How Can Teams Put the FDA–EMA AI Principles Into Practice?

In January 2026, the FDA and EMA released ten shared principles for good AI practice in drug development. A practical team question is which workflows need an additional record, check or accountable owner. Applying the principles to a defined use makes them useful in daily work.

The shared document addresses people’s interests, risk, standards, intended use, expertise, data, model practices, performance, ongoing management and essential information. It does not certify a product or prescribe one accuracy threshold for every tool. The list below offers this article’s implementation suggestions; it is not a regulator-mandated form.

• People’s interests and rights: Identify who is affected and preserve professional accountability
• Controls proportional to consequences: Set review intensity and escalation according to potential harm
• Applicable requirements: Check jurisdiction, work type and existing quality procedures
• Defined purpose and limits: Specify inputs, outputs, supported decisions and excluded uses
• Appropriate expertise: Involve relevant clinical, statistical, data, quality and operational staff
• Managed data and records: Keep sources, versions, permissions and important transformations
• Suitable development practices: Understand task fit, configuration or training basis and limitations
• Evaluation suited to risk: Test representative material and consequential errors
• Ongoing oversight: Define change, monitoring, reassessment and stopping conditions
• Usable information: Explain capabilities, uncertainties and how to review outputs

Start with a use statement.

For example, a tool reads public registry records and organizes eligibility differences for a development meeting. It does not decide patient eligibility or automatically revise a formal protocol. Define the source scope, output, reviewer and handling of missing material. A specific purpose is easier to evaluate than a promise to transform all development.

Make acceptance checks executable.

For that tool, test omitted restrictions, negation, time windows, source locations and version changes. Define separately the errors that could change population judgments. Establish representative cases, human reference material and failure handling before testing. Internal thresholds should fit the use and consequences; do not claim the shared principles mandate a particular numerical target.

Documentation should be sufficient to explain how a consequential result arose. Consider retaining input versions, processing rules, model or configuration version, outputs and corrections. Reassess requirements within the relevant quality system if outputs support formal study materials or regulatory evidence.

Treat updates as changes requiring judgment.

A model, retrieval collection, prompt or input process may change while the product name stays the same. Determine which changes affect evaluated tasks, when to retest and how to detect unusual output. A first demonstration cannot permanently establish every later version’s suitability.

Distinguish document status too. The FDA’s January 2025 AI guidance for regulatory decisions was still labeled draft when checked and discusses credibility and context. Shared principles, draft guidance, regulations and internal procedures are different levels of authority.

Chinese teams also need to check local obligations and quality requirements. International principles are useful references without replacing them. Relevant functions can agree what belongs in procedures, what remains pilot documentation and what requires dedicated discussion.

A practical minimum is a use statement, representative evaluation record, source and version record, and ownership of changes and exceptions. Link each to an actual task so the principles help the team decide when to rely on output and when to return to manual review.