were shaped by people who inhabit it. The technologist who spends six months in clinical environments before beginning to design a clinical tool is not wasting time. They are investing in the implementation success that validation studies alone cannot produce.
The most technically sophisticated AI tool that no clinician uses has zero clinical impact. The simplest tool that every clinician uses consistently has transformative potential. Design for use, not for validation.
💬 What clinical AI tool in your practice has achieved the highest sustained adoption — and what design or implementation feature do you believe most explains its success?