Part 11 sets the rule, not the number
21 CFR Part 11 does not contain a retention period of its own. Two of its closed-system controls deal with retention. Section 11.10(c) requires protection of records to enable their accurate and ready retrieval throughout the records retention period. Section 11.10(e) requires secure, computer-generated, time-stamped audit trails to independently record the date and time of operator entries and actions that create, modify or delete electronic records, says that record changes shall not obscure previously recorded information, and says that audit-trail documentation shall be retained for a period at least as long as that required for the subject electronic records and be available for FDA review and copying.
The period itself therefore comes from the predicate rule: the FDA regulation that requires the record in the first place. FDA's 2003 guidance on the scope and application of Part 11 says the agency intends to exercise enforcement discretion on the Part 11 requirements for validation, audit trails, record retention and record copying, including 11.10(c) and the audit-trail requirements of 11.10(e). It is explicit that this does not relax the underlying rules: persons must still comply with all applicable predicate-rule requirements for record retention and availability, and that FDA can take regulatory action for noncompliance with those predicate rules. It also notes that Part 11 remains in effect and that the enforcement discretion applies only as identified in the guidance. The guidance suggests basing decisions on how to maintain records on the predicate rule, on a justified and documented risk assessment, and on a determination of the value of the records over time. The same guidance also narrowed which records are subject to Part 11: whether an agent log is itself a Part 11 record depends on whether a predicate rule requires it and whether you rely on it to perform regulated activities. The predicate-rule retention period for the underlying record applies either way.
For an AI agent this has a practical consequence. Where an agent creates, modifies or deletes an electronic record that a predicate rule requires, we recommend treating the entries that show what it did as part of that record's audit trail or supporting metadata, and keeping them for the record's retention period. Whether a particular agent log is itself the Part 11 audit trail depends on system design, for example whether the regulated system's own audit trail already captures the change, and should be decided and documented in validation. A log the agent writes about its own actions may not meet the expectation that an audit trail records actions independently; the regulated system's own audit trail is normally the primary trail. FDA's guidance says that, even where it does not intend to enforce 11.10(e), persons must still comply with predicate-rule requirements on the date, time or sequencing of events and on changes not obscuring previous entries, citing 21 CFR 58.130(e) as an example for dates, and that audit trails can be particularly appropriate when users are expected to create, modify or delete regulated records during normal operation. A log deleted on a short operational schedule while the record it explains is kept for years can leave an inspector unable to reconstruct how the record came to be. Where an agent only reads or proposes, and a person or another system makes the change, its log is, in our reading, generally not the 11.10(e) audit trail. Its output may still be a record in its own right, for example GLP raw data or a derived result you rely on, and then carries that record's retention period. Either way it is evidence of how the record came to be; our companion page on audit trails for AI agents explains why that evidence matters for reconstruction.
US predicate-rule retention periods
The table summarises the retention clauses most often relevant to research and development work. It paraphrases the eCFR text as published on 20 October 2026; the linked section is the authority, and other FDA regulations may apply to your records. Section 58.195(a) itself says its periods do not supersede retention requirements in other regulations in the chapter. The fourth column is our reading, not regulatory text. This page does not cover the separate retention rules for licensed biologics under 21 CFR Part 600 (see 600.12), for advanced therapy medicinal products under EU product legislation, or for EU and OECD GLP; check those where they apply.
Two scope notes for early development. Under 21 CFR 210.2(c), production of an investigational drug for a phase 1 study is exempt from Part 211, though it remains subject to the statutory CGMP requirement in 21 U.S.C. 351(a)(2)(B); the exemption ends once the investigational drug has been made available for use by or for the sponsor in a phase 2 or 3 study, or has been lawfully marketed. Investigational new drug products are also exempt from expiration dating under 21 CFR 211.137(g) if they meet appropriate standards shown by stability studies during clinical use; where a phase 2 or 3 batch has no labelled expiry, the 211.180(a) clock has no fixed start. We suggest setting the trigger in your retention schedule and, as a floor, keeping the record at least as long as the related IND records under 312.57(c). In the EU, batch records for investigational medicinal products follow EU GMP Chapter 4, paragraph 4.11, described below.
| Rule | Records covered | Minimum retention (paraphrased) | Our reading for agent logs |
|---|---|---|---|
| 21 CFR 58.195(b), nonclinical laboratory studies (GLP) | Documentation records, raw data and specimens pertaining to a nonclinical laboratory study and required by Part 58 | Whichever of these is shortest: at least 2 years after FDA approves the research or marketing application the study supported (this branch does not apply to studies supporting an IND or IDE; under 58.195(b)(1) those records are governed by (b)(2)); at least 5 years after the study results are submitted to FDA in support of an application; or, where the study is not submitted, at least 2 years after the study is completed, terminated or discontinued. 58.195(c) allows shorter retention for certain fragile specimens and test-article samples; 58.195(g) permits true copies | The log of an agent that writes or changes GLP study data, or that would be needed to reconstruct how raw data became reported results, should be kept with the study archive for the same period. 58.190 requires archives with an identified responsible individual, indexed for expedient retrieval |
| 21 CFR 211.180(a) and (b), drug product manufacturing (CGMP) | Production, control and distribution records specifically associated with a batch; records for components, containers, closures and labeling | Batch records: at least 1 year after the expiration date of the batch. Components, containers, closures and labeling: at least 1 year after the expiration date (commonly read as that of the last lot using them). For certain OTC products exempt from expiration dating, 3 years after distribution of the batch or of the last lot | Retention runs from batch expiry, so the period ends at a different date for each batch an agent touched. 211.180(c) requires the records to be readily available for inspection during that period; records that can be retrieved immediately by computer from another location meet this. 211.68(b) separately requires a backup file of data entered into a computer system, with backups designed to be exact, complete and secure from alteration, inadvertent erasure or loss |
| 21 CFR 312.57(c), IND sponsors | Records and reports required by Part 312 | 2 years after a marketing application is approved for the drug; or, if an application is not approved, until 2 years after shipment and delivery of the drug for investigational use is discontinued and FDA has been notified | The end date depends on a future regulatory event that is unknown when the agent acts, so the log has to be held open-ended until that event is recorded |
| 21 CFR 312.62(c), clinical investigators | Records the investigator must keep under Part 312, including drug disposition records and case histories | 2 years after a marketing application is approved for the drug for the indication being investigated; or, if no application is to be filed or it is not approved for that indication, until 2 years after the investigation is discontinued and FDA is notified | Applies to agents used at or for an investigator site, for example in preparing case-history data; again the clock is event-driven |
| 21 CFR 812.140(d), investigational device studies (IDE) | Records investigators and sponsors must keep under Part 812, subpart G | During the investigation and for 2 years after the later of the date the investigation is terminated or completed, or the date the records are no longer required to support a PMA, a product development protocol notice of completion, an HDE, a 510(k) or a De Novo request | For agents that create or change records in a device investigation; the end date again depends on later regulatory events |
European retention periods
In the EU, GMP retention is set in Chapter 4 (Documentation) of the EU GMP Guide, EudraLex Volume 4. Paragraph 4.10 says it should be clearly defined which record relates to each manufacturing activity and where it is located, and that secure controls must be in place to ensure the integrity of the record throughout the retention period, validated where appropriate. Paragraph 4.11 says batch documentation must be kept for one year after expiry of the batch or at least five years after certification of the batch by the Qualified Person, whichever is longer. For investigational medicinal products, batch documentation must be kept for at least five years after the completion or formal discontinuation of the last clinical trial in which the batch was used. Paragraph 4.11 also notes that legislation for specific product types, such as advanced therapy medicinal products, may set longer periods for certain documents. Under paragraph 4.12, critical documentation, including raw data (for example relating to validation or stability) supporting information in the marketing authorisation, should be retained while the authorisation remains in force; the same paragraph says it may be acceptable to retire certain raw data where it has been superseded by a full set of new data, but the justification should be documented and should take batch-record retention into account; for example, process validation raw data should be kept at least as long as the records of every batch released on the strength of that validation. In our reading, agent logs that support such raw data follow the same limit.
For clinical trials, Article 58 of Regulation (EU) No 536/2014 requires the sponsor and the investigator to archive the content of the clinical trial master file for at least 25 years after the end of the trial, unless other Union law requires a longer period; subjects' medical files are archived under national law. Article 58 also requires that the media used for archiving keep the content complete and legible throughout that period, and that any alteration be traceable. EMA's 2023 guideline on computerised systems and electronic data in clinical trials adds practical expectations in section 6.11: keep an inventory of essential data and documents with their retention periods, make sure the files and any software needed to read them remain accessible throughout the retention period, safeguard data integrity for the periods required in any region where the data may be used for regulatory submissions, not only the country where they were generated, and archive data in a read-only state. The same section reminds sponsors and investigators that retention periods should respect the data-protection principle of storage limitation, which matters when prompts or outputs contain personal data.
Revised versions of Chapter 4 and Annex 11 went to a European Commission stakeholder consultation from 7 July to 7 October 2025, alongside a new draft Annex 22 on artificial intelligence. When we checked the EudraLex Volume 4 page on 20 October 2026, the 2011 texts were still the ones listed; re-check before relying on paragraph numbers.
Readable for the whole period: Annex 11 sections 7 and 17
Keeping a log is not enough; it has to be usable at the end of the period. Annex 11 is a GMP text; for EU clinical trials the closer reference is EMA's 2023 guideline, covered above and below, and for GLP it is OECD GLP Advisory Document No. 17. Annex 11's storage and archiving clauses are still a clear statement of what readable for the whole period means. EU GMP Annex 11 section 7.1 says data should be secured by physical and electronic means against damage, stored data should be checked for accessibility, readability and accuracy, and access to data should be ensured throughout the retention period. Section 7.2 asks for regular back-ups, with the integrity and accuracy of back-up data and the ability to restore checked during validation and monitored periodically; in the US, 21 CFR 211.68(b) sets a comparable backup requirement for CGMP computer systems. Section 17 says data may be archived, archived data should be checked for accessibility, readability and integrity, and if relevant changes are to be made to the system, such as computer equipment or programs, the ability to retrieve the data should be ensured and tested.
Agent platforms change faster than most validated systems. Over a retention period that can run for many years, the model behind an agent may be retired, the orchestration software replaced and the log format revised. We suggest treating each of those as a relevant system change in the sense of section 17, which asks that the ability to retrieve the data be ensured and tested when such changes are to be made, rather than assuming the archive still works. Practical risks include logs that store only a model identifier the provider no longer serves, outputs stored only by reference to a vendor console, and encoded or compressed payloads that need software you no longer run. EMA's ALCOA++ definition of legible asks that compression, encryption and coding be completely reversible, which is a useful test for any agent log you archive.
In the US, FDA's 2003 guidance says FDA does not intend to object if required records in electronic format are archived to non-electronic media such as microfilm, microfiche and paper, or to a standard electronic file format such as PDF, XML or SGML, provided the predicate-rule requirements are fully met and the copies preserve the content and meaning of the records. For clinical trial data, EMA's guideline is stricter. Section 6.12 says that when a trial database is decommissioned, archived formats should allow the database to be restored, including its dynamic functionality and metadata such as audit trails, event logs and user logs; where that is no longer possible, the data and metadata should be available as dynamic data files, and static formats of dynamic data will not be considered adequate. The sponsor should establish which audit trails and logs a system holds and how they will be retained, and address this in contracts where a service provider is involved. If you plan to archive agent logs to a static format, decide first which regime applies and what the reviewer will need to see.
What to retain for each agent action
The retention period applies to the content that lets someone reconstruct the action. The table lists what we suggest keeping for each agent action that touches a regulated record, and the requirement it answers to. None of the cited documents prescribes an agent log schema; the list is our interpretation and should be confirmed through your own risk assessment and validation. The field-by-field reasoning is on our page on audit trails for AI agents in regulated research.
| Item | What to keep for the full retention period | Why |
|---|---|---|
| Prompt or input reference | The task request and the version or hash of the instructions in force, with the instructions themselves held in a controlled store | Without it a reviewer cannot tell what the agent was asked to do; ALCOA++ complete |
| Model and version | Provider, exact model identifier or dated snapshot, and settings that affect output | The same request to another model version can give a different result, and the model may be retired before the period ends |
| Data accessed | Identifiers, versions and hashes of the records and documents the agent read | Lets the conclusion be re-derived from the evidence; the source record stays the original |
| Output | The output as produced, or a hash plus a pointer to a controlled copy that is itself retained | A later re-run is not a true copy; keep what was actually produced |
| Record changes | Record, field, previous and new value, and the reason | 21 CFR 11.10(e) changes must not obscure previously recorded information; Annex 11 section 9 reason for change |
| Reviewer | Identity of the person who accepted, edited or rejected the output, the decision, the time and any linked electronic signature | Separates the agent's contribution from human approval; 21 CFR 11.50 and 11.70 |
| Identity, time and integrity | Agent identity and accountable person, synchronised timestamps, and the integrity protection used | Secure, computer-generated, time-stamped audit trail under 11.10(e) |
| Retention class | The retention rule and trigger event inherited from the most demanding record the action touched | So the entry is not disposed of before the record it explains |
Which clock applies to an agent log
A single agent run can touch records governed by different rules, for example study data under Part 58 and a document that later supports an IND under Part 312. The simplest defensible rule is to give the whole run's log the longest retention, and the latest trigger event, of anything it touched. Splitting one run's log across several retention schedules is possible but makes it easy to destroy part of the evidence for a record that is still within its period.
Several of the periods above start from an event that has not happened when the agent acts: an approval, a submission, a batch expiry or the end of the last trial that used a batch. A practical approach is to record a retention class and trigger at the time the entry is written, link it to your records-retention schedule, and dispose of entries only after the trigger has been recorded and the period has run. Paragraph 4.10 of EU GMP Chapter 4 and section 6.11 of EMA's guideline both expect you to know which record relates to which activity and where it is kept; an inventory that lists agent logs alongside the records they explain, with who can access them, supports that expectation and makes the retention decision auditable.
Do not use a model provider's own logs as the retained copy. Providers set their own retention and access terms, which can change during a period that may run for decades in a clinical trial, and those logs may record only the API call rather than the regulated context. Keep the retained trail under your own quality system and record in your agreements what the provider keeps. Exploratory discovery work outside regulated records does not carry these periods, but decide and document the boundary before the work starts, because an agent's analysis can later support a regulated record or submission.
Knowing how long to keep agent logs, and keeping them readable for that long, is a records and data integrity job as much as an AI one. 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 (AI discovery, data-flow and vendor-terms mapping, and a NIST AI RMF gap report) 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 GxP scope, retention schedules, validation and compliance stay with your quality function.
Does 21 CFR Part 11 set a retention period for AI agent logs?
No. 21 CFR 11.10(e) says audit-trail documentation must be kept at least as long as the subject electronic records, and 11.10(c) requires those records to be protected for accurate and ready retrieval throughout their retention period. The number of years comes from the predicate rule, such as 21 CFR 58.195, 211.180, 312.57 or 312.62.
FDA uses enforcement discretion on Part 11 record retention. Can we keep agent logs for less time?
Not safely. FDA's 2003 guidance applies enforcement discretion to the Part 11 retention and audit-trail requirements, but says Part 11 remains in effect, persons must still comply with predicate-rule requirements for record retention and availability, and FDA can take regulatory action for noncompliance with those predicate rules. Where a predicate rule requires a record and its history, the agent's contribution to that history should be kept with it.
Can we archive agent logs as PDFs?
Possibly, for US records. FDA's 2003 guidance does not object to archiving electronic records to paper, microfilm or a standard electronic file format, naming PDF, XML and SGML as examples, if predicate-rule requirements are met and content and meaning are preserved. For EU clinical trial data, EMA's guideline expects archived formats to allow a decommissioned trial database to be restored with its audit trails and logs, or failing that to keep them as dynamic data files, and says static formats of dynamic data will not be considered adequate, so a PDF alone is unlikely to be enough there.
How long are EU clinical trial records kept?
Article 58 of Regulation (EU) No 536/2014 requires the sponsor and investigator to archive the clinical trial master file content for at least 25 years after the end of the trial, unless other Union law requires longer. Agent logs that form part of the trial record should be planned for that horizon.
What if the model an agent used is retired before the retention period ends?
The log should still be usable without it. Retain the output as it was produced and the inputs it relied on, so the record can be reviewed without re-running the model. Annex 11 section 17 asks that retrieval be ensured and tested when relevant system changes are to be made.
Do agent logs from exploratory research need GxP retention?
Only where they create or change records required by a predicate rule, or support a regulated submission or decision. Decide and document which agent workflows are in scope before the work starts, because exploratory output can later become part of a regulated record.
- eCFR: 21 CFR 11.10 - Controls for closed systems (11.10(c) retrieval through the retention period; 11.10(e) audit-trail retention)
- eCFR: 21 CFR 58.195 - Retention of records (GLP)
- eCFR: 21 CFR 58.190 - Storage and retrieval of records and data (GLP)
- eCFR: 21 CFR 211.180 - Records and reports, general requirements (CGMP)
- eCFR: 21 CFR 312.57 - Sponsor recordkeeping and record retention
- eCFR: 21 CFR 312.62 - Investigator recordkeeping and record retention
- eCFR: 21 CFR 812.140 - Records (investigational devices; 812.140(d) retention period)
- eCFR: 21 CFR 211.68 - Automatic, mechanical, and electronic equipment (211.68(b) backup)
- eCFR: 21 CFR 211.137 - Expiration dating (211.137(g) investigational new drug products)
- eCFR: 21 CFR 600.12 - Records (licensed biologics; outside the scope of this page)
- eCFR: 21 CFR 210.2 - Applicability of current good manufacturing practice regulations (210.2(c) phase 1 exemption)
- eCFR: 21 CFR 58.130 - Conduct of a nonclinical laboratory study (58.130(e) changes to automated data entries)
- eCFR: 21 CFR Part 11 - Electronic Records; Electronic Signatures (including 11.50 and 11.70)
- FDA guidance: Part 11, Electronic Records; Electronic Signatures - Scope and Application (2003)
- European Commission: EudraLex Volume 4, Chapter 4 - Documentation (paragraphs 4.10 to 4.12, retention of documents)
- European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems (sections 7, 9 and 17)
- European Commission: EudraLex Volume 4 - Good Manufacturing Practice guidelines
- European Commission (EudraLex Volume 1): Regulation (EU) No 536/2014 on clinical trials on medicinal products for human use (Article 58, archiving of the clinical trial master file)
- EMA: Guideline on computerised systems and electronic data in clinical trials (EMA/INS/GCP/112288/2023), sections 4.5 ALCOA++, 6.11 Archiving and 6.12 Database decommissioning
- European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22 (2025)
- OECD Series on Principles of GLP No. 17: Application of GLP Principles to Computerised Systems (hosted by BfR, German GLP authority)
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 →