Effective dates, as published
E6(R3) has three parts: overarching principles, Annex 1 for interventional clinical trials, and Annex 2 with additional considerations for trials with decentralised or pragmatic elements and real-world data sources. The dates below are as stated by ICH, EMA and FDA. FDA's notice for the final guidance gives no separate effective date, and FDA guidance documents state that they contain nonbinding recommendations.
| Body | Principles and Annex 1 | Annex 2 |
|---|---|---|
| ICH | Step 4: endorsed by the Regulatory Members of the ICH Assembly and adopted on 6 January 2025 | Step 4: endorsed on 3 June 2026; ICH's consolidated guideline combining the principles and both annexes is dated 16 June 2026 |
| EU (EMA / CHMP) | Final adoption by CHMP 12 December 2024; came into effect 23 July 2025 | Final adoption by CHMP 25 June 2026; comes into effect 15 January 2027. EMA also lists a consolidated Step 5 version (EMA/CHMP/ICH/162879/2026) with a legal effective date of 15 January 2027 |
| US (FDA) | Final guidance for industry, dated September 2025; Federal Register notice of availability 9 September 2025 (90 FR 43460, docket FDA-2023-D-1955) | Draft guidance announced 30 December 2024 (docket FDA-2024-D-5601), comment period reopened 13 March 2025. We found no Federal Register notice of a final FDA Annex 2 as of 4 November 2026; check FDA's guidance pages before relying on this |
Why E6(R3) reaches AI tools even though it never names them
We searched the ICH Step 4 text and the June 2026 consolidated guideline: neither uses the terms artificial intelligence or machine learning. FDA's announcement says the revision "incorporates flexible, risk-based approaches and embraces innovations in trial design, conduct, and technology." The guideline governs AI through the general categories it already defines.
Two of those categories matter. First, computerised systems: principle 9.3 says "Computerised systems used in clinical trials should be fit for purpose (e.g., through risk-based validation, if appropriate)", and section 4.3 sets out procedures, training, security, validation and user management for them. E6(R3) does not define the term, but in our reading an AI tool that captures, reviews, analyses or changes trial data, or produces trial records, is a computerised system used in the trial. Second, service providers: the glossary defines a service provider as a person or organisation "providing a service used by either the sponsor or the investigator to fulfil trial-related activities". A vendor that provides an AI-assisted service or a hosted AI tool for trial work generally fits that definition; licensed software the sponsor runs itself is a less clear case.
The mapping below is our reading of how the guideline's text applies to an AI tool or agent. It is not ICH, FDA or EMA guidance on AI, and how much of it applies to a given tool depends on that tool's risk. It covers the principles and Annex 1 only; it does not address Annex 2's additional considerations for decentralised and pragmatic trials or real-world data sources.
Requirement to what it means for an AI tool
Section numbers refer to the principles and Annex 1 of the ICH E6(R3) Step 4 guideline of 6 January 2025; Annex 2 has its own numbering. FDA's version uses its own outline and US spelling; for Annex 1 it shows the ICH section number in parentheses after each heading, and the principles keep ICH's numbering. The left column summarises the guideline, which is written in terms of what parties should do; the right column is our interpretation for an AI tool used in trial operations.
| E6(R3) requirement | What it means for an AI tool |
|---|---|
| Principle 9.3 and 4.3.4(a): computerised systems should be fit for purpose; the validation approach should rest on a risk assessment of intended use, the importance of the data, and the system's potential to affect participants and the reliability of results | Write down the tool's intended use in the trial before choosing how much to validate. A tool that drafts text a qualified person rewrites and signs carries different risk from one that decides which data get queried |
| 4.3.4(a): "The responsible party is responsible for the validation status of the system throughout its life cycle." 4.3.4(g): this includes systems developed by other parties, and the responsible party should ensure validation documentation is maintained and retained | Buying an AI tool from a vendor does not transfer the validation obligation. The sponsor or investigator still needs evidence the tool is fit for purpose for this trial and, under 4.3.4(g), should ensure the validation documentation is maintained and retained |
| 4.3.4(c) and (d): systems should be validated prior to use; later changes should be validated based on risk, considering both previously collected and new data, in line with change control; periodic review may be appropriate to keep systems in a validated state | A new model version, a changed system prompt or standing instruction set, or a changed configuration is a system change to assess under change control. If an update re-processes earlier data, consider its effect on data already collected. Plan periodic review, because a model provider can change behaviour between versions |
| 4.3.4(e): standard system functionality and protocol-specific configurations and customisations, including automated data entry checks and calculations, should be validated; interfaces between systems should be defined and validated; different degrees of validation may be needed for bespoke, configurable and unaltered systems | We treat the tool's trial-specific set-up (rules, system prompts, thresholds) as protocol-specific configuration to validate, to a degree that fits the kind of system it is. Its connections to EDC, CTMS, eTMF or safety systems are interfaces to define and validate |
| 4.2.2: for metadata associated with data of higher criticality, the responsible party should evaluate the system so that, among other things, "Systems record and maintain workflow actions in addition to direct data entry/changes into the system", and should ensure audit trails and logs are not disabled and are interpretable | Where the tool touches data of higher criticality, consider logging what it read, what it proposed or changed, which version ran, and who accepted or rejected the output, in a form a reviewer can follow |
| 4.2.4: "Corrections should be attributed to the person or computerised system making the correction", justified and supported by source records | If an AI tool applies a correction, the correction should be attributed to the tool and justified. If it only proposes one, the person who makes the correction is the one attributed. In our reading, recording the tool's involvement and who approved its proposal is good practice beyond what 4.2.4 states |
| 4.2.3: review of trial data, audit trails and other relevant metadata is planned and risk-based | Include the tool's logs and outputs in the planned review, weighted to the data the tool touches that matter most |
| 4.3.8: access controls limit access to authorised users and ensure attributability to an individual; permissions follow duties, functions, blinding arrangements and organisation, and are revoked when no longer needed. 4.1.2: for example, in blinded trials, sponsor staff or service providers involved in operating the trial who directly or indirectly interact with investigator site staff should not have access to unblinding information except when justified by the trial design | Give the tool its own scoped access rather than a shared account, and keep each tool action attributable to the individual who ran, configured or approved it. In a blinded trial, check that the tool cannot read treatment allocation and then surface it, directly or indirectly, to people who should stay blinded, for example in a summary or draft |
| 4.2.5 and 4.3.3: data transferred between systems keeps its integrity and confidentiality; security controls are maintained across the data life cycle | Sending trial data to an externally hosted model is a data transfer. Document where data go, how integrity and confidentiality are kept, and what the provider may retain |
| 3.11.4.2: centralised monitoring is an evaluation of accumulated data "by the sponsor's qualified and trained persons"; centralised data analytics can help identify systemic or site-specific issues. 3.11.4.3: the monitoring plan should describe the monitoring methods and tools to be used and the rationale for their use | An AI tool can support centralised monitoring analytics. Name it in the monitoring plan with the reason for using it. Our reading is that the evaluation and resulting decisions stay with qualified, trained people, whose review is documented |
| 3.6.4, 3.6.6 and 3.6.7: transferred activities are documented in an agreement; ultimate responsibility stays with the sponsor; the sponsor assesses the provider's suitability | An AI vendor or a CRO using AI for trial work should have an agreement that names the activities. Assess the vendor's suitability for the specific use as part of selecting it |
| 3.6.8 and 3.6.9: the sponsor should have access to relevant information such as SOPs and performance metrics, and oversee important transferred activities, "including activities further subcontracted by the service provider" | Ask which AI tools a provider uses on your trial and on what model, and for its procedures and performance information. Consider whether a model provider behind the vendor amounts to a further subcontracted activity |
| 2.3.1 and 2.3.3: the investigator retains ultimate responsibility for delegated activities, oversees them proportionately, and keeps a record of persons and parties to whom activities are delegated | A site that uses an AI tool for trial tasks keeps responsibility for the output. The tool is not a delegate: the delegation record covers the persons and parties who use it, documented in proportion to the significance of the activity |
Responsibility does not move to the tool or the vendor
Principle 10.2 states the point plainly: "Where activities have been transferred or delegated to service providers, the responsibility for the conduct of the trial, including quality and integrity of the trial data, resides with the sponsor or investigator, respectively." Section 3.9.5 adds that the selection and oversight of service providers are fundamental features of sponsor oversight, and that oversight includes quality assurance and quality control of their trial-related activities.
E6(R3) also allows proportion. Section 3.6.10 says a service provider's existing quality management processes may meet the relevant GCP requirements even if they were not designed specifically to be GCP-compliant, provided they are fit for purpose in the context of the trial. For an AI vendor, that means the question is whether its development, testing and change processes are adequate for the use you intend, not whether it carries a GCP label.
Questions to put to a service provider that uses AI
Sections 3.6.7 to 3.6.9 ask the sponsor to assess a provider's suitability, to have access to information such as SOPs and performance metrics, and to oversee important activities, including subcontracted ones. Section 3.6.6 asks providers to report incidents that might affect participant safety or trial results. For a CRO, vendor or technology provider that uses AI on your trial, those duties translate into questions like these. They are our suggestions, not a checklist from the guideline.
- Which trial-related activities on this trial use an AI tool, and which tool and model version performs each one?
- What is each tool's intended use, and what testing shows it is fit for that use? Can we see the validation documentation, or a summary of it?
- How are changes to the model, system prompts or standing instructions, and configuration controlled, and will you tell us before a change that affects our trial?
- What trial data does the tool receive, where is it processed, and does any third-party model provider retain or reuse it?
- How are the tool's outputs and actions logged, and can our monitors, auditors and inspectors see those logs?
- Who reviews the tool's output before it affects trial data or decisions, and how is that review recorded?
- How will you report an AI-related incident that might affect participant safety or trial results?
What investigators and sites should control
E6(R3) gives investigators the same structure on a smaller scale. Section 2.3.1 says the investigator may delegate trial-related activities to other persons or parties but retains ultimate responsibility and should oversee them in proportion to the importance of the data and the risks to participants and data reliability. Section 2.3.4 says agreements with service providers for trial-related activities should be documented.
For a site, the practical points are: know which AI tools staff use for trial tasks, keep trial data out of tools the site has not approved for that purpose, make sure a qualified person reviews AI output before it enters a source record or a trial system, and, by analogy with the delegation records in 2.3.3, document the tool's use in line with the significance of the activity. Section 2.3.3 notes that delegation documentation may not be required where activities are performed as part of clinical practice. Where the sponsor identifies a service provider, such as an AI vendor, to carry out activities under the investigator's responsibility, section 3.6.5 says the sponsor should give the investigator information on that provider, and under 2.3.1 the investigator makes the final decision on whether it is appropriate.
Three uses, three levels of control
These examples illustrate the risk-proportionate approach; they are our examples, not cases from the guideline. In each, the deciding factors are the ones 4.3.4(a) names: intended use, importance of the data, and potential to affect participants or the reliability of results.
- Document drafting. An AI tool drafts a monitoring visit report or a section of a trial document, and a qualified person reviews, edits and approves it. The person, not the tool, owns the content, and site monitoring reports are among the essential records listed in Appendix C. Controls centre on that review, on keeping the version the person approved, and on what trial data the tool is given.
- Data review and centralised monitoring. An AI tool screens accumulated data and flags possible issues for data managers or monitors. Missing a signal or raising a false one affects data quality decisions, so expect documented testing against the intended use, logged outputs, version control, a mention in the monitoring plan, and a person evaluating each flag. Centralised monitoring reports are also among the essential records listed in Appendix C.
- Writing into a system of record. An AI tool proposes or applies corrections in EDC or another trial system. This touches data the trial depends on, so expect the fullest controls: validation of the configuration and interface, attribution of each correction to whoever made it with any approval recorded, justification, and change control for any model or system-prompt update. Under 3.16.1(i), the sponsor should not make changes to data entered by the investigator or trial participants unless justified, agreed in advance by the investigator and documented, and 3.16.1(l) says the sponsor should not have exclusive control of data captured in data acquisition tools. Using an AI tool does not change either point.
How this fits with Part 11, AI validation and FDA's credibility framework
E6(R3) is a good clinical practice guideline. It sits alongside, and does not replace, the record and signature rules in 21 CFR Part 11 for in-scope US records, the general practice of computerised system validation, and FDA's separate draft framework for establishing the credibility of AI models whose output supports regulatory decisions. Those topics have their own answers on this site, linked below; this page covers only what E6(R3) adds for trial operations.
See: Does 21 CFR Part 11 apply to AI tools? · How do you validate an AI agent used in a GxP process? · How does FDA's AI credibility assessment framework work?
Where to start
List every AI tool used on a trial, by the sponsor, by sites and by service providers, with its intended use and the trial data it touches. Rank them by the risk factors in 4.3.4(a), then apply the controls in the table in proportion. Make sure agreements with service providers name AI-assisted activities. This page is a reading of public guidance, not legal or regulatory advice; decisions on GCP compliance, validation and regulatory strategy belong to your clinical quality and regulatory functions.
Knowing which AI tools touch trial data, and under whose control, is where E6(R3) oversight starts. Atlas Bio's Secure AI for Life Sciences starts with an AI risk assessment of current AI use and data exposure. Status, October 2026. In use in Atlas Bio's own agent platform today: a separate verification step that checks agent work, and a hold that stops a task until a person decides. Built and tested on synthetic data, not yet used with a client: the AI risk assessment tooling and a record of what data an AI used, what it produced and who reviewed it. In development: a customer-controlled AI environment in your own Azure tenant, an AI gateway, agent permissions, evidence linkage and notifications. Each engagement states which controls are built. Decisions on GCP compliance, validation and regulatory strategy stay with your clinical quality and regulatory functions.
Does ICH E6(R3) mention artificial intelligence?
No. Neither the January 2025 Step 4 guideline nor the June 2026 consolidated guideline uses the term. In our reading, AI tools come within the guideline's general provisions on computerised systems, data governance and service providers.
When did ICH E6(R3) take effect?
ICH adopted the principles and Annex 1 on 6 January 2025. In the EU they came into effect on 23 July 2025. FDA published them as final guidance in September 2025, with a Federal Register notice on 9 September 2025. Annex 2 reached ICH Step 4 on 3 June 2026 and comes into effect in the EU on 15 January 2027; FDA had issued it only as a draft as of 4 November 2026.
Is an AI vendor a service provider under ICH E6(R3)?
If it provides a service used by the sponsor or investigator to fulfil trial-related activities, it generally fits the guideline's definition of a service provider. The activities should be documented in an agreement, the sponsor should assess the vendor's suitability and oversee its work, and responsibility for trial data stays with the sponsor or investigator.
Does every AI tool used in a trial need full validation?
No. E6(R3) asks for risk-based validation, scaled to the tool's intended use, the importance of the data and its potential to affect participants or the reliability of results, and 4.3.4(e) says different degrees of validation may be needed for different kinds of system. In our reading, a drafting aid whose output a qualified person reviews needs less than a tool that changes data in a trial system.
Can an AI tool perform centralised monitoring under ICH E6(R3)?
E6(R3) describes centralised monitoring as an evaluation of accumulated data by the sponsor's qualified and trained persons, and says centralised data analytics can help identify issues. An AI tool can support that analysis; our reading is that the evaluation and decisions stay with qualified people.
Does ICH E6(R3) replace 21 CFR Part 11?
No. E6(R3) is a good clinical practice guideline; Part 11 is a US regulation on electronic records and signatures and still applies to in-scope records. See our separate answer on Part 11 and AI tools.
- ICH E6(R3) Guideline for Good Clinical Practice, Step 4, adopted 6 January 2025 (principles and Annex 1, PDF)
- ICH E6(R3) Guideline for Good Clinical Practice, Step 4 consolidated guideline (principles, Annex 1 and Annex 2), June 2026 (PDF)
- EMA: ICH E6 Good clinical practice - Scientific guideline (effective dates for the principles, Annex 1 and Annex 2)
- EMA: ICH E6(R3) Guideline for good clinical practice, Step 5, EMA/CHMP/ICH/135/1995 (PDF)
- EMA: ICH E6(R3) Guideline for good clinical practice - Annex 2, Step 5, EMA/CHMP/ICH/495903/2024 (PDF)
- FDA: E6(R3) Good Clinical Practice (GCP), final guidance, September 2025 (guidance page)
- FDA: E6(R3) Good Clinical Practice (GCP), final guidance (PDF)
- Federal Register, 9 September 2025: E6(R3) Good Clinical Practice; Guidance for Industry; Availability (90 FR 43460)
- Federal Register, 30 December 2024: E6(R3) Good Clinical Practice: Annex 2; Draft Guidance for Industry; Availability
- Federal Register, 13 March 2025: E6(R3) Good Clinical Practice: Annex 2; Draft Guidance; Reopening of the Comment Period
Talk to Atlas Bio.
Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.
Contact Us →