MHRA widens AI Airlock as lifecycle oversight becomes the central regulatory test
The Medicines and Healthcare products Regulatory Agency is turning its AI Airlock from an exploratory pilot into a more ambitious regulatory instrument. Launched in spring 2024, the sandbox is designed to expose the hardest problems in regulating AI-enabled medical devices before those problems spread across the wider market. Phase 3, announced on 6 October 2026, shifts the emphasis toward post-market surveillance and broader oversight across the full device lifecycle.
That shift matters because AI-enabled devices do not behave like static products. Their performance can change as data inputs evolve, workflows shift, software is updated, and clinical settings differ from the environment in which they were first assessed. A conventional premarket review can miss those moving parts. The AI Airlock is built to surface them early, using direct testing with regulators, clinicians, industry partners, academics, and healthcare system stakeholders.
Phase 3 moves from concept testing to longer-term regulatory stress testing
The third phase is supported by multi-year funding from the Department for Health and Social Care, giving the programme room to move beyond short exploratory exercises. The MHRA says the next wave will center on more ambitious, longer-term testing models aimed at producing tangible outputs and advancing agency priorities. In regulatory terms, that is a meaningful escalation: it allows the sandbox to probe how AI devices behave after deployment, how updates should be tracked, and what evidence should be expected once software starts interacting with real clinical environments.
The call for applications is open to innovators who want to work through those questions. Eligible applicants include medical device manufacturers, academics, and developers of related innovative solutions. The programme is no longer framed as a narrow trial of a few technologies; it is becoming a mechanism for shaping how Great Britain regulates AI-enabled medical devices in practice.
A regulatory framework for adaptive software cannot rely on static assumptions.
Pilot and phase 2 cohorts tested a wider set of clinical use cases
The reporting history of the AI Airlock shows how quickly the programme expanded. The pilot phase closed in April 2025 and covered four products across a range of medical device regulatory issues. Phase 2 completed in May 2026 and involved seven innovators across three regulatory challenges. Together, the first two phases engaged 11 innovators across seven regulatory challenges.
The second phase widened the clinical span of the programme substantially. It included technologies for AI-powered clinical note taking, advanced cancer diagnostics, eye disease detection tools, and obesity treatment support systems. Those categories are not interchangeable from a regulatory standpoint. Diagnostic tools raise questions about analytical validity, clinical validity, and false reassurance. Documentation systems raise concerns about transcription accuracy, workflow drift, and the possibility that an error could propagate into the medical record. Obesity support tools, by contrast, move closer to longitudinal behavior change and ongoing clinical oversight, where performance may depend on patient engagement and iterative model behavior.
By bringing those use cases into one sandbox, the MHRA created a controlled setting in which different regulatory problems could be compared rather than treated as isolated one-offs. The programme focused on three core issues: managing evolving AI applications, regulating AI-powered diagnostics effectively, and implementing robust post-market surveillance for AI medical devices. Those are the pressure points that usually become visible only after a product is in routine use.
Simulation workshops turned regulatory theory into operational questions
The sandbox did not rely solely on written submissions or case study reports. Simulation workshops formed a central part of the testing process. These were roundtable sessions bringing together representatives from industry, clinical practice, regulation, academia, and technology, including experts from the MHRA, NICE, the NHS, and partner organisations. Their purpose was to stress-test regulatory scenarios from multiple angles at once.
That format is especially relevant for AI-enabled devices because evidence alone rarely settles the practical questions. A device may perform well in validation, yet still create uncertainty around human oversight, documentation standards, accountability when outputs change, and how to respond when performance degrades after deployment. Multi-stakeholder workshops help surface those tensions before they harden into policy gaps.
The workshop reports from the pilot are now publicly available. They are described as insights from the simulation sessions rather than formal MHRA guidance, but they still matter because they show how the agency is translating real-world product experience into regulatory design.
A broader regulatory framework is taking shape around real products, not abstractions
The AI Airlock was established to proactively investigate the regulatory challenges of AI-enabled medical devices. It is not positioned as a substitute for formal regulation; instead, it functions as an evidence-generating environment that can inform future guidance. The outputs are intended to support a fit-for-purpose framework for Great Britain, aligned with recommendations from the National Commission into the regulation of AI in healthcare.
This alignment is important. AI regulation in health care fails when policy is built too far from the point of use. Devices behave differently in emergency medicine, outpatient care, radiology, ophthalmology, and chronic disease management. They also behave differently depending on the maturity of the underlying product and the robustness of the developer’s quality systems. The Airlock’s value lies in forcing those differences into the open, where they can be examined under regulatory supervision rather than discovered after market entry.
What the program reveals about the next phase of AI regulation
The page’s update history shows a programme that has moved through several distinct stages: launch, pilot cohort selection, publication of reports, expansion into phase 2, and now phase 3 with a more explicit focus on lifecycle oversight. That progression suggests the MHRA is no longer treating AI devices as a niche category requiring ad hoc responses. It is building an operating model for recurring scrutiny, especially where algorithms continue to evolve after deployment.
For manufacturers and developers, the immediate implication is practical. Regulatory success will depend not only on premarket performance but on how well a product can be monitored, updated, and explained once it is in use. Post-market surveillance is no longer a back-end administrative task; for AI-enabled devices, it becomes part of the core evidence package.
The remaining challenge is whether this sandbox can be scaled without blunting the specificity that makes it useful. Phase 3 is designed to answer that by testing more ambitious models and generating tangible outputs, but the public reports available so far are still described as non-binding and not formal guidance. The next decisive question is how quickly those findings can be converted into durable regulatory expectations for the devices already moving into clinical practice.