AI regulation in biotechnology: beyond the EU AI Act

Futuristic AI brain with blue circuits representing AI regulation, governance, and compliance in biotechnology

Why GxP, quality systems, and decision accountability may matter more than most board discussions about AI compliance.

Ask most executives what regulates their use of AI and the short answer is usually: the EU AI Act.

That answer is understandable. But it is also incomplete, and in 2026 it became more incomplete than it was. Specifically, for biotech organisations, the most immediate governance and compliance pressure rarely comes from AI specific legislation. It comes from the frameworks already governing quality, manufacturing, clinical, and regulatory submissions: frameworks that were written before anyone used the term large language model and that apply regardless.

Understanding where AI stands therefore means looking past the EU AI Act first.

What the EU AI Act actually changes

The EU Artificial Intelligence Act, Regulation (EU) 2024/1689, was never a single switch. It arrived in stages, and only some of them have landed.

The first prohibitions on certain practices, and the AI literacy obligation in Article 4, have applied since 2 February 2025. Governance rules and the obligations attaching to general purpose AI models have applied since 2 August 2025.

The high-risk regime was due to follow in August 2026. But it did not. Regulation (EU) 2026/1744, the Digital Omnibus on AI, was adopted on 8 July 2026 and entered into force on 27 July, on the reasoning that standards, common specifications, and national competent authorities were not ready and that holding the original date would raise costs without justification. The obligations on high-risk AI systems now apply from 2 December 2027 for the use cases listed in Annex III, and from 2 August 2028 for AI embedded in regulated products, a category that expressly includes AI-based medical software.

Two practical consequences follow, and both are routinely missed.

The AI literacy obligation did not move, but its content changed. The Digital Omnibus replaced Article 4 rather than postponing it. The original text required providers and deployers to ensure AI literacy among their staff. The replacement requires them to take measures to support the development of it, judged against people existing knowledge, experience, and training and the context in which the systems are used; it states expressly that nobody must guarantee a specific level for any individual. The duty became lighter, not optional, and it has been live since February 2025.

What was postponed concerns products, not internal use. The high-risk regime governs systems placed on the market. For a biotechnology company that mainly means what it sells: a device, a diagnostic, or software with a medical purpose. It says very little about the AI the company runs on itself.

Why biotechnology should not focus only on the EU AI Act

The AI the company runs on itself is governed elsewhere, and that is the point most board discussions miss.

AI inside a biotechnology organisation touches batch records, deviation handling, laboratory data, trial conduct, and regulatory submissions. Those activities are governed by good manufacturing practice, by good clinical practice, by the quality management system, by data integrity expectations, and by validation requirements for computerised systems. None of those obligations has a 2027 date attached. They apply now, they are inspected now, and they were designed for an era when a computerised system behaved the same way twice.

There is a deeper structural reason not to read a biotechnology company’s position off the EU AI Act. As a The Conversation article has observed, laws are generally organised around particular technologies, organisms, or activities; so what happens when a new biological capability emerges from a combination of technologies regulators have always treated separately? AI, CRISPR-based gene editing, and synthetic biology have converged to the point where that is a live question rather than a thought experiment. The demonstration that made it concrete was benign and striking: AI-designed bacteriophages, viruses that attack only bacteria, which rapidly overcame antibiotic-resistant strains of E. coli.

The EU AI Act classifies by intended purpose and risk category. It does not classify biological capability, and it was never meant to.

Manufacturing: the significance of draft Annex 22

In the manufacturing discipline (i.e. GMP), the European Commission has consulted on three documents as a package: a revision of Chapter 4 on documentation, a revision of Annex 11 on computerised systems, and an entirely new Annex 22 on artificial intelligence. The consultation is closed. Annex 22 remains a draft and has not yet been adopted.

Read it anyway, because it is considerably stronger than the EU AI Act.

Annex 22 covers AI models used in critical applications, meaning those with direct impact on patient safety, product quality, or data integrity. Inside that scope it draws three lines, and none of them concerns how much validation effort a company is willing to spend.

Firstly, it applies to static models, those that do not adapt during use. Models that continuously and automatically learn and adapt fall outside it and, in the draft’s own words, should not be used in critical GMP applications.

Secondly, It applies to models with a deterministic output, where identical inputs give identical outputs. Models whose output may vary fall outside it and, again, should not be used in critical GMP applications.

And finally, it states unequivocally that it does not apply to generative AI or large language models (LLMs), and that such models should not be used in critical GMP applications at all. Where they are used for non-critical purposes, a suitably qualified person remains responsible for deciding whether the output is fit for its intended use.

This now contrasts with the proposals reaching executive committees. The technology in most of them is generative: the assistant that writes the deviation summary, reviews the batch record, or answers the question about the procedure. On the Annex 22 draft logic, that is precisely the class of model that does not belong in critical applications.

Whether or not Annex 22 changes before adoption, it offers valuable insight into how regulators are increasingly thinking about AI in GMP environments. A company that sets its own model-class rule now will have a far easier conversation later than one that lets a supplier set it.

Clinical development: the rules already exist

A clinical-stage company may reasonably conclude that the section above belongs to someone else, because it manufactures nothing itself. That conclusion does not hold, for two separate reasons.

The first is that the clinical discipline has had its own instrument for years. The EMA Guideline on computerised systems and electronic data in clinical trials has been in effect since 2023, and artificial intelligence sits among its own listed keywords. It was written by the GCP Inspectors Working Group, which is to say by the people who carry out the inspections.

Alongside it, ICH E6(R3) came into effect on 23 July 2025 for its overarching principles and its Annex 1. The Annex 2, covering pragmatic and decentralised designs and trials drawing on real-world data, was adopted in June 2026 and takes effect on 15 January 2027.

ICH E6(R3) rests on proportionate, risk-based quality management rather than uniform procedure, which makes it a better foundation for governing a model than any AI-specific text available today. It asks the question that actually scales: what could go wrong for the participant or for the reliability of the result, and what is being done about it.

So AI governance in a clinical environment is not unsupported. It is already carried by EMA guidance, by computerised systems expectations, by risk-based quality management, and by ICH E6(R3). What is usually missing is not a framework but the decision to apply the one already in force.

The outsourcing challenge nobody discusses

The second reason is structural, and it is the gap most likely to be found by someone else first.

Most biotechnology companies do not perform their own manufacture, and many do not run their own trials. They contract both. A contract moves the activity; it does not move the obligation. A marketing authorisation holder carries the consequence when a contract manufacturer puts a model into batch release, environmental monitoring, or deviation triage. A sponsor may delegate trial-related duties to a contract research organisation, but retains responsibility for the quality and integrity of the trial data.

Meanwhile the contractors are adopting AI at least as quickly as their clients, and often faster, because efficiency is their margin.

Very few quality agreements signed before 2025 mention AI at all. Fewer still answer the questions that matter: which model classes are acceptable in critical applications, who approves a change to a model, what happens when model performance drifts, what has to be notified and within what period, and what evidence the sponsor is entitled to see on request.

Those are not technical clauses. They are governance clauses, and in most agreements of that generation they are simply absent, across contract manufacturers, contract development and manufacturing organisations, contract research organisations, and the growing number of specialist suppliers whose service is itself AI-enabled.

What regulators are converging around

The instruments above were written by different authorities, in different disciplines, for different purposes. They converge on the same question, which is the most useful fact in this whole subject.

The EU AI Act classifies by intended purpose. Draft Annex 22 opens with a chapter on intended use, and makes acceptance criteria, test data, explainability, and confidence all follow from it. The FDA draft guidance Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products builds a risk-based credibility assessment framework whose first two steps are to define the question of interest and then the context of use, with the weight of evidence scaling to the model’s risk within it.

Not one of them asks whether an organisation uses AI. Each asks what specific decision a given model influences, and what evidence justifies relying on it for that decision.

Which is why governance designed around the technology keeps failing. A policy about what staff may type into a chat assistant answers none of the three. An inventory organised by software licence answers none of them either. Governance organised around the decision answers all three at once, because the decision is the only thing all three are asking about.

Four actions to take now

  1. Discharge the AI literacy obligation as it now stands. It has applied since February 2025 and the Digital Omnibus made it lighter, not optional. Demonstrable measures, proportionate to role and context, covering the people who deploy and use these systems and the people approving their purchase. Nobody has to be certified. Something has to exist.
  2. Inventory by decision, not by tool. For every use, record the decision the model influences, the discipline that governs that decision, and the consequence if the output is wrong. A list of software licences is not an inventory and will not survive a conversation with an inspector.
  3. Decide the model-class rule now, in writing. Which applications are critical. That critical applications use static, deterministic models. That generative models are confined to non-critical work with a named person responsible for the output. Deciding this before a supplier decides it is the difference between governance and paperwork.
  4. Put AI into the quality agreements. Model change control, performance monitoring, drift thresholds, notification duties, and the sponsor’s right to evidence. Where manufacture or trial conduct is contracted, the agreement is the only place the obligation can actually be exercised.

None of those four actions requires a new platform, a budget cycle, or a deadline in 2027.

Finally

The important question is not whether an organisation uses AI.

The important questions are where it is used, for what purpose, under which quality framework, on what evidence, and under whose accountability.

Organisations that answer those five questions now will find compliance considerably easier as AI-specific regulation continues to mature, because the answers are what every instrument discussed here is asking for in its own vocabulary.

Those that wait for regulatory certainty are likely to discover something less comfortable: that the obligations already in force were sufficient to require action all along.

The question of who owns an AI influenced decision, and how accountability is maintained once these systems are embedded in daily work. This is the subject of an Akility article: AI governance in biotechnology: govern the decision, not the technology.


Working with Akility

Akility establishes pragmatic governance across IT, AI, data, quality, and security for Life Sciences and Healthcare organisations, so that innovation delivers measurable value while accountability and human oversight hold. Akility also takes over IT Director or CIO roles, interim, fractional, or long-term. Akility sets the technology vision and leads its delivery, from vision to value.


 

Sources

 

For more references: consult the Regulation and Guidelines page

Pictures Credits: Pixabay

Articles on this site may be drafted with the assistance of AI tools. Every article is read, checked, and approved by its named author before publication.