News · 16 April 2026

What a regulatory sandbox teaches developers: the MHRA AI Airlock

On 8 April 2026 the UK Government announced £3.6 million of funding, £1.2 million a year for three years, to expand the MHRA's AI Airlock. The Airlock is a regulatory sandbox for AI as a Medical Device: a way for developers, regulators, the NHS and approved bodies to test real products and learn where regulation needs to adapt. It began as a pilot in spring 2024, and a second round of projects opened in 2025 covering topics such as AI diagnostic tools and predetermined change control plans.

Why sandboxes help

AI medical devices raise questions that traditional frameworks handle awkwardly. How should a manufacturer describe a device that learns or is updated? What evidence shows that performance holds across hospitals? How should a change to the model be assessed? A sandbox lets those questions be explored with a real product and a regulator in the room.

What developers can learn from it, even without joining

Describe the intended use precisely. Much regulatory friction starts with a vague claim. "Assists radiologists by flagging suspected nodules on chest CT in adults" is testable. "AI for lung health" is not.

Plan change from the start. If you expect to update a model, define in advance what may change, how it will be validated and what triggers re-review. This is the idea behind predetermined change control plans.

Treat explainability as evidence. A regulator, a clinical safety officer and an end user all benefit from seeing what drives a device's output and where it fails. That evidence can support both approval and post-market monitoring.

Involve clinicians early. Performance on a benchmark is not the same as usefulness at the point of care. Workflow, alert fatigue and interpretation all matter.

Collect real-world performance data. Post-deployment monitoring is how safety is maintained, and regulators increasingly expect it.

A note for small companies

Regulation can look like a barrier that only large firms can climb. Programmes like this exist partly to reduce that barrier. If you are building in medical AI, follow what the sandbox publishes: its reports are a free source of practical expectations, written by the people who will assess your product.

Written in April 2026; details reflect the situation at that time.

All news