Part 11 does not regulate AI as a technology. It regulates electronic records and electronic signatures that FDA rules require you to keep or that you submit to the agency. So an AI tool falls under 21 CFR Part 11 when it creates, modifies, maintains or transmits records of that kind, or when your team relies on its output as the record of a regulated activity. If it only helps someone think, search or brainstorm, and nothing it produces becomes a required record, Part 11 is usually not the question. What you must be able to show is the scope decision itself.
This guide walks through that decision, the controls Part 11 asks for once an AI tool is in scope, and a sequence your clinical, regulatory and quality teams can follow without treating every AI tool as a validation project.
TL;DR
- Part 11 applies to electronic records required by FDA predicate rules (GMP, GCP and similar) or submitted to FDA, and to the signatures on them. It does not apply to a tool simply because the tool uses AI.
- Decide scope per use case, not per product, and write the decision down.
- For in-scope systems, expect validation for intended use, secure time-stamped audit trails, access controls, record retention and signature-to-record linking.
- Validate on risk: intended use, data importance, and effect on participant safety, product quality and record integrity.
- A cloud or third-party AI vendor does not move your responsibility. You still have to show records stay authentic, secure and available for inspection.
Table of Contents
- When Part 11 applies to an AI-enabled system
- Predicate rules: the part teams skip
- What Part 11 requires once a system is in scope
- Validating an AI-enabled system
- Managing model changes and AI-generated records
- Cloud-hosted and third-party AI
- A practical sequence for your team
- Our perspective
- Where we can help
- FAQ
- Sources
When Part 11 applies to an AI-enabled system
The regulation's own scope text is the starting point. Part 11 covers electronic records that are created, modified, maintained, archived, retrieved or transmitted under FDA record requirements, and electronic records submitted to the agency under the FD&C Act or the PHS Act. Nothing in that language mentions a technology type. A rules-based spreadsheet, a laboratory system and a large language model are all judged the same way: what record does it touch, and is that record required or relied on?
FDA's October 2024 guidance on electronic systems in clinical investigations (Revision 1) puts it in practical terms. Part 11 applies to records required under predicate rules and kept electronically in place of paper, relied on for regulated activities, or submitted electronically. Examples it gives include records under 21 CFR Parts 312 and 812, the investigational drug and device rules.
That gives you a simple test with three questions:
- Is there a predicate rule that requires this record or this signature?
- Do we keep or rely on the record in electronic form instead of paper?
- Does the AI tool create, change, or decide the content of that record?
If the answer to all three is yes, the AI-enabled system is in scope for that use. If the answer to the first is no, Part 11 is not triggered for that record, although good practice may still apply.
Same tool, different scope
Scope belongs to the use, not the product. Consider one drafting assistant used in two ways:
- A regulatory writer asks it to summarize literature for an internal brainstorm. Nothing it outputs is a required record. The summary is a working note.
- The same writer uses it to draft sections of a document that will be filed, then approves the text. The final document is a regulated record, and the system that produced and stored it, along with the approval signature, sits in the Part 11 conversation.
The second case does not mean you cannot use AI. It means the way the AI is used must leave a record you can defend. Our guide to AI in regulatory document drafting covers where drafting support fits and where human authorship has to stay in place.
Predicate rules: the part teams skip
Predicate rules are the underlying FDA requirements, such as good manufacturing practice or good clinical practice, that say which records you must create and keep. FDA's guidance describes Part 11 as the layer that gives the criteria for when those electronic records and signatures are valid, and it says electronic records and signatures that meet Part 11 are considered equivalent to paper records and handwritten signatures, as established in the 1997 final rule.
This matters for AI because most of the confusion comes from skipping the predicate step. A team hears "AI" and "FDA" and assumes full validation of everything. FDA's 2003 scope and application guidance points the other way: determine which records are subject to predicate rules and whether you rely on the electronic versions to carry out regulated activities. That guidance sets out a narrower reading of Part 11, and it expects you to base decisions on risk.
Write a one-page scope record for each AI use case. It should name the predicate rule, the record, whether the electronic version is the one you rely on, and the person who made the call. When an inspector asks why a tool was not validated to the same level as your clinical data system, that page is the answer.

What Part 11 requires once a system is in scope
The controls are the same ones any electronic system must meet. What changes with AI is how you prove them. The table maps each requirement to the question an AI tool raises.
| Part 11 control | What it requires | The AI-specific question |
|---|---|---|
| System validation (11.10(a)) | Validation that the system performs as intended | What is the intended use, and how do you show the model performs consistently for it? |
| Audit trails (11.10(e)) | Secure, computer-generated, time-stamped trail that does not obscure earlier entries | Can you tell machine-suggested content from human edits and approvals? |
| Access controls | Only authorized individuals can use the system and change records | Who can change prompts, configuration, or model versions? |
| Record retention | Records stay accurate and retrievable for the required period | Are prompts, outputs and versions retained along with the final record? |
| Signature manifestation (11.50) | Printed name, date and time, and the meaning of the signature | Does the approval screen make clear the human is approving the AI-assisted content? |
| Signature linking (11.70) | Signatures cannot be excised, copied or transferred to falsify a record | Is the signature bound to the exact version of the record that was approved? |
The full text of Part 11 on the eCFR is short and worth reading in full. It requires validation of systems to ensure intended performance, secure time-stamped audit trails for changes, access controls and record retention, and it requires signed records to show the signer's name, the date and time, and the meaning of the signature, such as review, approval or authorship. Signatures must be linked to their records so they cannot be removed or reused to falsify a record.
Audit trails for AI-assisted work
The audit-trail rule asks that changes to electronic records do not obscure previously recorded information, and that entries are time-stamped. For AI, the practical reading is that a reviewer should be able to reconstruct who or what produced each change: the model output, the human edit, the approval. If your tool overwrites a draft and keeps only the final version, you have a gap that no amount of policy writing closes.
We go into design choices, including tamper-evident approaches, in our piece on inspection-ready FDA audit trails.
Electronic signatures and human review
A signature means a person took responsibility. When AI helped produce the content, the signer still carries that responsibility, and the signature meaning (review, approval, authorship) should reflect what the person actually did. Build the workflow so that approval is a deliberate step on a specific version, not a click-through at the end of a generated draft.
Validating an AI-enabled system
FDA's 2024 guidance recommends risk-based validation that documents a system can consistently fulfill its specified requirements throughout its life cycle. The factors it names are the intended use, the purpose and importance of the data, and the potential effect on participant safety and the reliability of trial results. It also expects you to document the validation and the system requirements.
For an AI system, a workable validation file usually contains:
- Intended use statement. What the tool does, for which records, and what it must never do without human review.
- Requirements. Accuracy expectations, access rules, audit-trail behavior, retention, and how outputs reach the record.
- Risk assessment. What happens if the output is wrong, and who would catch it. A tool that proposes wording for a person to rewrite carries a different risk than one that populates fields in a submission.
- Test evidence. Representative test cases drawn from your real documents and data, with acceptance criteria set before testing.
- Explainability notes. How a reviewer can see why the system produced an output, so "the model suggested it" is never the only rationale. Our article on model explainability for U.S. regulators goes deeper.
- Change control. What counts as a change, who approves it, and what re-testing follows.
FDA's 2003 guidance adds a useful point: even where no predicate rule requires validation, it may still be important to validate a system given its potential effect on product quality, safety and record integrity. The same guidance describes enforcement discretion for specific requirements on validation (11.10(a)) and audit trails (11.10(e)) in certain situations, but it is explicit that you must still meet the documentation and data integrity requirements of the applicable predicate rules. Treat discretion as a reason to size your effort to risk, never as a reason to skip records.
If your quality group works to a GxP lifecycle, our seven-step GxP validation lifecycle for AI shows how these pieces fit together.
Managing model changes and AI-generated records
The hardest AI-specific problem is that the system can change without anyone editing a line of code. A vendor updates a model, a retrieval source changes, a prompt template gets tweaked. Each can alter outputs on records you rely on.
Treat these as controlled changes:
- Pin and record versions. Know which model, configuration and prompt template produced each output, and keep that information with the record.
- Define what triggers re-validation. A new model version used on regulated records should trigger a documented impact assessment, scaled to risk.
- Separate suggestion from decision. Keep machine-generated text visibly distinct from human-authored and human-approved text in the record history.
- Set human review expectations in the procedure. State who reviews, what they check, and what they sign.
- Monitor performance after release. Validation is not a one-time event. Sample outputs, track errors and feed findings back into change control.
These are our recommendations for working practice, built on the audit-trail and validation requirements above, not additional regulatory text.
Cloud-hosted and third-party AI
Most biotech teams use AI through a vendor's cloud service. Part 11 does not exempt you for that. FDA's guidance says regulated entities should ensure cloud-hosted records remain authentic, integral, confidential, secure, traceable and available for inspection, and that you may review an IT service provider's documentation to judge whether its system is fit for purpose.
In practice, that means your vendor review should ask for:
- Documentation of how the service handles audit trails, access control and retention.
- How and when models and configurations change, and how you will be told.
- Where records live and how you can retrieve them during an inspection.
- Your ability to export records and their history in a usable form.
If your records include protected health information, there is a separate question about privacy rules. Our article on why no AI tool is HIPAA compliant by itself covers that side.
A practical sequence for your team
- List your AI use cases. Not tools. Uses, such as drafting a document section, triaging monitoring signals or summarizing site feasibility data.
- Run the three-question scope test on each, and record the predicate rule where there is one.
- Sort by risk. Consider the effect on participant safety, product quality and record integrity, plus how likely a human is to catch an error.
- Write the intended use and requirements for each in-scope system.
- Validate in proportion. Heavier evidence for high-risk uses, lighter for low-risk ones, all documented.
- Wire in audit trails and signatures that separate machine and human contributions and bind approvals to a specific version.
- Put change control and monitoring in place before go-live, including vendor model updates.
- Train reviewers on what they are signing and what to look for in AI output.
- Review the scope record periodically and whenever a use expands.
For teams working in clinical operations, a bounded pilot is often the cleanest way to test this sequence. See how we structure one in our 90-day partnered pilot for AI-enabled risk-based monitoring.
Our perspective
Most Part 11 trouble with AI starts upstream of the software. A team buys a tool, then asks what compliance requires. We work the other way. We start from the milestone you need to reach, trace back through the regulatory and clinical steps that stand in the way, and only then decide where AI belongs and what it must leave behind as a record.
That order changes the outcome. When the workflow is designed first, the audit trail, the human review step and the signature meaning are part of the process from day one, not a retrofit before an inspection. It also keeps the scope honest: the tasks that touch regulated records get governed automation, and the tasks that do not stay light.
We also believe adoption is a compliance control. A well-validated system that reviewers do not understand will still produce weak records. Our approach embeds senior expert teams in your operations, redesigns the workflow around your regulatory and scientific needs, and stays with you through governance and monitoring, so the controls are used and not just documented.
John
Where we can help
If you are deciding which AI uses are in scope, how to validate them, and how to keep records inspection-ready without slowing your timeline, that is the work we do with biotech and life sciences teams. We start with a diagnostic of where your operational bottlenecks are, then redesign the affected workflows with governance built in, including regulatory drafting and clinical site activation.

To talk about an engagement, see our partnership page, or look at the sectors we work in. For a view of where AI helps in regulatory work and where people must stay in charge, read our regulatory affairs overview.
FAQ
Does 21 CFR Part 11 apply to every AI tool we use?
No. It applies when the tool creates, modifies, maintains or transmits electronic records required by an FDA predicate rule, or records submitted to FDA, and when you rely on those electronic records. A tool used for general research or brainstorming, whose output never becomes a required record, is generally outside that scope. Document the reasoning so you can show it later.
Do we have to fully validate an AI system under Part 11?
Validation for intended performance is a Part 11 requirement for in-scope systems, and FDA recommends a risk-based approach. That means the depth of evidence depends on intended use, the importance of the data and the potential effect on safety, quality and record integrity. Even where no predicate rule requires it, FDA says validation may still be important for those reasons.
What should an AI audit trail capture?
Part 11 requires secure, computer-generated, time-stamped audit trails that do not obscure previously recorded information. For AI-assisted records, we recommend also capturing the model and configuration version and separating machine-suggested content from human edits and approvals, so every change is attributable.
Are electronic signatures still valid when AI helped draft the record?
Yes, provided the signature meets Part 11: it shows the signer's printed name, the date and time, and the meaning of the signature, and it is linked to the record. The signer remains responsible for what they approve, so the workflow should make clear which version was reviewed and approved.
Does using a cloud AI vendor shift our compliance responsibility?
No. FDA expects regulated entities to ensure cloud-hosted records stay authentic, secure, traceable and available for inspection, and allows you to review a provider's documentation to judge fitness for purpose. Build vendor review and change notification into your agreement and procedures.
Sources
- FDA: Electronic Systems, Electronic Records, and Electronic Signatures in Clinical Investigations (Revision 1, October 2024)
- eCFR: 21 CFR Part 11, Subpart A, General Provisions
- eCFR: 21 CFR Part 11, Electronic Records; Electronic Signatures
- FDA: Part 11, Electronic Records; Electronic Signatures, Scope and Application
