Recently, a hospital worker at a Somerset NHS trust accessed up to 200 patient records without permission and shared a screenshot of one patient's medical information with their partner, according to documents obtained by the BBC. A person with legitimate access and no one examining how it was used.

The pressure to adopt AI is real and understandable. Tight budgets, ageing infrastructure and staffing shortages make any tool that reduces administrative load or speeds access to information hard to resist. But when the pressure to deploy outpaces governance, AI extends access to sensitive data and critical systems without anyone having decided that it should. Poorly configured or introduced without approval, they put patient data and critical systems at risk.

The first priority is visibility. You can't govern what you can't see. In healthcare terms, think of it as the X-ray or MRI for AI security: you can't treat what you haven't scanned. Providers need to know which AI tools are in use, what each can reach and who is accountable. The question for senior leaders is whether the organisation can show clear ownership, oversight and governance when AI touches patient care.

Shadow AI is a decision nobody sees

One of the biggest obstacles to visibility is shadow AI: tools and capabilities operating outside formal procurement, approval or review. In healthcare, that often looks like a member of staff using a publicly available AI tool to clear a backlog or speed up an administrative task.

Mimecast research found unsanctioned AI tools active within 98% of the organisations surveyed, while only 18.5% of employees said they were aware of their organisation's AI-use policy. Adoption also travels between home and work, in both directions. Pasting personal details into a chatbot can become pasting patient details. Personal data hygiene is a patient safety measure.

Those tools may access patient information, internal communications, workforce data and applications that support care delivery. Every tool outside your monitoring is also an easier route in for a threat actor. Disruption to clinical workflows becomes a patient safety issue.

A complex environment makes it harder

Healthcare providers run on interconnected systems, clinical applications, patient records, communications and workforce management. That makes it difficult to see where an AI capability is operating, what data it can reach and who owns it.

AI agents may now operate across those systems. Each connection is another route for exposure. That means approving an application once is no longer enough. Platforms reviewed 12 months ago may now include new AI functions, integrations or data access.

Why this is a patient care issue

Healthcare already knows what happens when technology fails. WannaCry showed the immediate cost of losing access to critical systems. The more recent Synnovis incident showed how disruption at a single third-party provider can spread across multiple NHS organisations, interrupting services and affecting patients.

Technology governance can't be treated as a back-office exercise in healthcare. Policies only reduce risk when staff understand them, systems are monitored and accountability is clear.

Accountability stays human

Healthcare already knows how to triage. An AI scheduling tool carries different risk from a platform that accesses patient records or supports clinical decisions. The closer a system sits to patient care, the more scrutiny it needs.

That focus must include the people behind the technology. AI agents are deployed by people, connected to human identities and often inherit their permissions. An organisation that governs only the tool has addressed half the problem.

People decide whether to deploy it, what it can access and how its outputs are used. Responsibility must be assigned across the teams that approve, configure, operate and oversee it. ‘The AI did it' is not a finding.

Apply what you already have, continuously

Healthcare doesn't need to invent a new discipline. Clinical risk management, supplier assurance and the Data Security and Protection Toolkit (DSPT) already ask what can this system reach, what does it connect to and what happens if it fails? AI can change faster than review cycles so those checks must run continuously.

Start by identifying what's already running, including AI features added to previously approved systems. Give each capability a named owner. Include AI into existing training and incident reporting. Treat staff as part of the control, not the problem as fear of blame discourages disclosure.

None of this is a one-off exercise. Approval is a snapshot and an inventory that is a quarter out of date may already be unreliable.

The questions for the next executive meeting should be simple and direct. For every AI capability that touches patient data or care delivery, leaders should be able to name the accountable person, which systems and data each tool can access. Clear answers to these questions turn AI governance from a policy exercise into meaningful protection for patient care.