The UK Medicines and Healthcare products Regulatory Agency opened a call for evidence on 22 September 2026 for its Beyond ADMET: AI for medicines safety survey. The agency is asking organisations across toxicology, preclinical research, clinical testing, life sciences, biotechnology and AI to explain how they use or plan to use AI for medicines-safety evaluation.
This is a compliance story, but not because it creates a new rule. MHRA says the evidence will inform a Beyond ADMET regulatory sandbox and future regulatory work. The useful signal is that the regulator is moving from general AI principles toward the records needed to evaluate an AI model in a specific safety context.
What MHRA is asking
ADMET stands for absorption, distribution, metabolism, excretion and toxicity. These are core questions in medicines development because they help teams understand how a medicine behaves in the body and where safety problems may appear.
MHRA says AI may help by integrating traditional preclinical data with clinical and real-world data across populations, disease states and clinical conditions. The agency is not asking only whether models perform well. It is asking about model development, validation, data access, data sharing and regulatory considerations.
That list matters. For regulated AI, an impressive result is not the same as acceptable evidence. A medicines-safety model may need to show where its data came from, which patient groups are represented, how performance was tested, what limitations remain, who owns the evidence and when the model must be reassessed.
The sandbox signal
The survey is connected to a future regulatory sandbox. A sandbox is not final guidance. It is a controlled way for regulators and innovators to test practical cases before expectations become fixed.
That can be useful for medicines-safety AI because context changes the evidence burden. A model used to prioritise additional review is not the same as a model used to support a regulatory submission or guide a safety decision. The closer the AI output sits to a consequential decision, the stronger the validation and change-control record needs to be.
AIforPharma independently framed the initiative in similar terms, noting that the central question is what evidence a regulator would need before relying on AI output. It also highlighted data credibility, lifecycle control and context of use as the hard problems behind the survey.
What teams should prepare
The immediate action for pharmaceutical, biotech and AI teams is not to claim that MHRA has endorsed any model. It has not. The practical action is to organise evidence around the model's intended use.
Useful records include the model purpose, the decision it supports, data provenance, population coverage, validation method, known limits, monitoring plan, retraining trigger, owner, approval history and a path for explaining changed results. The Maetra AI compliance evidence checklist gives a reusable structure for turning those records into current, owned evidence rather than scattered documents.
For agentic systems, the evidence problem gets sharper. An AI workflow that searches data, calls tools or routes recommendations needs an inventory of the systems it touches and an audit trail of what happened. The Maetra agent inventory guide is a useful starting point because teams cannot validate a workflow they have not mapped.
What remains uncertain
The public source is a call for evidence. It does not establish a new compliance requirement, approve any AI model, define final validation criteria or say that AI-generated medicines-safety evidence will be accepted in submissions.
It also does not resolve the data-sharing problem. Medicines-safety evidence often spans clinical, preclinical, pharmacovigilance and real-world datasets. Linking those sources can raise privacy, provenance, representativeness and interoperability issues that a high-performing model cannot erase.
Maetra analysis
MHRA's survey is important because it makes the next stage of AI regulation more concrete. The question is no longer only whether AI should be transparent, safe or human overseen. It is what evidence proves that an AI output is credible for one defined safety use.
Teams that want to use AI in medicines safety should treat the survey as a prompt to build an evidence file now. Map the intended use, connect every dataset to provenance and coverage, define validation and reassessment rules, and preserve a reviewable history of changes. If future sandbox work turns into guidance, the teams with current evidence will be in a much better position than the teams with a model card and a slide deck.