MHRA regulates artificial intelligence (AI) used in UK clinical research primarily through two channels that work together rather than as separate systems: the standard medical-device regulatory pathway under UK law for any AI tool that qualifies as Software as a Medical Device (SaMD), and the AI Airlock, a regulatory sandbox launched in Spring 2024 specifically to work through the novel regulatory questions AI raises before formal guidance catches up. There is no separate “AI trial approval” process distinct from the UK’s existing clinical-research and medical-device frameworks — AI-enabled tools are regulated as devices (or as part of a trial’s investigational product) using the same statutory instruments as any other technology, with the Airlock layered on top to resolve specific open questions.
AI Airlock at a Glance
| Phase | Status | Timing | Scope |
|---|---|---|---|
| Pilot | Complete | Spring 2024 – April 2025 | 5 technologies selected; one participant (Lenus Health) withdrew, leaving 4 tested products across varying regulatory maturity levels |
| Phase 2 | Complete | Applications closed July 2025; concluded May 2026 | 7 innovators working through 3 regulatory challenges: managing evolving/adaptive AI applications, regulating AI-powered diagnostics, and post-market surveillance for AI devices |
| Phase 3 | In design | Details expected later in 2026 | Not yet published |
| Programme funding | Secured | Announced April 2026 | £3.6 million over three years, moving the Airlock off annual budget cycles |
Last verified 16 August 2026 against MHRA’s own AI Airlock collection page on GOV.UK. Phase timelines and cohort details are the fastest-moving part of this page — recheck the primary source before citing a specific date more than a few months out from this verification.
What Actually Regulates AI-Enabled Devices in the UK
Medical devices in Great Britain — including standalone software and AI that meets the legal definition of a medical device — are governed by the UK Medical Devices Regulations 2002 (SI 2002/618, as amended), not by the EU Medical Device Regulation (EU MDR 2017/745) directly. This is a retained, domestic UK framework, a distinction that regularly trips up teams used to working under EU MDR. Compliant devices carry UKCA (UK Conformity Assessed) marking to be placed on the Great Britain market.
Because UKCA-certification capacity is still building, the UK government has extended transitional recognition of CE-marked devices in Great Britain: devices certified under the old EU Medical Devices Directive/Active Implantable Medical Devices Directive (MDD/AIMDD) are accepted until 30 June 2028 (or certificate expiry, if sooner), and devices certified under the current EU MDR, or under the EU In Vitro Diagnostic Regulation (IVDR), are accepted until 30 June 2030. MHRA has also indicated it intends to consult on indefinite recognition of CE-marked devices beyond those dates. A sponsor or manufacturer bringing an AI-enabled SaMD into a UK trial needs to know which certification route the device is actually using before assuming UKCA is the only option.
This device-side framework sits alongside, not instead of, the clinical-trial regulations that already govern the trial itself — see CASRAI’s guide to CTIMP vs non-CTIMP classification for how a trial’s regulatory category is determined, since an AI-enabled diagnostic or decision-support tool used within a trial may itself need separate device authorisation independent of the trial’s own approval route.
The Software and AI as a Medical Device Change Programme
MHRA’s Software and AI as a Medical Device Change Programme, published 16 September 2021, is the umbrella policy initiative behind the UK’s current approach to regulating software and AI medical devices. Its stated aim is to modernise device regulation for software specifically, so the framework provides “a high degree of protection for patients and public” while positioning the UK as a viable location for responsible AI medical-device development. The programme operates through the AI and Digital Regulations Service, a multi-agency partnership between MHRA, NICE, the Care Quality Commission (CQC), and the Health Research Authority (HRA) that helps developers navigate the regulatory, health-technology-assessment, and NHS-adoption requirements that apply to a given product — useful context for research administrators because it means an AI-enabled trial technology can touch four different bodies’ requirements, not just MHRA’s.
The AI Airlock: MHRA’s Regulatory Sandbox for AI as a Medical Device
The AI Airlock is described by MHRA as the UK’s first regulatory sandbox purpose-built for AI as a Medical Device (AIaMD) — a “proactive, collaborative, agile” mechanism for testing real products against genuinely unresolved regulatory questions, rather than waiting for finished guidance to be written in the abstract. Participation brings together MHRA, UK Approved Bodies (the UK’s post-Brexit equivalent of EU Notified Bodies), the NHS, and in some cases other regulators, working through a specific product’s regulatory question in a structured, time-bound way, with the findings intended to inform guidance MHRA subsequently publishes for the whole sector.
It is not an approval or certification route in itself — a product going through the Airlock still has to obtain UKCA marking (or rely on transitional CE recognition) through the ordinary route once the regulatory question it’s testing has been resolved. The Airlock exists to work out how an unresolved question should be answered, not to grant market access.
The pilot phase (Spring 2024 – April 2025) tested 4 completed products (5 were originally selected; Lenus Health withdrew) spanning a range of medical disciplines and regulatory maturity levels. Phase 2 (applications closed July 2025, concluded May 2026) worked with 7 additional innovators on technologies including AI-powered clinical note-taking, cancer diagnostics, eye-disease detection, and obesity-treatment support systems, organised around three named regulatory challenges: managing AI applications that continue to evolve after deployment, regulating AI-powered diagnostics, and building robust post-market surveillance for AI medical devices. In April 2026, MHRA secured £3.6 million in funding over three years specifically to move the Airlock beyond single-year budget cycles, and a Phase 3 design was in progress as of this page’s last verification, with details expected later in 2026.
Where This Intersects With Clinical Research Specifically
For a UK clinical-research team, the AI Airlock becomes relevant at a specific, narrow point: when a trial (or a device intended to generate evidence supporting a device submission) uses an AI-enabled SaMD whose regulatory classification, intended-use boundary, or change-management approach is not clearly resolved by existing MHRA guidance. In that situation, a sponsor or manufacturer — not typically the research administrator alone, but usually in coordination with regulatory affairs — can seek a place in a future Airlock cohort to work through the specific question with MHRA directly, rather than making a unilateral classification call and hoping it survives scrutiny at submission.
This sits alongside, not in place of, the trial’s ordinary ethics and regulatory pathway. Human Research Authority (HRA) approval via IRAS and research ethics committee review still apply to the trial itself; the Airlock addresses the device-classification and evidence-generation question for the AI tool, not the trial’s own ethical or scientific review. Similarly, Good Clinical Practice obligations under ICH E6(R3) — see CASRAI’s guide to what changed under E6(R3) and the broader GCP overview — continue to govern trial conduct regardless of whether the AI component involved is also working through the Airlock. Sites and sponsors already subject to routine MHRA GCP inspection should expect an AI-enabled data-collection or decision-support tool used in a trial to be examined the same way any other trial system would be — for data integrity, audit trail, and validation — independent of its device-classification status.
SaMD Classification for AI Under UK Law
AI that meets the definition of a medical device in the UK is classified, like any other SaMD, on a risk-based tier system under UK MDR 2002 (Class I, IIa, IIb, or III, broadly mirroring but legally distinct from the EU MDR’s classification structure) — the higher the potential harm from an incorrect or delayed output, the higher the class and the more rigorous the conformity-assessment route. For general background on how software qualifies as a medical device and how classification tiers work, see CASRAI’s guide to Software as a Medical Device (SaMD), which also covers the shared Good Machine Learning Practice (GMLP) principles that FDA, Health Canada, and MHRA jointly articulated — the same underlying development principles apply whichever jurisdiction’s classification system a given product is being assessed under.
What AI adds on top of ordinary SaMD classification is the question the Airlock’s Phase 2 was explicitly built to test: an AI model that continues to learn, or that a developer wants to update post-deployment, does not fit neatly into a classification and conformity-assessment system designed around a fixed, unchanging product. The UK does not yet have a published, finalised equivalent to FDA’s Predetermined Change Control Plan (PCCP) framework for pre-authorising a defined envelope of post-market AI updates — managing evolving AI applications was one of the Airlock’s three named Phase 2 challenges, meaning the UK’s approach to this specific problem is still being worked out through real cases rather than settled guidance. Teams planning an adaptive or continuously-trained AI tool for use in or alongside a UK trial should treat this as an open regulatory question to raise proactively with MHRA, not an already-solved compliance checkbox.
MHRA’s AI Airlock vs. FDA’s AI/ML Framework
Because many device developers and CROs operate across both the UK and US, it helps to see the two regulators’ approaches side by side. For the full FDA picture, see CASRAI’s FDA AI guidance guide.
| Dimension | UK – MHRA | US – FDA |
|---|---|---|
| Core legal framework | UK MDR 2002 (SI 2002/618, as amended); UKCA marking | FD&C Act; 510(k)/De Novo/PMA pathways under FDA’s device authorities |
| Mechanism for unresolved AI questions | AI Airlock — a time-bound regulatory sandbox with named cohorts and published findings | Predetermined Change Control Plan (PCCP) — a pre-authorised, product-specific plan filed as part of a marketing submission |
| Status as of this page’s verification | Pilot and Phase 2 complete; Phase 3 in design; no finalised general framework yet published | PCCP framework finalised and in active use for cleared/authorised devices (see CASRAI’s FDA AI guidance guide for current status) |
| Shared development principles | Good Machine Learning Practice (GMLP) — jointly issued with FDA and Health Canada | Same GMLP principles, jointly issued |
| Multi-agency coordination | AI and Digital Regulations Service (MHRA, NICE, CQC, HRA) | Primarily CDRH for devices; separate CDER/CBER track for AI in drug development decision-making |
Practical Checklist for UK Clinical Research Teams
- Confirm whether the AI tool is a medical device at all. Not every AI application used around a trial (e.g. purely administrative or scheduling tools) meets the legal definition — only tools intended for a medical purpose (diagnosis, prevention, monitoring, treatment, prediction, or prognosis) typically qualify.
- Identify the certification route in play. Determine whether the device will carry UKCA marking or is relying on transitional CE recognition, and confirm which recognition deadline (30 June 2028 for MDD/AIMDD-certified devices, 30 June 2030 for MDR/IVDR-certified devices) applies before assuming either route is settled indefinitely.
- Separate the device question from the trial question. HRA/IRAS approval and REC review govern the trial; UK MDR/UKCA and, where relevant, the AI Airlock govern the device. Both may need to run in parallel, not sequentially.
- Flag adaptive/continuously-trained models early. If the AI component is designed to update after deployment, raise the post-market change-control question with regulatory affairs and consider whether the product is a candidate for a future AI Airlock cohort, given the UK has not yet published a finalised general framework for this specific problem.
- Treat AI-enabled trial systems as subject to ordinary GCP inspection. Data integrity, audit trail, and validation expectations under MHRA GCP inspection apply to AI-driven data-capture or decision-support tools the same way they apply to any other trial system.
Frequently Asked Questions
What is the MHRA AI Airlock?
The AI Airlock is a regulatory sandbox MHRA launched in Spring 2024, purpose-built for AI as a Medical Device (AIaMD). It brings together MHRA, UK Approved Bodies, the NHS, and sometimes other regulators to work through unresolved regulatory questions using real, selected products, with the findings intended to inform guidance for the wider sector. It is not itself a market-authorisation route.
Is the AI Airlock mandatory for AI-enabled medical devices used in UK trials?
No. Most AI-enabled SaMD goes through the ordinary UK MDR 2002 classification and UKCA (or transitional CE) conformity-assessment route without ever entering the Airlock. The Airlock is reserved for products facing a genuinely unresolved regulatory question that existing guidance does not clearly answer, and participation has so far run through defined, competitive cohorts (the pilot and Phase 2) rather than continuous open enrolment.
How does the UK’s device framework differ from the EU’s for AI-enabled devices?
Since Brexit, Great Britain regulates medical devices, including AI/software, under the domestic UK Medical Devices Regulations 2002 (SI 2002/618, as amended), with UKCA marking as the compliance mark — not directly under the EU Medical Device Regulation (EU MDR 2017/745). The UK government has extended transitional recognition of CE-marked devices (until 30 June 2028 or 30 June 2030 depending on which EU legislation certified the device) while UKCA-certification capacity builds out.
How does MHRA’s approach to adaptive AI compare to FDA’s Predetermined Change Control Plan (PCCP)?
FDA has a finalised PCCP framework that lets a manufacturer pre-authorise a defined envelope of post-market AI/ML changes as part of a marketing submission. The UK has not yet published an equivalent finalised framework; “managing evolving AI applications” was one of the three named regulatory challenges the AI Airlock’s Phase 2 cohort worked on, meaning the UK’s approach to this specific problem is still being developed through sandbox cases rather than settled policy as of this page’s last verification.
Who can apply to join a future AI Airlock cohort?
Historically, developers of AIaMD products facing a genuine, unresolved regulatory question have applied during defined application windows tied to each cohort (the pilot and Phase 2 both ran this way). MHRA has indicated a Phase 3 is in design, with details expected later in 2026 — check MHRA’s own AI Airlock collection page on GOV.UK for the current application status before assuming a window is open.
Sources: MHRA/GOV.UK AI Airlock: the regulatory sandbox for AIaMD (collection page, accessed 16 August 2026); GOV.UK AI Airlock pilot cohort (published 4 December 2024, updated 28 February 2025); GOV.UK Software and AI as a Medical Device Change Programme (published 16 September 2021, updated 14 June 2023); GOV.UK Regulating medical devices in the UK. This page was last verified against those primary sources on 16 August 2026 — the Airlock’s phase status, cohort details, and any UKCA transitional deadlines are the parts most likely to change; recheck GOV.UK directly before relying on a specific date more than a few months past that verification.
