Start with a precise task
A system that summarises documents, a model that estimates risk and software that recommends an action have different evidence and regulatory needs. Broad claims about ‘AI for medicine’ hide these differences.
The intended user, setting, input data and action triggered by the output should be explicit.
Evidence before scale
Technical accuracy alone is insufficient. Developers need representative data, external validation, human-factors testing and an evaluation of clinical benefit and harm in the intended workflow.
Performance may drift as populations, documentation or clinical practice change, so post-deployment monitoring matters.
Responsible use
Access controls, privacy, cybersecurity, auditability and incident response should be designed with the clinical workflow. A qualified person must be able to review, challenge and override an output.
Patient communication should state that the tool supports—not replaces—professional assessment.
Quick answers
Frequently asked questions
Can AI diagnose cancer from uploaded documents?
A general-purpose interface should not be treated as a diagnostic service. Validated tools have a defined intended use and remain part of a qualified clinical workflow.
What should an IT company validate first?
The clinical problem, intended user, workflow, data quality, failure modes and evidence needed for the stated claim.
Evidence base
Primary and authoritative sources
Use the linked source documents for professional or current interpretation.