Healthcare systems are at an inflection point that I have watched develop over the past several years. The architecture that served hospitals for decades, monolithic EMR systems, siloed data stores, manual clinical processes, is meeting AI capabilities that demand something different.
The data problem is the first obstacle. AI in healthcare is only as good as the data it can access, and healthcare data is fragmented, inconsistently structured, and protected by appropriate but complex regulations. FHIR (Fast Healthcare Interoperability Resources) has made progress on the structure and sharing problem. The HL7 standards ecosystem provides the plumbing. But the implementation across trusts and providers remains incomplete.
The clinical workflow problem is more subtle. AI tools that do not fit naturally into clinical workflows do not get used, regardless of their technical capability. Clinicians have limited time and high cognitive load. An AI that requires additional steps, additional logins, or additional mental switching is more likely to be avoided than adopted. The most successful healthcare AI implementations I have seen were built by teams who spent as much time understanding clinical workflow as they did building technology.
The regulatory and liability questions are genuinely difficult. When an AI system contributes to a clinical decision, who is responsible if that decision contributes to harm? The regulatory frameworks for AI medical devices (class IIb, UKCA marking in the UK, FDA clearance in the US) apply to defined medical device software but the boundaries are not always clear. Healthcare organisations have been cautious about AI-assisted clinical decisions for good reasons.
The areas where healthcare AI has delivered most clearly in practice: administrative automation (prior authorisation, coding, scheduling), clinical documentation reduction (ambient AI scribing, structured note generation), diagnostic image analysis where there is CE/FDA cleared software, and operational efficiency (bed management, staffing optimisation).
The areas where the gap between promise and delivery has been largest: clinical decision support that is not embedded in workflow, population health predictions that are not actionable, and diagnostic AI that was not evaluated on representative populations.
My view is that the healthcare AI story is genuinely important and will play out over decades rather than years. The potential to reduce the administrative burden on clinicians, improve diagnostic accuracy, and personalise treatment is real. The path to realising it requires rigorous evaluation, genuine clinical partnership, and the kind of patient infrastructure investment that is difficult but necessary.