NHS England is rolling out Microsoft 365 Copilot to roughly 500,000 clinicians and support staff, making it one of the largest workplace AI deployments in the public sector. The governance picture around it is alarming: 68% of physicians say the NHS lacks the digital infrastructure to deploy AI effectively, and a ShareGate survey of IT and security leaders found that 93% felt confident in their Microsoft 365 governance frameworks while 29% had already experienced AI-surfaced data exposure incidents<\/a> involving contracts, employee records, and strategic plans.<\/p>\n The NHS rollout is a stress test that every large enterprise should read as a mirror. The core failure mode isn’t a misconfigured firewall or a breached perimeter. It’s permission models designed for human search behavior that were never built to withstand prompt-based retrieval. An employee who technically has read access to a folder of legacy HR records poses minimal risk when finding those records requires knowing they exist. Copilot removes that friction entirely.<\/p>\n The 93\/29 confidence gap in that ShareGate survey is the most important number in this piece. Nearly a third of organizations that believed they were governance-ready had already experienced exposure. That’s not a minor calibration error. It means most enterprise AI governance programs are measuring the wrong thing: they’re auditing policies on paper rather than testing what a well-crafted prompt actually returns. Confident governance frameworks built on stale permission inventories are not governance. They’re liability with good documentation.<\/p>\n The signal worth watching: how quickly the NHS incident rate (if disclosed) forces Microsoft to accelerate its Purview and permission-scoping tooling. If a 500,000-seat public sector deployment produces visible patient data exposure, the regulatory and reputational blowback will reshape enterprise Copilot procurement requirements globally, not just in UK healthcare. I’d revise this view if NHS England publishes a credible pre-deployment data access audit, but there’s no evidence one exists at the required depth.<\/p>\n Permission sprawl is the accumulation of data access rights that were granted for legitimate reasons but were never revoked when roles changed, projects ended, or teams restructured. It exists because provisioning access is fast and revoking it is slow, unglamorous work. Think of it as a building where every employee still holds a keycard for every office they’ve ever worked in. For human workers, sprawl is a latent risk. For an AI assistant querying the entire corpus on their behalf, it becomes an active retrieval surface. Every Copilot deployment inherits whatever mess already exists.<\/p>\nWhat this means for your business<\/h2>\n
Concept deep-dive: Permission sprawl<\/h2>\n