Saturday, November 14, 2026

EMA Reveals Its AI Governance Model- Risk Assessment, Guardrails and Human-Supervised AI by 2030


Introduction⇒ EMA has published its operational AI governance model, including use-case-specific risk assessment, real-time guardrails, multidisciplinary oversight and a policy framework for human-supervised AI-enabled regulatory work by 2030.

The European Medicines Agency has published a detailed description of how it intends to govern artificial intelligence inside the European medicines regulatory network. Published on 11 November 2026, the new EMA resource describes use-case-specific AI risk assessment, technical guardrails, multidisciplinary oversight and a policy framework intended to support a transition toward AI-enabled regulatory work under human supervision by 2030.

For pharmaceutical companies, this is important even though the framework governs EMA’s own AI use: it shows how one of the world’s most influential medicines regulators is translating high-level AI principles into operational controls.

EMA Is Moving from AI Principles to Operational Governance

EMA states that artificial intelligence is becoming increasingly important for working with large volumes of regulatory and health data.

Its AI capabilities may support:

  • personal productivity;
  • process automation;
  • improved analysis of data;
  • regulatory decision support.

The important development is that EMA is not treating AI deployment simply as an IT project.

It has established a governance structure intended to determine:

  • which AI systems may be deployed;
  • which risks need to be assessed;
  • which controls are required;
  • how legal and ethical requirements are addressed;
  • how AI remains under appropriate human supervision.

A Dedicated AI Coordination Group Provides Strategic Oversight

EMA has established an AI coordination group responsible for strategic direction of:

  • AI activities within EMA;
  • EMA activities supporting the broader European medicines regulatory network.

The group is intended to align AI work with other EMA initiatives and with the Network Data Steering Group.

This is an important organisational lesson.

AI governance requires more than technical validation of individual models.

A mature framework also needs:

  • defined ownership;
  • strategic coordination;
  • risk escalation;
  • policy development;
  • cross-functional decision-making.

For pharmaceutical companies, the parallel would be an enterprise AI governance structure linked to the Pharmaceutical Quality System, IT governance, data governance and business-process ownership.

EMA Has Created an AI System Risk Management Group

EMA’s AI System Risk Management Group is a multidisciplinary operational and advisory subgroup of the AI coordination group.

Its role is to develop principles and processes for safe and compliant AI use.

EMA states explicitly that AI systems are not fully deployed without an assessment process.

This assessment is performed at the level of the specific use case.

That is significant.

The regulatory philosophy is not:

AI technology = one fixed risk category.

Instead:

AI System + Model + Data + User + Intended Use = Specific Risk Profile

Risk Assessment Is Use-Case Specific

EMA identifies several factors relevant to AI risk assessment:

  • the AI system itself;
  • the model used by the system;
  • the data processed by the system;
  • the intended user.

This is closely aligned with the broader pharmaceutical concept of Context of Use.

The same model may create very different risks depending on whether it is used to:

  • summarise meeting notes;
  • search regulatory precedents;
  • prioritise safety information;
  • support a scientific regulatory assessment.

The required control should therefore reflect how the AI affects the decision.

Data Protection and Information Security Are Explicit Assessment Areas

EMA identifies several priority areas when assessing AI tools.

These include:

  • data protection;
  • information security;
  • ethical risk;
  • legal compliance;
  • necessity and proportionality.

Data protection may require a formal Data Protection Impact Assessment.

This is an important reminder for pharmaceutical companies.

AI validation should not be treated only as a question of model accuracy.

A technically accurate system may still be unsuitable if:

  • confidential data are processed inappropriately;
  • security controls are inadequate;
  • the supplier retains sensitive information;
  • the intended use is legally inappropriate.

Necessity and Proportionality Are Part of AI Governance

One particularly interesting EMA principle is necessity and proportionality.

EMA expects an AI tool to be genuinely necessary for a legitimate purpose and not to exceed what is required for that purpose.

This is different from asking:

“Can AI perform this task?”

The stronger governance question is:

“Is AI appropriate and proportionate for this task?”

For regulated pharmaceutical applications, that can prevent unnecessary use of complex AI where simpler and more predictable technology would provide adequate control.

EMA Defines Guardrails as Real-Time Technical Controls

EMA explicitly refers to:

AI guardrails.

It describes them as real-time technical controls that operate on prompts or requests to constrain:

  • what the AI system can accept;
  • what it can generate;
  • what it can do.

This is an important development because guardrails are sometimes discussed only as an abstract GenAI concept.

EMA is treating them as part of the operational control environment.

Guardrails Can Restrict Both Inputs and Outputs

For a pharmaceutical AI system, guardrails could potentially control:

  • entry of confidential information;
  • access to prohibited data sources;
  • generation of unsupported claims;
  • use of AI outside its approved Context of Use;
  • execution of unauthorised actions.

Conceptually:

User Request

↓

Input Guardrail

↓

AI Model

↓

Output Guardrail

↓

Human Review or Controlled Action

This is stronger than relying entirely on user training.

Guardrails Do Not Replace Validation

A guardrail can prevent certain known undesirable behaviours.

It cannot demonstrate that the complete AI system is suitable for its intended use.

A controlled AI architecture still requires consideration of:

  • performance;
  • data suitability;
  • failure modes;
  • human oversight;
  • change control;
  • lifecycle monitoring.

Guardrails should therefore be considered one layer in a defence-in-depth architecture.

EMA Is Developing an AI Policy Framework

EMA states that it is preparing a formal AI policy framework.

The objective is to support:

EMA’s full transition toward AI-enabled work under human supervision by 2030.

EMA expects to publish this policy framework by the end of 2026.

The framework is intended to support safe and ethical AI development and use while remaining compliant with the EU Artificial Intelligence Act.

Human Supervision Is Explicitly Part of the Target Operating Model

The phrase “under human supervision” is important.

EMA is not describing a future in which autonomous AI replaces regulatory assessors.

Instead, the proposed operating model is closer to:

AI Capability + Defined Controls + Human Supervision + Regulatory Accountability

The balance between AI autonomy and human review may vary by use case.

However, responsibility for regulatory decision-making remains human and organisational.

EMA Is Also Building an Ecosystem of AI Tools and Agents

EMA separately describes development of an ecosystem involving:

  • AI tools;
  • AI assistants;
  • AI agents;
  • validated reusable prompts;
  • knowledge-mining systems.

More than 60 potential AI use cases have been identified across the regulatory network.

These include:

  • knowledge mining;
  • personal productivity;
  • process automation;
  • pharmacovigilance triage;
  • scientific literature screening;
  • data protection and anonymisation.

This means governance will increasingly need to address portfolios of interacting AI systems rather than isolated individual models.

This Has Important Implications for Pharmaceutical AI Governance

EMA’s internal operating model is not an industry requirement.

However, pharmaceutical companies can learn from its structure.

A mature corporate model could similarly include:

Enterprise AI Governance

↓

AI Use-Case Inventory

↓

Context-of-Use Assessment

↓

Risk Classification

↓

Technical and Procedural Controls

↓

Human Accountability

↓

Operational Monitoring

↓

Lifecycle Review

AI Risk Assessment Should Extend Beyond GxP Classification

Pharmaceutical companies often begin by asking:

“Is this AI system GxP or non-GxP?”

That remains important, but EMA’s framework suggests a broader assessment.

Other risks may include:

  • privacy;
  • cybersecurity;
  • confidentiality;
  • ethical risk;
  • legal compliance;
  • model behaviour;
  • supplier dependency.

A system can therefore be low impact from a direct GMP perspective while still creating significant organisational risk.

The AI Act and Pharmaceutical Regulation Will Operate Together

EMA’s governance framework explicitly connects its AI work with compliance under the EU AI Act.

For pharmaceutical companies operating in Europe, a similar dual assessment may become necessary:

AI Act Obligations

+

Pharmaceutical GxP Requirements

These frameworks address different regulatory questions.

Compliance with one does not automatically demonstrate compliance with the other.

Why This Development Matters

The significance of EMA’s new governance description lies in its operational detail.

The regulatory discussion is moving beyond:

fairness → transparency → human oversight

toward concrete controls such as:

  • governance groups;
  • use-case risk assessment;
  • data-protection impact assessment;
  • information-security assessment;
  • real-time guardrails;
  • deployment approval;
  • human supervision.

This provides a useful indication of what mature AI governance can look like inside a regulated medicines organisation.

Regulatory Status Note

The EMA Artificial Intelligence Governance and Policies webpage was first published on 11 November 2026.

It describes EMA’s internal governance structure and policy direction for the use of AI by EMA and the European medicines regulatory network.

It is not GMP guidance and does not create new compliance requirements for pharmaceutical companies.

Its importance for industry lies in demonstrating how a major medicines regulator is operationalising AI risk management, governance and human supervision.

Sources

EMA – Artificial intelligence governance and policies, first published 11 November 2026:

https://www.ema.euro … -governance-policies

EMA – Artificial intelligence tools and technology:

https://www.ema.euro … nce-tools-technology

EMA – Artificial intelligence guidance:

https://www.ema.euro … ntelligence-guidance