ICH E6(R3) was endorsed by the ICH Assembly at Step 4 on January 6, 2025, and came into effect in the EU on July 23, 2025. It is the first revision to the GCP guideline in nearly a decade, and it arrives at the same moment that AI tools are proliferating across clinical trial operations. The timing is not coincidental: the revision was explicitly motivated by the need to address risk-based and technology-enabled approaches to clinical trial conduct, including the role of computerized systems in data management and oversight.
The guideline does not mention large language models or agentic AI. It does not have to. Section 4 on data governance, section 4.2.2 on audit trails, and section 4.3.4 on computerized-system validation apply to an AI tool the same way they apply to any other electronic system that creates, modifies, or manages trial records.
What follows is a practical reading of the E6(R3) requirements that are most relevant to sponsors using AI tools in their clinical data management pipeline.
Part 11 asks whether the electronic record can be reconstructed. This note asks the GCP question: did a qualified person oversee the process, with governance and validation matched to the risk. Use the Part 11 checklist for the record. Use this page for the oversight. The Trust Center states the public posture: architected for E6(R3), a validation package on request, and no certification we do not hold.
What changed from E6(R2)
E6(R3) reorganized the guideline. The change that matters for clinical data management is section 4, Data governance, investigator and sponsor. E6(R2) treated data handling inside the sponsor and investigator sections. Section 4 is where E6(R3) states the data life cycle, the metadata, and the expectations for computerized systems.
The specific additions relevant to AI-assisted processes:
Data governance framework. Sponsors must have documented processes for ensuring data quality throughout the trial lifecycle. For AI-generated data artifacts, this means the governance framework must address how AI outputs are generated, reviewed, and approved, not just how data is entered by site staff.
Computerized system validation. Section 4.3.4 says the approach to validation is based on a risk assessment that considers the intended use of the system and the importance of the record it holds. For AI systems that generate SDTM mappings or Define-XML, that intended use is a submission record, so the validation has to match.
Audit trails. Section 4.2.2 covers relevant metadata, including audit trails. The glossary defines an audit trail as the metadata that shows the activity, the initial entry and later changes, by whom, when, and, where applicable, why. Section 4.2.2 says systems document the initial entry and any later change or deletion, including the reason where appropriate, and that audit trails are not disabled. For an AI system, the actor can be the computerized system that made the change, which section 4.2.4 already names, and the action includes the inference that produced the output.
The human oversight requirement
Sponsor oversight is section 3.9. The sponsor ensures that the processes undertaken, and the information and data generated, are of sufficient quality for reliable results and participant safety. Oversight of investigators and service providers is part of that duty. Section 4.1 is a different provision. Its title is "Safeguard blinding in data governance." It does not assign sponsor responsibility for data integrity, and Annex 1 has no section 4.9.
The practical implication sits in the data life cycle. Section 4.2.3 expects a planned review of trial data, audit trails, and other relevant metadata. Section 4.2.4 says a correction is attributed to the person or the computerized system that made it, justified, and supported by source records. For an SDTM mapping, a DMP entry, or a Define-XML definition that will sit in the trial record, a qualified person's review should be documented, attributed, and timestamped. The audit trail for that review is the record the glossary and section 4.2.2 describe.
Risk-based validation for AI systems
One of E6(R3)'s central contributions is the formalization of risk-based approaches to GCP compliance. For computerized systems, this means validation effort should be scaled to the risk level of the system's intended use. A system that produces outputs that go directly into regulatory submissions without additional human verification requires more rigorous validation than a system used for internal planning.
For AI-assisted SDTM mapping, the risk level is high. The outputs directly influence the submission package. A mapping error that survives to submission can cause a Pinnacle 21 failure or trigger a deficiency letter. Under E6(R3)'s risk-based framework, this use case requires a validation package that specifically addresses the AI components' accuracy, consistency, and behavior under edge cases, not just the application infrastructure they run on.
The validation package should include: a definition of intended use and scope, qualification testing against known cases with verified correct outputs, a description of confidence scoring and how it relates to the human review decision, and a documented process for updating the validation when the underlying model is changed.
What "data integrity" means for AI-generated records
Section 4 is the data-governance section. For an AI-generated record, the part a reviewer can check is the chain of custody: a value proposed by an agent, revised by a DM lead, and approved by a qualified person has to show that chain, not only the final value.
This is what distinguishes an audit-trail-compliant AI system from a system that merely produces correct outputs. Correct outputs without the metadata in section 4.2.2 leave an inspector unable to see who changed the record and why. The chain of custody from AI proposal to human approval to committed record is the compliance story. It has to be documentable on request.
What an inspector should be able to find
Read together, the glossary and section 4.2.2 ask for who acted, what changed, when, and, where a change was made, why. Section 3.9 keeps sponsor oversight on the process and the data. Section 4.1 is the blinding safeguard. Applied to an AI-assisted artifact, that is a short list. If a row is missing, the oversight story is not finished.
| Show this | What documented means |
|---|---|
| Who | The agent that proposed the value, and the qualified person who approved it. The model is not an attribution. |
| What | The inference, any human revision, and the approval. A trail of clicks that skips the AI proposal does not record the action that created the record. |
| When | A timestamp on each of those events, including the model version that was current then. |
| Why | The source passages, the controlled terminology consulted, and the text of a revision. Provenance, not a confidence number on its own. |
| Oversight | The approval is structural. Section 3.9 keeps the sponsor responsible for the quality of the process and the data. A configuration that lets the queue be skipped leaves that oversight undocumented. |
| Validation | Section 4.3.4 scales validation to intended use and to the importance of the record. State the intended use, the tests against known-correct cases, how confidence relates to the review decision, and what is revalidated when the model changes. SDTM and Define-XML are high-risk uses. The package has to say so. |
The electronic-records version of this list is the Part 11 checklist. How the trail is stored is building an audit trail you would show a regulator. Where the gate sits in the product is human-in-the-loop, on Core.
Practical steps for sponsors using AI DM tools
Given the E6(R3) requirements, sponsors evaluating or currently using AI tools in their data management pipeline should verify the following:
First, confirm that the AI tool's audit trail captures agent-level operations, not just human interactions. Ask the vendor for an example audit report from a study showing both agent inference records and human approval records.
Second, confirm that a validation package exists for the AI components. The validation package should address the system's intended use in SDTM mapping, Define-XML generation, or whatever data management function it performs, not just the application layer.
Third, confirm that the human approval step is architecturally enforced. A system where the approval step can be bypassed by configuration does not meet E6(R3)'s human oversight requirement: the oversight must be structural, not optional.
Fourth, document the data governance for AI-generated artifacts. The document should describe how AI outputs are generated, reviewed, and approved, and how the chain of custody is maintained from proposal to committed record. That is the record an inspector follows under section 4, and under section 4.2.3 for review of data and metadata.
E6(R3) does not prohibit AI in clinical data management. It sets the bar for what compliant AI-assisted data management looks like. The bar is achievable, but it requires AI tools that were designed for the regulated environment, not tools designed for general productivity that have been deployed in a regulated context without the necessary compliance architecture.