The Medicines and Healthcare products Regulatory Agency has opened expressions of interest for a London regulatory sandbox intended to test artificial intelligence medical devices in clinical settings.
The programme will bring manufacturers together with NHS provider organisations so selected systems can be examined as they move from development environments into operational healthcare. Up to ten projects are expected to receive regulatory support, access to clinical settings, and structured evidence gathering.
Applicants will need to show how their technology could deliver a patient or service benefit while addressing safety, effectiveness, data governance, and regulatory compliance. The sandbox is designed to answer questions that cannot be resolved fully through laboratory validation, retrospective datasets, or tightly controlled trials.
Eligible technologies may include diagnostic support systems, imaging tools, risk prediction models, workflow software, monitoring platforms, and systems that help clinicians interpret complex information. Their regulatory position depends on the product’s intended purpose, the risk created by its output, and the extent to which clinical decisions rely on it.
Moving artificial intelligence systems into routine care exposes variations that are often absent during development. Data quality differs between hospitals, patient populations may not resemble training datasets, clinical workflows impose time constraints, and users can interpret the same output in different ways.
By placing selected products in supervised clinical environments, the programme will allow manufacturers to develop evidence around performance and usability while NHS organisations assess integration, operational burden, information governance, training, and the conditions required for safe deployment.
The initiative builds on the MHRA’s AI Airlock programme, which established a supervised environment for addressing uncertainty around artificial intelligence medical devices. The agency has received £3.6 million over three years to expand that work and make regulatory sandboxes a more regular part of product development.
Earlier engagement can reveal evidence gaps before they become barriers to approval or procurement. A technically capable system can still face delays when its clinical evaluation, quality documentation, cyber security controls, human factors work, or monitoring plan does not align with the way the product will be used.
Those disciplines become more demanding as manufacturers enter several markets and continue to update software after launch. The relationship between design controls, supplier management, configuration, verification, and regulatory records is central to resilient medical device manufacturing, particularly when hardware and software revisions proceed on different schedules.
Artificial intelligence adds complexity because performance may depend on model behaviour and data quality rather than a fixed mechanical or electrical characteristic. Developers must define the approved configuration, manage updates, control access to training data, document known limitations, and determine when a change requires further assessment.
Clinical testing can expose performance differences between demographic groups, hospitals, imaging systems, and local practices. It can also identify automation bias, where users place excessive confidence in a generated recommendation, and alert fatigue, where frequent low value warnings reduce attention to important outputs.
Manufacturing readiness remains part of the assessment even for products delivered mainly through software. Repeatable release processes, configuration control, cyber security maintenance, complaint handling, incident reporting, and traceability all form part of the operating system around the device.
Products that incorporate sensors, imaging equipment, or dedicated hardware add component sourcing, calibration, production test, servicing, and obsolescence management. A software update may also affect the performance of physical equipment already installed across several hospitals.
NHS participation should keep the selected projects within realistic operating constraints. Hospitals need technology that works with existing estates, networks, staffing models, procurement processes, and clinical responsibilities, rather than systems that depend on continuous support from their developers.
Regulation, procurement, and implementation rarely move at the same pace. A device may meet safety and performance requirements but still struggle to secure adoption when integration costs are unclear, benefits are difficult to measure, or responsibility for monitoring has not been assigned.
Structured testing can resolve those questions before deployment expands across multiple sites, while evidence gathered from several projects should also reveal recurring obstacles that future guidance can address more consistently.
Expressions of interest are now open to eligible manufacturers and London NHS providers. The selected projects will test whether supervised clinical development can shorten the route from promising software to a product that clinicians can use, manufacturers can support, and regulators can oversee throughout its operating life.
