Answers · By Hugh Donatello, Atlas Bio · Updated 2026-11-11

AI Agent Audit Trails: 21 CFR Part 11 and Annex 11

How do you make AI agent actions auditable in regulated research? Record every agent action as a secure, computer-generated, time-stamped entry: which agent acted, which person owns it, the authority it held, the inputs, model and version it used, what it created or changed and why, and who approved the result. Protect entries from change, review them on a risk basis, keep them as long as the records they touch, and link each conclusion to its evidence so someone who was not there can reconstruct it.

Download: Audit-trail and access-control checklist for AI agents in GxP research (PDF)
One printable PDF checklist. Free to copy and use inside your organization. General guidance, not regulatory advice.
Download: AI-agent audit-trail record field dictionary (Excel)
Download: AI-agent audit-trail record JSON Schema (JSON)

How this maps to EU GMP Annex 11

This page is built around FDA's 21 CFR Part 11. If you also work under EU GMP, the table gives the closest Annex 11 clause for each Part 11 control discussed here. Annex 11 is a GMP text: for EU clinical trials the closer reference is EMA's 2023 guideline on computerised systems and electronic data in clinical trials, and for GLP it is OECD GLP Advisory Document No. 17, both covered further down. The pairing is our reading, not an official crosswalk, and the two texts are not equivalent. Part 11 is a US regulation that lists required controls (FDA's 2003 guidance narrows how it enforces some of them, as explained below), while Annex 11 is part of the EU GMP Guide (EudraLex Volume 4) and leaves more to a documented risk assessment. The clause-by-clause section further down covers most of these clauses in more detail.

Three differences matter most for agents. First, Annex 11 clause 9 asks firms to consider building in an audit trail on the basis of risk, where Part 11 lists it as a control; in practice FDA's 2003 guidance also applies enforcement discretion to Part 11 audit trails and recommends deciding on them by documented risk assessment, while still enforcing the predicate rules. What Annex 11 adds in its own text is that the reason for any change or deletion of GMP-relevant data should be documented and that audit trails should be regularly reviewed. Second, Annex 11 covers topics Part 11 does not, including suppliers and service providers, periodic evaluation, incident management, systems that record Qualified Person certification and batch release, and business continuity. Third, Part 11 is more specific about electronic signatures: each must be unique to one individual, and a signed record must show the signer's printed name and the meaning of the signature. As of our check on 17 October 2026, the current Annex 11 text is revision 1, in operation since 30 June 2011; a revised draft went to public consultation in 2025, as described below.

Part 11 controlClosest Annex 11 clauseWhat this means for an AI agent under GMP (our interpretation)
11.10(a) Validation1 Risk management; 4 ValidationThe agent, its model version, tools and logging are validated for the intended use, to an extent justified by a documented risk assessment
11.10(b) Accurate and complete copies for inspection8.1 Printouts; 9 audit trails convertible to a generally intelligible formLogs can be exported in human-readable and electronic form for a reviewer or inspector, not only as raw traces
11.10(c) Protection and retrieval through the retention period7 Data storage; 17 ArchivingLogs and outputs are backed up, restorable and readable for the full retention period, including after a model or platform change
11.10(d) and (g) Access limits and authority checks12.1 to 12.3 Security; access authorisations recordedThe agent has its own least-privilege credentials, and every grant or revocation is logged
11.10(e) Audit trails9 Audit trails; 12.4 operator identity with date and timeA secure, computer-generated, time-stamped entry for every create, change or delete, naming the agent and the accountable person, that keeps the previous value rather than overwriting it and is retained for at least the retention period required for the record; for GMP-relevant changes and deletions, the reason, and evidence of regular review
11.10(i) Education, training and experience2 PersonnelA named process owner and system owner for each agent; people who build, run or review agents are qualified for that role
11.10(k) Controls over systems documentation, including revision and change control of that documentation4.3 Inventory and system description; 10 Change and configuration management (wider: covers the system itself)The agent is in the system inventory, and its system description, prompts and configuration records are version-controlled. Changes to the agent itself (model version, prompts, tool permissions) go through change control under Annex 11 clause 10; in the US that expectation comes mainly from validation (11.10(a)) and the predicate rules rather than from 11.10(k)
11.50 and 11.70 Signature manifestation and linking14(b) and (c) Electronic signatureHuman signatures carry the date and time, stay permanently linked to the agent-prepared record they approve, and for Part 11 also show the printed name and meaning; the agent does not sign
11.100(a) Signature unique to one individual; 11.100(c) certification to FDA of handwritten equivalenceNo direct counterpart for uniqueness; 14(a) same impact as a hand-written signature within the companyPart 11 defines an electronic signature as executed, adopted or authorised by an individual (11.3(b)(7)), so an agent should not hold one; approval stays with a named person
No direct counterpart (the Qualified Person is an EU role)15 Batch releaseAn agent may prepare release records, but the system lets only a Qualified Person certify, identifies that person and uses an electronic signature
No direct counterpart3 Suppliers and service providersFormal agreements with model or hosting providers that process GMP data, stating their responsibilities
No direct counterpart (for incidents, 11.300(c) and (d) cover only compromised tokens or cards and attempted misuse of passwords or identification codes)11 Periodic evaluation; 13 Incident management; 16 Business continuityAgents are re-evaluated on a schedule, unexpected agent actions are handled as incidents, and a manual or alternative route exists where the agent supports a critical process

What regulators already expect

The binding GxP rules were not written with AI agents in mind, but their audit-trail expectations attach to regulated records and to the computerised systems that create them, so they apply whether a person or a program made the change. In the United States, 21 CFR 11.10(e) requires secure, computer-generated, time-stamped audit trails that independently record the date and time of operator entries and actions that create, modify or delete electronic records. Record changes must not obscure previously recorded information, and the audit trail must be retained at least as long as the subject records and be available for FDA review and copying.

FDA's 2003 guidance on the scope and application of Part 11 says the agency intends to exercise enforcement discretion on some Part 11 requirements, namely validation, audit trails, record retention, record copying and legacy systems. It still expects compliance with the underlying predicate rules, recommends deciding on audit trails through a documented risk assessment, and continues to enforce the other provisions, including access limits, authority checks and the electronic-signature controls. FDA's 2018 data-integrity guidance for CGMP defines an audit trail as a secure, computer-generated, time-stamped electronic record that allows the course of events relating to the creation, modification or deletion of an electronic record to be reconstructed. It also expects processes to be designed so that required data cannot be modified without a record of the modification.

In Europe, EU GMP Annex 11 asks firms to consider, on the basis of risk, building in a system-generated record of all GMP-relevant changes and deletions, with the reason documented, and to review that record regularly. For clinical trials, EMA's 2023 guideline on computerised systems and electronic data says an audit trail should be enabled for the original creation and every later modification of electronic data. It should show what was changed, by whom, when and, where applicable, why. It should be protected against change and deletion, and normal users should not be able to switch it off. ICH E6(R3), the international good clinical practice guideline, says the same in its own words: audit trails should not be disabled, should be interpretable, and should be reviewed on a planned, risk-based schedule.

Nonclinical laboratory studies have their own rule. Under 21 CFR 58.130(e), in automated data collection systems the individual responsible for direct data input must be identified at the time of input, and any change to automated entries must not obscure the original entry, must state the reason, must be dated and must identify the responsible individual. The OECD GLP advisory document on computerised systems (No. 17) applies the GLP Principles to such systems. On our reading, for an agent writing GLP study data, a named person must be identifiable as responsible for each input the agent makes.

When an AI agent creates or changes a regulated record, all of this applies to the agent's actions. The practical question is not whether agents need audit trails, but what an agent audit trail has to contain so that it meets expectations written for people and conventional software.

Where the AI agent sits in the record

Regulators attribute data to the person or system that generated it. EMA's clinical-trial guideline says data should be attributable to the person and/or system generating them and, depending on how critical the data are, traceable to the system or device that captured them. For an AI agent that means two identities on every entry: the agent itself, as a distinct, registered system identity, and the person who is accountable for it or on whose behalf it acted.

Part 11's access controls and authority checks, 21 CFR 11.10(d) and (g), are written in terms of authorized individuals. An agent is not an individual, so its permissions should be granted by, and traceable to, a named person who is accountable for what it is allowed to do.

Agents should never borrow a person's login, and people should never share an agent's credentials. FDA's CGMP data-integrity guidance explains why shared login accounts are a concern: they make it impossible to attribute an action to one individual. An agent running under a scientist's account produces an audit trail that says the scientist did something the scientist never did. That is an attribution failure, and it would also undermine any later investigation.

Electronic signatures are a separate matter. Under 21 CFR 11.100 each electronic signature must be unique to one individual, and Part 11 requires a signed record to show the signer's name, the date and time, and the meaning of the signature. An agent can prepare a record for signature, but approval, review and release signatures belong to people. The audit trail should show the agent's contribution and the human signature as separate events, and the agent should not be able to apply a signature for anyone.

Which agent work is in scope

Part 11 applies to records in electronic form that FDA's predicate rules require to be kept, and to electronic records submitted to FDA. FDA's 2003 guidance reads this narrowly: predicate-rule records kept electronically instead of on paper, predicate-rule records kept electronically as well as on paper when the electronic version is relied on to perform regulated activities, applicable electronic submissions, and electronic signatures intended to be equivalent to handwritten signatures. Hybrid arrangements need a clear decision on which version is the regulated record. If the electronic record is the regulated record, signatures executed to it must be linked to it under 21 CFR 11.70. If a wet-signed printout is the official record, its relationship to the electronic source is controlled through the predicate rule, procedures and data-integrity controls. If the electronic version is in practice relied on to perform regulated activities, Part 11 applies regardless of the printout. FDA's January 2025 draft guidance on AI used to support regulatory decision-making for drugs and biological products covers AI that produces information or data supporting decisions about safety, effectiveness or quality. It explicitly does not address AI used in drug discovery, or AI used for operational efficiencies such as internal workflows or drafting a submission, where that use does not affect patient safety, drug quality or the reliability of nonclinical or clinical study results. The guidance is still a draft and is not for implementation, but its scope statement is a useful boundary marker.

That leaves a grey zone that research organisations have to decide for themselves. Exploratory discovery work can usually be governed more lightly than GxP work. However, an agent's analysis can migrate into a regulated decision later: a target rationale whose supporting analysis is later submitted in an IND, a dataset cleaned by an agent that becomes study data, or a safety signal triaged by an agent. If the provenance was never captured, it cannot be reconstructed afterwards. A defensible approach is to decide, in writing and before the work starts, which agent workflows produce or touch regulated records, apply the full audit-trail standard to those, and keep at least the identity, input, model-version and output fields for everything else.

The January 2026 guiding principles of good AI practice in drug development, published jointly by FDA and EMA, point the same way. Principle 6 asks that data source provenance, processing steps and analytical decisions be documented in a detailed, traceable and verifiable manner, in line with GxP requirements. Principle 2 asks for validation and oversight proportionate to the context of use and model risk.

The AI-agent audit-trail fields

The table below lists the fields an agent audit-trail entry should carry when the agent touches regulated records, and the requirement or guidance each field answers to. The anchors are the closest primary-source expectation; none of these documents prescribes an agent log schema, so the field list is an engineering interpretation that each organisation should confirm through its own risk assessment and validation. NIST SP 800-53, the US federal security control catalogue, is not a GxP regulation, but its control AU-3 is a useful reference for typical audit-record content: what type of event occurred, when, where, its source, its outcome, and the identity of the people, subjects or entities involved.

FieldWhat to recordWhy it is needed
Event ID and sequenceA unique, never-reused identifier and a monotonically increasing sequence number per logEngineering practice that lets a reviewer show no entry is missing or out of order; supports EMA 6.2.1 protection against deletion and NIST SP 800-53 AU-3
TimestampDate and time in UTC or with a fixed offset, from a clock synchronised to an external reference; for buffered actions, both action time and write time21 CFR 11.10(e); EMA ALCOA++ 'contemporaneous' - time set by an external standard; NIST SP 800-53 AU-8
Agent identityRegistered agent ID, agent software version and configuration versionAttributable to the system that generated the data (EMA 4.5); analogous to the operator identity required by Annex 11 clause 12.4
Accountable personThe named owner of the agent and, if different, the person on whose behalf it acted in this sessionAttribution to a person; Annex 11 clause 2 defined responsibilities; no shared accounts (FDA data-integrity Q&A)
Authority in forceThe permission set the agent held at the time, and any approval reference it relied on21 CFR 11.10(d) and (g) access limits and authority checks; Annex 11 clause 12.3 access authorisations recorded
Model and versionModel provider, exact model identifier or snapshot, and generation settings that affect outputReconstruction: the same request to a different model version can produce a different result; FDA AI draft guidance on life-cycle maintenance
Instructions versionVersion identifier or cryptographic hash of the system prompt, task instructions and any retrieved policy textWithout it a reviewer cannot tell what the agent was told to do; FDA-EMA principle 6 on documented processing steps
InputsIdentifiers and versions of the records, documents and datasets read, with content hashes, and the retrieval results actually returnedRe-derivation of the conclusion; FDA-EMA principle 6 on data provenance
Tool callTool or interface name, target system, parameters and the response receivedShows every action, not only the final output; ICH E6(R3) 4.2.2 on recording workflow actions as well as data changes; FDA CGMP data-integrity Q&A on system-level actions
Record changeRecord identifier, field, value before and after, and whether it was a create, modify or delete21 CFR 11.10(e) changes must not obscure earlier values; EMA 6.2.1 initial entry and changes
Reason for changeThe stated justification, including the instruction or finding that triggered itAnnex 11 clause 9; EMA 6.2.1 'why'
OutputThe content produced, or a hash plus a pointer to the controlled copy, with its storage locationRetain it as the record of what the agent actually produced at the time; do not rely on a later re-run to recreate it. Where the agent copies or derives values from source data, the source stays the original
OutcomeSuccess, failure, refusal or error, with the error detailNIST SP 800-53 AU-3 outcome; failed and aborted actions should be visible, as FDA's CGMP data-integrity Q&A expects for aborted or incomplete chromatography injections
Human decisionReviewer identity, decision (accepted, edited, rejected), timestamp and, where signed, a link to the electronic signature recordSeparates the agent's contribution from human approval; 21 CFR 11.50 and 11.70 signature manifestation and linking
Integrity protectionA hash chain or digital signature over the entry, and where used, a trusted timestamp tokenTamper evidence; NIST SP 800-53 AU-9 and AU-10; IETF RFC 3161 time-stamp protocol
Retention classThe retention rule inherited from the most demanding record the action touched21 CFR 11.10(e) retention at least as long as the subject records

Worked example: one AI-agent action, one audit-trail record

Illustrative example only. The record below is invented to show most of the fields from the table above filled in for a single agent action; some are combined. It is not taken from any real system, study or organisation, and the identifiers, names and values are placeholders. The record is the agent's own entry together with the linked entries for the human review, the signature and the applied change, which in this example are written as separate time-stamped events. None of the documents cited on this page prescribes a record format, so the third column is our reading of the closest clause, not an official crosswalk.

The scenario: in a nonclinical (GLP) study, an AI agent compares a working summary table in the study's electronic data system against its source dataset, finds a mean that disagrees with the source, and proposes a correction. This is a table still being prepared, not the signed final report; a correction to a signed GLP final report has to be an amendment by the study director under 21 CFR 58.185(c). The table is changed only after the study director, or a scientist they have designated under the study's procedures, reviews the proposal and signs it. The third column gives the 21 CFR Part 11 clause, the closest EU GMP Annex 11 clause if the same record were GMP-relevant, and the ALCOA+ attribute (as used by WHO, MHRA and PIC/S) that each field supports.

For a GLP study the EU and OECD reference for computerised systems is OECD GLP Advisory Document No. 17, not Annex 11; we show Annex 11 because it is the clause map used elsewhere on this page. Do not read the example as an approved GMP pattern: a correction to GMP data with a direct impact on data integrity would likely be a critical application under the consultation draft of Annex 22 described below, which says large language models should not be used in critical GMP applications.

FieldIllustrative value (not a real record)Closest clause (our reading)
Event ID and sequenceEVT-000123; sequence 4812 in this logNo Part 11 or Annex 11 clause names it; engineering practice that lets a reviewer show no entry is missing or out of order. ALCOA+: Complete
TimestampAgent proposal 2026-03-02 14:07:31 UTC, entry written 14:07:31 UTC; change applied to T-4 2026-03-02 15:22:14 UTC; server clock synchronised to an external time sourcePart 11: 11.10(e), time-stamped audit trail recording the date and time of entries and actions. Annex 11: 12.4, date and time with operator identity. ALCOA+: Contemporaneous
Agent identityRegistered agent ID agent-example-01, running under its own credentials; agent software version 2.3.0; configuration version cfg-17Part 11: 11.10(e), records operator entries and actions (Part 11 does not define an agent; on our reading the agent's action is recorded here alongside the accountable person). Annex 11: 12.4 by analogy (the clause speaks of operators). ALCOA+: Attributable
Accountable person and authorityResponsible individual for this change: [named study scientist] (see the reviewer row below); system owner [named person, Study Data Management], accountable for the agent's configuration and permissions, not for the data; agent permission set ps-4: read source datasets, propose changes to summary tables, no signing rightsPart 11: 11.10(d) and (g), access limits and authority checks are written for authorized individuals, so the agent's permissions trace to a named person. GLP: 21 CFR 58.130(e), the individual responsible for a change to automated data entries is identified (here the reviewing scientist; on our reading the system owner answers for the system, not for the data). Annex 11: 2 defined responsibilities; 12.3 access authorisations recorded (this entry references the permission set in force). ALCOA+: Attributable
Model and versionModel provider [name]; exact model identifier including the dated snapshot; generation settings recordedPart 11: 11.10(a), shows the validated configuration was the one in use (no clause names a model version). Annex 11: 4 validation; 10 change and configuration management. ALCOA+: Complete
Input referenceTask request REQ-0457 from the study data manager; instructions version prompt-set v12, recorded with its SHA-256 hashPart 11: 11.10(a), under which changes to the instructions are controlled as part of the validated system; 11.10(k)(2) applies only to the extent the instructions are kept as systems documentation. Annex 11: 10, prompt and instructions changes handled as configuration changes. ALCOA+: Complete
Data accessedSource dataset DS-0091 version 3 and summary table T-4 version 2, each with its SHA-256 hash; read-only accessNo Part 11 or Annex 11 clause requires reads to be logged (11.10(e) covers create, modify and delete); recorded so the result can be re-derived. The read-only limit itself is an access control under 11.10(d) and (g) and Annex 11 12.1 to 12.3. ALCOA+: Complete; the source dataset, traceable through its own audit trail to the first-captured raw data, remains the original record, and the summary table is a derived record checked against it
Tool callRead DS-0091 version 3; read T-4 version 2; submit the proposed change to T-4 for review; each response hashedPart 11: 11.10(e) only where the call creates, modifies or deletes a record; reads and proposals are logged for reconstruction. ALCOA+: Complete
Record changeModify: table T-4, row 7, dose-group mean, 41.8 to 42.3; the earlier value 41.8 is kept in the trail, not overwritten; written by the study data system, not by the agent, 2026-03-02 15:22:14 UTC, only after the signed approval below, as a separate time-stamped entry attributed to the agent as proposer and to the approving scientist as the responsible individualPart 11: 11.10(e), record changes shall not obscure previously recorded information; 11.10(f), operational system checks enforce the permitted sequence (no write before a signed approval). Annex 11: 9, record of GMP-relevant changes and deletions. ALCOA+: Accurate (the derived table now agrees with its source); the earlier value stays visible rather than overwritten
Reason for changeTranscription error: the table value did not match the mean recalculated from DS-0091 version 3 (42.3) using the same animals and the protocol's calculation rules; the reviewer confirmed that no protocol-defined exclusion or unscheduled death explains the difference. Flagged by the agent; cause confirmed and correction approved by the reviewerPart 11: 11.10(e) does not itself require a reason; predicate rules can, for example 21 CFR 58.130(e) for GLP. Annex 11: 9, reason documented for change or deletion of GMP-relevant data. ALCOA+: Complete
OutputThe proposed correction and a short explanation as the model returned them, stored as controlled record OUT-0457 with its SHA-256 hash; the log entry holds the hash and a pointerPart 11: 11.10(c) protection and ready retrieval through the retention period; 11.10(b) then lets accurate and complete copies be produced for inspection. Annex 11: 7 data storage. ALCOA+: Original (the stored response is the record of what the agent produced; it is not recreated by a later re-run)
OutcomeSuccess: proposal submitted and awaiting review; no errors or retriesNo clause names it; failed or aborted actions should be as visible as successful ones (NIST SP 800-53 AU-3). ALCOA+: Complete
Human reviewer and decision[named study scientist, designated by the study director]; decision: accepted without edits; 2026-03-02 15:22:10 UTCPart 11: 11.10(e), the reviewer's action is a separate time-stamped entry; 11.50(a)(1) and (2), printed name and date and time on the signed record. Annex 11: 12.4, identity of operators confirming data, with date and time. ALCOA+: Attributable
Signature meaningApproval: "I have checked this correction against the source dataset and approve it"; electronic signature record ESIG-0889, linked to table T-4 version 3; the agent applied no signaturePart 11: 11.50(a)(3), meaning such as review, approval, responsibility or authorship; 11.70 signature linked to its record; 11.100(a) unique to one individual. Annex 11: 14(b) and (c), permanently linked, with time and date (Annex 11 does not separately require the meaning). ALCOA+: Attributable
Integrity and retentionEntry appended to write-once storage outside the agent runtime and hash-chained to the previous entry; retention class: the longest of the records the action touched (here, the study record)Part 11: 11.10(e), secure audit trail retained at least as long as the subject records; 11.10(c). Annex 11: 7.1 data secured against damage; 17 archiving. ALCOA+: Enduring and Available

What the worked example shows

Three things the example shows. First, the agent and the person appear as separate identities in separate events: the agent proposed, the reviewer approved and signed, the change was applied only after that, and the agent never held a signature. Second, the record keeps the output as it was produced and the earlier value of the changed field, so a reviewer can reconstruct what happened without re-running the model. Third, several fields, such as model version and instructions version, answer to no clause by name; they are there because without them the course of events cannot be fully reconstructed, which is the purpose FDA's 2018 CGMP data-integrity guidance gives an audit trail. Which fields your own records carry, and at what depth, should come from your risk assessment and be confirmed by validation.

See: This record as a field dictionary (Excel) · Machine-readable JSON Schema the example validates against

ALCOA++ mapped to agent behaviour

FDA's data-integrity guidance describes reliable data as attributable, legible, contemporaneously recorded, original or a true copy, and accurate (ALCOA). WHO's 2021 data-integrity guideline, like MHRA's 2018 GxP guidance and PIC/S PI 041-1, uses ALCOA+, which adds complete, consistent, enduring and available. EMA's clinical-trial guideline uses ALCOA++, which adds traceable. The table applies each attribute to an AI agent. The middle column paraphrases EMA's definitions, which are the most detailed of the three.

PrincipleWhat regulators meanWhat it means for an AI agent
AttributableData attributable to the person and/or system that generated themEvery entry names the agent identity and the accountable person; no shared or borrowed credentials
LegibleReadable in original context; any compression, encryption or coding fully reversibleLogs exportable in human-readable form; encrypted or encoded payloads decodable by an authorised reviewer
ContemporaneousRecorded at the time of the action, with time set by an external standardThe entry is written when the tool call happens, not reconstructed later from a summary; synchronised clocks
OriginalThe first capture, or a certified copy; dynamic data kept dynamicStore the actual model output and tool responses, not a later paraphrase; keep logs as queryable data, not only PDFs
AccurateA true representation of what happened; transfers validatedLogging is validated so that what was recorded is what the agent actually did, including retries and errors
CompleteEnough data and metadata to reconstruct the eventInputs, instructions version, model version, every intermediate tool call, and failed attempts, not only the final answer
ConsistentRecorded the same way, in the expected sequenceOne schema across agents; sequence numbers and timestamps that agree with each other
EnduringKept on durable media for the retention periodWrite-once or append-only storage, independent of the agent runtime and of any model provider's own retention
Available when neededReadily retrievable for review throughout the life cycleReviewers, auditors and inspectors can search and export the trail without engineering help
TraceableChanges traceable, not obscuring the original, and explainedEach change shows before and after values and the reason, and links back to the instruction that caused it

EU GMP Annex 11 clause by clause

Annex 11 applies to all computerised systems used as part of GMP-regulated activities, and an agent that creates, changes or relies on GMP data is part of such a system. The current text is revision 1, which came into operation on 30 June 2011. The clauses below are the ones that bear most directly on agents; the right-hand column is our interpretation, not Commission text.

The European Commission ran a stakeholder consultation from 7 July to 7 October 2025 on a revised Chapter 4 (Documentation), a revised Annex 11 and a new draft Annex 22 on artificial intelligence. When we checked the EudraLex Volume 4 page on 17 October 2026, it still listed Annex 11 at its January 2011 revision and did not list an Annex 22. Treat the drafts as a signal of direction, not as binding text, and re-check the EudraLex page before relying on either version.

That direction matters for agents. The consultation draft of Annex 22 applies to static machine-learning models with deterministic output used in critical GMP applications. It states that dynamic models, models with probabilistic output, and generative AI and large language models should not be used in critical GMP applications, meaning those with direct impact on patient safety, product quality or data integrity. Where such models are used in non-critical GMP applications, it says qualified and trained personnel should always be responsible for confirming that the outputs suit the intended use. If that text is adopted as drafted, an agent built on a large language model would be limited to non-critical GMP manufacturing work with a person in the loop, however good its audit trail.

Annex 11 clauseWhat it says, in briefImplication for AI agents
1 Risk managementExtent of validation and data-integrity controls based on a justified, documented risk assessmentAssess each agent workflow by what it can change and how critical the data are; the depth of logging and review follows from that
2 PersonnelProcess owner, system owner, Qualified Persons and IT cooperate, with defined responsibilitiesName a process owner and a system owner for every agent; the accountable-person field points to them
3 Suppliers and service providersFormal agreements with third parties, including those processing data; audits based on riskA model or hosting provider used to provide, maintain or retain a GMP system, or to process its data, falls under these supplier controls; agreements should define data handling, model changes and log access as applicable
4 ValidationValidation documentation covering the life cycle, justified by risk assessment; user requirements traceableThe central question for agents: the agent, its model version, tools and logging are validated for the intended use, with changes assessed under change control and retested or re-validated where risk and impact justify it
4.3 InventoryUp-to-date list of systems and their GMP functionality; system description for critical systemsAgents belong in the system inventory, with their data flows, tools and interfaces described
4.8 Data migrationValidation checks that data are not altered in value or meaning when transferredApplies when agents move or transform data between systems; log before and after values
6 Accuracy checksCritical data entered manually get an additional check, by a second operator or validated electronic meansThe clause speaks of manual entry; by analogy, where an agent enters or transcribes critical data, decide by risk assessment whether the agent is itself a validated electronic means or whether an independent check by a person or validated means is needed
7 Data storageData secured against damage; accessibility, readability and accuracy checked through the retention periodAgent logs and outputs stored, backed up and restorable for the full retention period
9 Audit trailsRisk-based record of all GMP-relevant changes and deletions, reason documented, regularly reviewedThe field table above; record a reason for every agent change or deletion of GMP-relevant data
10 Change and configuration managementChanges to systems and configurations made under a defined procedureA new model version, prompt or tool permission is a configuration change and goes through change control
11 Periodic evaluationSystems periodically evaluated to confirm they remain in a valid stateRe-evaluate agent performance and log completeness on a schedule; consistent with FDA-EMA principle 9 on monitoring and drift
12 SecurityAccess restricted; access authorisations recorded; identity of operators recorded with date and timeAgents get least-privilege credentials of their own; grants and revocations are logged
13 Incident managementAll incidents reported and assessed; root cause of critical incidents identifiedUnexpected agent actions are incidents; preserve the relevant log segment before investigating
14 Electronic signatureSignatures permanently linked to their record, with time and dateAgents do not sign; human signatures are linked to the agent-prepared record they approve
15 Batch releaseOnly Qualified Persons certify batch release, identified and recorded, using an electronic signatureAn agent may prepare release documentation, but certification is the Qualified Person's alone
17 ArchivingArchived data checked for accessibility, readability and integrity; retrieval tested after system changesArchived agent trails must stay readable when the agent platform or model provider changes

Why generative models change the logging problem

Many conventional validated systems are deterministic under controlled configuration, so re-running them can support an investigation. A large language model may not give the same output twice, and the provider can retire or update the model behind an identifier. Re-running a request months later is therefore not a reliable way to reconstruct what the agent did. The audit trail has to capture the output itself, together with the exact model version and the instructions in force, at the time of the action. In ALCOA terms, the model's response should be treated as the original record of what the agent produced. Where the agent transcribes, extracts or derives values from source data, the source remains the original, and the agent's output is a transcribed or derived record whose accuracy has to be checked against the source.

This also changes what 'reconstruction' means. The aim is not to reproduce the model's output byte for byte. The aim is that a reviewer can see what the agent was asked, what evidence it read, what it produced, what was done with the result and who accepted it, and can then check the conclusion against the evidence independently. FDA's draft AI guidance frames credibility as evidence gathered for a particular context of use and notes that model performance can change over time or across deployment environments. That is why the model-version and instructions-version fields matter: they let you find which outputs a model change could have affected.

Making the trail tamper-evident

An agent should be able to write to its audit trail but never alter or delete entries, and the administrators who run the agent platform should not be able to do so silently either. EMA's guideline expects that if an administrator can deactivate an audit trail, the deactivation itself is logged. ICH E6(R3) allows audit trails to be modified only in rare circumstances, such as removing personal information entered by mistake, and only with a log of the action and its justification.

Common engineering patterns include append-only or write-once storage held outside the agent's own runtime, a hash chain in which each entry includes a cryptographic hash of the previous one so that any deletion or edit breaks the chain, and periodic anchoring of the chain head with a trusted timestamp under IETF RFC 3161. NIST SP 800-53 control AU-9 covers protecting audit information from unauthorised access, modification and deletion, and NIST SP 800-92 covers log management practice more broadly. No regulation requires a particular technique. What matters is that you can show, and have validated, that audit information is protected against unauthorised alteration or deletion, that any permitted administrative action is controlled and logged, and that tampering would be detected.

Reviewing agent audit trails

An audit trail no one reads does not do its job. For CGMP records, FDA's 2018 data-integrity guidance says the people responsible for reviewing a record should review the audit trail of changes to that record as they review the rest of it. Where CGMP regulations set a review frequency for the data, the same frequency applies to the audit trail; otherwise it should be set by risk assessment. For clinical investigations, FDA's question-and-answer guidance on electronic systems, records and signatures in clinical investigations addresses the equivalent expectations, and ICH E6(R3) asks for planned, risk-based review of audit trails and other metadata. EMA expects trial-specific, risk-based audit-trail review, focused on critical data, that is proactive and ongoing. It notes that technology can help review large datasets.

For agents, useful review questions include: did any agent act outside its permission set; did any agent change regulated data without a recorded reason; did an agent approve or validate its own output; were there bulk changes, unusual hours or repeated retries; did any tool call reach a system or endpoint not on the approved list; and did any human approval come too quickly to have involved reading what was approved. Automated checks can flag these patterns. When an AI tool assists the review, a person still owns the conclusion, and the reviewing tool's own actions belong in the trail.

Retention, archiving and the model provider

Under 21 CFR 11.10(e), audit-trail documentation must be kept at least as long as the electronic records it covers. A single agent run can touch records with different retention rules, so the simplest safe rule is to give the run's log the longest retention of anything it touched. EMA's guideline expects audit trails and metadata to stay with the data and, when a system is decommissioned, to remain available as dynamic files rather than only as static printouts.

Do not rely on a model provider's logs as your audit trail. Providers set their own retention and access terms, which can change, and those logs usually record the API call rather than the regulated context around it. Keep your own trail under your own quality system, and record in your agreements what the provider retains, for how long and whether you can get it. Annex 11 clause 3 expects formal agreements with service providers that process data, with the provider's responsibilities clearly stated.

From logs to evidence

Logs alone are not an audit trail a reviewer can use. Each conclusion an agent contributes to should link to the evidence and test records behind it, so that someone who was not present can re-derive it. When something goes wrong, the relevant log segment, outputs and inputs should be preserved first, before anyone knows what they will show, so that a later determination rests on evidence collected without hindsight.

A practical build order is: register every agent with an identity and an owner; decide which workflows touch regulated records; define one log schema covering the fields above; write entries to protected storage outside the agent runtime; validate that the log records what actually happened, including failures; put model, prompt and permission changes under change control; set a risk-based review schedule; and test retrieval from the archive. The FDA-EMA guiding principles call for risk-based validation, documented data provenance and processing, and life-cycle monitoring and re-evaluation; the EMA reflection paper on AI in the medicinal product life cycle and the NIST AI Risk Management Framework set out broader governance in the same spirit. The audit trail is where that governance becomes something an inspector can check.

A defensible audit trail for AI agents is a 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, validation and compliance stay with your quality function.

Is a chat history an audit trail?

No. A chat history shows the text exchanged. An audit trail records who or what acted, under which authority, on which inputs and model version, with what effect on which records, in a form that is protected against change and kept for the required retention period.

Does every AI use in a life-science company need Part 11 controls?

Part 11 applies to electronic records required by FDA's predicate rules or submitted to FDA. AI that creates or changes such records should meet those expectations; where AI output supports a regulatory decision, FDA's draft AI credibility framework is also relevant. Exploratory discovery work outside regulated records can be governed more lightly, but decide and document the boundary before the work starts, because outputs can migrate into regulated decisions later.

How long should agent logs be kept?

At least as long as the records the agent touched. 21 CFR 11.10(e) requires audit-trail documentation to be kept for at least the retention period of the subject records. If a run touched records with different rules, apply the longest.

Should we log the full prompt and the full model output?

You need to be able to reconstruct both. Many teams store the content in a controlled repository and put a hash and a pointer in the log entry, which keeps the trail compact and lets access to sensitive content be restricted. Where prompts or outputs contain personal data, apply your data-protection controls to that store too.

Is a hash chain or blockchain required?

No regulation requires a particular technology. Regulators expect audit trails to be secure, computer-generated, time-stamped and protected from change or deletion. Append-only storage, hash chaining and trusted timestamps are common ways to show this. Whichever you choose has to be validated.

Does the EU AI Act add logging requirements?

Article 12 of Regulation (EU) 2024/1689 requires high-risk AI systems to allow automatic recording of events (logs) over their lifetime, once the Act's high-risk obligations apply. Whether a particular research or GxP agent is high-risk under the Act has to be assessed case by case. The GxP audit-trail expectations apply regardless.

Can an AI agent review audit trails?

It can help, for example by flagging unusual patterns across large logs, and EMA's clinical-trial guideline notes that technology can support review of large datasets. The person responsible for the record still owns the review conclusion, and the reviewing agent's own actions should be logged.

What additional validation and audit trails are required when agentic AI interacts with regulated medical devices or robots?

No rule below is agent-specific; as good practice, log the instrument, command and response. For FDA-required electronic records, 21 CFR 11.10 lists validation, audit trails, sequencing checks and checks on the source of operational instructions; FDA's 2003 guidance applies enforcement discretion to validation and audit trails, not predicate rules. 211.68 (drug CGMP) adds calibration under a written program, input/output checks and authorized-only record changes; 58.63 (GLP) adds calibration and maintenance records. For a device you manufacture, Part 820 duties depend on its classification regulation; 820.10(c) design controls cover class II, III and software-automated class I. eCFR retrieved 10 Nov 2026.

  1. eCFR: 21 CFR Part 11 - Electronic Records; Electronic Signatures
  2. FDA guidance: Part 11, Electronic Records; Electronic Signatures - Scope and Application (2003)
  3. FDA guidance: Data Integrity and Compliance With Drug CGMP - Questions and Answers (2018)
  4. FDA guidance: Electronic Systems, Electronic Records, and Electronic Signatures in Clinical Investigations - Questions and Answers
  5. FDA draft guidance: Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (January 2025)
  6. FDA and EMA: Guiding principles of good AI practice in drug development (January 2026)
  7. European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems (revision 1, 2011)
  8. European Commission: EudraLex Volume 4 - Good Manufacturing Practice guidelines
  9. European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22 (2025)
  10. EMA: Guideline on computerised systems and electronic data in clinical trials (EMA/INS/GCP/112288/2023)
  11. ICH E6(R3) Guideline for Good Clinical Practice
  12. WHO Technical Report Series 1033, Annex 4: Guideline on data integrity (2021)
  13. MHRA: 'GXP' Data Integrity Guidance and Definitions (Revision 1, March 2018)
  14. European Commission: Draft Annex 22 - Artificial Intelligence (consultation document, 2025)
  15. eCFR: 21 CFR 58.130 - Conduct of a nonclinical laboratory study
  16. OECD Series on Principles of GLP No. 17: Application of GLP Principles to Computerised Systems (hosted by BfR, German GLP authority)
  17. PIC/S PI 041-1: Good Practices for Data Management and Integrity in Regulated GMP/GDP Environments
  18. EMA: Reflection paper on the use of artificial intelligence in the medicinal product lifecycle
  19. NIST SP 800-53 Rev. 5: Security and Privacy Controls for Information Systems and Organizations (AU family)
  20. NIST SP 800-92: Guide to Computer Security Log Management
  21. IETF RFC 3161: Internet X.509 Public Key Infrastructure Time-Stamp Protocol (TSP)
  22. NIST AI 100-1: Artificial Intelligence Risk Management Framework (AI RMF 1.0)
  23. EUR-Lex: Regulation (EU) 2024/1689 (Artificial Intelligence Act)
  24. eCFR: 21 CFR 211.68 - Automatic, mechanical, and electronic equipment (drug CGMP)
  25. eCFR: 21 CFR 58.63 - Maintenance and calibration of equipment (GLP)
  26. eCFR: 21 CFR Part 820 - Quality Management System Regulation (medical devices)
  27. eCFR: 21 CFR 862.2050 - General purpose laboratory equipment labeled or promoted for a specific medical use

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 →