What’s changing in AI regulation for drug development - European Pharmaceutical Review

On 14 January, the European Medicines Agency (EMA) and US Food and Drug Administration (FDA) published joint guiding principles for good AI practice in drug development. This signals regulatory alignment on what ‘good’ looks like when AI supports evidence across the lifecycle, including clinical trials and safety monitoring.

The principles are not a new rulebook, but a reference point for how AI should be applied inside regulated processes. Guidelines are technology-neutral but operationally specific: define the context of use, apply risk-based controls, govern data and documentation, and manage the system across its lifecycle so oversight remains intact. They facilitate the shift from ‘can we use AI?’ to ‘can we operate AI reliably, with transparency and accountability?’

Why this matters: moving from capability to control

Over the past 18–24 months, many life sciences organisations have explored AI in ‘text-heavy’ and assistive use cases: summarising long documents, classifying content, extracting data points and drafting internal or external text.

The harder step is what comes next. As soon as AI output starts to influence regulated evidence, decisions, or actions, ‘it appears to work’ is no longer the standard. The standard is now evolving to become repeatable with oversight: clear accountability and an audit-ready record over time. That is what the EMA–FDA principles bring into focus.

Context of use becomes the hinge point

The most consequential phrase in the regulators’ framing is ‘context of use’. AI is not evaluated in the abstract, but in the context of what it is intended to do, what it influences and what risk it introduces.

A drafting assistant used to summarise internal notes is one thing. AI that uses the same summarisation method to prioritise safety case review, flag potential signals, or initiate steps in trial operations is another.

That difference matters because it changes what ‘good practice’ requires in the real world. When the context of use is higher risk, the surrounding expectations shift with it: clearer controls, stronger documentation, explicit human oversight and more disciplined change management. The practical implication is that AI programs will increasingly need an inventory of use cases mapped to risk, with defined boundaries and a shared understanding of accountability.

The unit of trust is the workflow and the record it produces

The guiding principles are technology-neutral, but they are not neutral about execution. This broadens attention beyond the model itself to the surrounding controls and documentation: data provenance, versioning, how outputs are reviewed and used, and what changed over time.

This is where a strategic design choice becomes critical: whether AI sits outside the regulated workflow as a separate tool, or whether it is embedded inside the systems where regulated work actually happens.

When AI is embedded in a regulated workflow, it can inherit what regulated environments already require: role-based permissions, standardised steps, controlled handoffs and audit trails by design. When AI sits outside, teams often end up reconstructing context and controls subsequently, copying information across tools and reconstructing the record during reviews and inspections. That is manageable in a pilot but is fragile at scale.

This ‘embedded’ direction is also where industry product strategy is heading. Many organisations are moving towards integrating AI directly into core operational systems across commercial, R&D and quality functions over the next few years. The underlying rationale is consistent: embedding AI inside the operational systems of record is structurally better aligned with governed data, auditability and repeatable oversight than treating AI as a separate overlay.

Lifecycle management stops being optional once AI moves into regulated workflows

Lifecycle management can sound abstract until you map it to how AI behaves in production.

Newer foundational models get released. Prompts derived from agent objectives and instructions evolve. Knowledge sources and transactional data change. Any of those changes can subtly alter AI outputs in ways that matter in a regulated context. The regulators’ emphasis on risk-based performance assessment and lifecycle management is, in practice, a reminder that ‘set it and forget it’ is not an option once AI influences regulated work.

What lifecycle management means in operational terms:

●Version control that is audit-relevant: not only for code and models, but also for prompts, agent instructions, templates and reference material that influences outputs

●Defined monitoring and drift triggers: what you monitor, how often and what constitutes meaningful change

●Change control proportional to risk: when a change requires documentation, review, re-qualification or rollback

●Clear accountability when AI output is accepted, overridden or escalated: so responsibility does not evaporate into ‘the algorithm said so’.

These are operating model disciplines. They are not optional add-ons delegated to a data science team. They sit at the intersection of quality, business process ownership and technology, which is exactly where regulated execution already lives.

Regulatory momentum is converging with operational expectations

The EMA–FDA principles arrive as part of a wider regulatory pattern that reinforces the same underlying message: trustworthy technology depends on trustworthy execution.

In Europe, the European Commission’s timeline for the EU AI Act makes governance expectations increasingly concrete, with the Act fully applicable from 2 August, alongside earlier milestones for prohibited practices and AI literacy (from 2 February 2025) and general-purpose AI obligations (from 2 August 2025).

In the US, FDA’s October 2024 guidance on electronic systems, electronic records and electronic signatures in clinical investigations clarifies what it means for electronic records and systems to be trustworthy and reliable. Despite not being AI guidance, it becomes directly relevant when AI output is part of evidence generation. If AI influences evidence and decisions, the integrity of the electronic record and the system that produced it becomes more central, not less.

None of this suggests regulators want to slow innovation. The joint principles are explicitly framed to engender responsible use. The operational bar is rising in a way that rewards organisations that make oversight practical, repeatable and scalable.

What leaders should do now

The practical response is not to slow down AI adoption, but to operationalise it, so each new use case does not trigger a fresh governance debate”

The practical response is not to slow down AI adoption, but to operationalise it, so each new use case does not trigger a fresh governance debate.

For most organisations, the near-term opportunity is to use the EMA–FDA principles as a design requirement for the AI operating environment, even while product capabilities across clinical, regulatory and safety continue to mature across the market. That means doing the foundational work to move from isolated AI applications to a governed capability in a regulated execution environment. This involves defining contexts of use, mapping them to risk, standardising what ‘essential information’ looks like for oversight, and putting lifecycle controls around what changes and when.

This is the moment to push vendors for specificity. Not ‘does it do AI?’ but: where is AI in the workflow, what record does it produce, what is versioned, what is auditable, what is monitored and what changes under change control? The more AI is embedded in the systems that already manage roles, permissions, workflows and audit trails, the easier it becomes to scale responsibly over time – without rebuilding evidence and oversight from scratch.

Operationalising AI for GxP environments

The EMA–FDA principles are changing how the industry approaches AI: not as a standalone add-on, but as part of the regulated foundation for evidence generation and monitoring. Biopharma’s that embed AI into controlled workflows, where decisions are reviewable and traceable and change can be managed with oversight, will be the ones that turn AI investment into real operational advantage.