Share with your CISO
The Race and Health Observatory is pushing to make ethnicity-disaggregated performance data a hard procurement requirement for NHS AI, not a post-deployment audit item. The paper, released 10 September, calls for vendors to supply accuracy figures broken down by ethnic group before contracts are signed, minimum performance thresholds below which a product can’t be purchased, mandatory patient disclosure when AI influenced a clinical decision, and a defined exit route for tools that cause harm. Nobody currently owns the decision to pull a harmful NHS AI tool, and the paper treats that gap as the central failure.
What this means for your business
Any vendor selling clinical AI into the NHS should treat this as a requirements shift, not a policy discussion. The paper has the MHRA, NHS England, and the DHSC in the room at the roundtable, which means this is closer to a regulatory signal than a think-piece. If you sell diagnostic, triage, or workflow AI into UK healthcare, the question isn’t whether ethnicity performance breakdowns become a tender requirement; it’s how long you have before they do, and whether your validation pipeline currently generates that data at all.
The accountability gap the paper identifies is the sharpest finding. Distributed responsibility, spread across manufacturers, providers, and regulators, is functionally the same as no responsibility, and regulators have a long history of resolving that ambiguity by pushing liability onto the vendor. The Observatory’s recommended fix, a pre-defined escalation and exit trigger attached to each deployed system, is the kind of thing that sounds procedural until it becomes the clause that terminates your contract. Organizations that wait for formal guidance before building disaggregated validation into their model development process will be retrofitting under deadline pressure.
The MHRA commission’s staged approval framework for clinical AI, published days before this paper, makes the two documents read as a coordinated architecture. Staged authorisation provides the on-ramp and off-ramp mechanism; disaggregated performance data provides the evidence standard that governs movement through those stages. That pairing tightens the compliance surface considerably. Vendors who assumed the MHRA’s staged model was the whole story now have a second dimension to meet. I’d revise this read only if the NHS procurement bodies decline to adopt the Observatory’s thresholds as formal tender criteria, which would leave the paper as influential but optional.
Based on reporting from NHS AI procurement should require ethnicity performance data, originally published 2026-09-14 06:19:00.
