MHRA AI Airlock (2024)

The AI Airlock is not a new regulatory regime. It is a sandbox: a controlled environment where the MHRA, an Approved Body and an AIaMD manufacturer co-develop the regulatory framing for a specific product whose features (adaptive learning, generative output, novel evidence base) fall outside settled guidance. Lessons learned feed back into future guidance and into the UK's wider Software and AI as a Medical Device Change Programme.

• AIaMD candidates with novel regulatory challenges — adaptive models, generative AI outputs, agentic workflows, foundation-model-derived products.
• Products with a clear intended UK deployment pathway in NHS England, NHS Scotland, NHS Wales or HSC Northern Ireland.
• Manufacturers with a credible Quality Management System (typically ISO 13485) and existing or planned UKCA marking strategy.

1. Intended purpose articulation — particularly hard for generative or agentic AIaMD where outputs are not bounded.
2. Change control for learning models — when does a model update become a 'significant change' triggering reassessment?
3. Real-world performance monitoring — what metrics, what cadence, what action thresholds.
4. Bias, equity and fairness — how to detect and mitigate population-subgroup performance drift.
5. Human-AI interaction — over-reliance, automation bias, alert fatigue, override logging.
6. Cybersecurity and model-integrity attacks — prompt injection, training-data poisoning.

The pilot ran a competitive selection in 2024 and confirmed candidates work with a dedicated MHRA Airlock Team and one of the Approved Bodies (BSI, DEKRA, SGS, TÜV SÜD, UL). Each candidate has a defined sandbox scope, success criteria, and a publication plan so the wider sector can learn from the case study.

• Not a fast-track UKCA approval — Airlock participation does not waive any regulatory requirement.
• Not a clinical trial sandbox — clinical investigations remain under the Medicines for Human Use (Clinical Trials) Regulations and MHRA device-trial pathways.
• Not a substitute for the underlying QMS — manufacturers still need ISO 13485 and IEC 62304 discipline.
• Not a permanent programme by default — the pilot is structured to inform statutory and guidance change.

The UK is reforming its medical-device statute (UK MDR 2002 as amended) under the SaMD Change Programme. The Airlock generates evidence and case-study learning that feeds the reform — particularly on pre-authorised change control (UK equivalent to FDA's PCCP), real-world performance monitoring requirements, and equity expectations. Manufacturers operating in the Airlock get earlier sight of where the regulator's thinking is heading.

• If the product is genuinely novel and a UK launch is in plan, Airlock is the cheapest way to de-risk the regulatory pathway.
• Even if not selected, the published case-study output is the best read on where MHRA expectations are going.
• Quality system, intended-purpose articulation, change-control SOP and real-world performance monitoring framework all need to be in mature shape to participate productively.
• Cross-jurisdictional manufacturers should map Airlock outputs to FDA PCCP and the EU AI Act so a single technical-file artefact serves all three regimes.