The MHRA has launched a first-of-its-kind AI regulatory sandbox to test whether artificial intelligence can predict how medicines behave in the body and catch safety risks that existing methods routinely miss.

UK Medicines Watchdog Bets on AI to Fix a Drug Development System With a 90% Failure Rate

The UK medicines regulator has moved to place artificial intelligence at the heart of how drugs are tested, assessed and ultimately approved. A shift with significant consequences for the patients most vulnerable to adverse reactions, including older people receiving care at home and those with complex, multiple conditions managed across community health settings.

Announced by Science Minister Lord Vallance, the MHRA’s new AI sandbox gives companies and researchers a regulator-supervised environment in which to test AI tools designed to predict how medicines behave inside the body, how they are absorbed, processed, and whether they are likely to cause harm. Up to five AI-driven approaches will be tested in the first phase, with industry and academic partners joining from summer 2026 to shape the programme’s design.

The announcement is the latest in a series of moves by the MHRA to position regulatory sandboxes as the primary mechanism for managing AI’s entry into healthcare. 

Its AI Airlock programme focused on AI-enabled medical devices completed its second phase earlier this year and secured renewed multi-year funding, with a third phase now in development. The medicines sandbox extends that model into pharmacology, and the regulator is clearly building a playbook.

Why Adverse Drug Reactions Matter Beyond the Hospital Ward

Approximately 250,000 people in the UK are admitted to hospital each year because of adverse drug reactions, a burden that costs the NHS more than £2 billion annually. 

For commissioners and providers working in home care and community health, those numbers are not abstract. Older adults managing multiple long-term conditions are disproportionately exposed to drug interaction risks, particularly when care is delivered across fragmented services with limited communication between prescribers.

The challenge is not simply clinical; it is systemic. Existing methods for predicting how medicines will behave in patients are limited, particularly for groups that have historically been underrepresented in clinical trials children, older people, and people from diverse ethnic backgrounds. 

The sandbox specifically commits to using clinical data to improve understanding of medicines across these groups, which is directly relevant to community health commissioners trying to serve populations whose needs are rarely reflected in the pivotal trial data on which prescribing decisions are based.

A 90% Failure Rate and What It Costs the Sector

Around nine in ten medicines fail during development, partly because the tools used to predict human response are inadequate. That failure rate has real consequences, promising treatments do not reach patients, development costs inflate across the industry, and the medicines that do progress carry residual uncertainty about how they will behave in real-world populations. For a health and care system already operating under sustained financial pressure, the costs of that uncertainty extend well beyond the pharmaceutical supply chain.

The MHRA’s sandbox is intended to address the problem at source testing whether AI models can generate more reliable predictions earlier in development, before costly late-stage failures occur. Dan O’Connor, Executive Director of Regulatory Policy at the Association of the British Pharmaceutical Industry (ABPI), described the initiative as a potential step toward the “safe harbour” the industry has been pressing for a trusted environment in which to develop AI tools that meet regulatory expectations before they reach formal assessment. The ABPI has stressed that success will depend on meaningful industry engagement and coherent alignment with the Centres of Excellence in Regulatory Science and Innovation, known as the CERSIs.

Regulator as Architecture, Not Obstacle

What distinguishes the sandbox model here as in the AI Airlock is the deliberate integration of regulatory oversight into the testing process itself, rather than treating regulation as something that happens at the end. Companies working within the sandbox will do so in direct collaboration with the MHRA, generating evidence on AI tool reliability that will ultimately inform how the regulator frames its expectations for safe use going forward. This is not a soft-touch arrangement; the outputs are intended to shape policy.

Professor Alastair Denniston, who chairs the National Commission into the Regulation of AI in Healthcare, described the sandbox as a practical means of establishing what credible evidence looks like for AI tools that predict drug safety and pharmacokinetics the science of how medicines move through the body. That framing is instructive. The regulator is not simply permitting innovation to happen; it is helping to define the evidential standards that will govern it.

Professor Chris Molloy, chief executive of the BioIndustry Association, welcomed the announcement, noting that AI models need to be “taught, tested and proven in a rigorous, safe space” before they can be relied upon in a regulatory context. Whether five initial approaches in phase one is sufficient to generate that rigour at scale remains an open question.

What the NHS’s 10 Year Plan Demands and What the Sandbox Can Deliver

The medicines AI sandbox sits within a broader architecture of government ambition. It supports the AI for Science Mission One, which commits the UK to developing new treatments faster, and reinforces the government’s drive to reduce reliance on animal testing, a policy direction with its own scientific and ethical momentum. It also contributes to the 10 Year Health Plan’s stated goal of making the UK the most AI-enabled healthcare system in the world.

Those ambitions are large, and the sandbox is for now a carefully scoped first step. The MHRA will begin working with partners in summer 2026, and the first-phase findings will take time to translate into regulatory guidance, let alone into medicines that reach patients in community or home care settings. The timeline from sandbox insight to clinical reality is rarely short.

But the architecture being built matters. A regulator that is actively shaping the evidence standards for AI in medicines development rather than waiting for industry to present finished tools for approval is a materially different proposition for the health and care sector. Whether that architecture can move quickly enough to match the pace of AI development is the question the sector will be watching closely.

The sandbox will not fix overnight the long-standing failure to include older people and those with complex care needs in clinical trial populations. But if it builds the evidence base to make that inclusion the norm rather than the exception, its consequences for care delivered at home and in the community could be substantial.