Status of the guidance this page relies on
No regulator document cited here mentions AI agents, prompts or large language model suppliers in the context of periodic review. This page applies general computerised-system review guidance to an AI agent, and where it does that it says it is our reading. Status as checked on 20 October 2026:
EU GMP Annex 11 (Computerised Systems, January 2011 revision, in operation since 30 June 2011) is current EU GMP guidance. The European Commission consulted from 7 July to 7 October 2025 on a revised Annex 11 and a new Annex 22 on artificial intelligence; the consultation page shows the consultation closed, and the EudraLex Volume 4 page still lists the 2011 Annex 11 and no Annex 22. Both are drafts and may change before adoption; this page quotes them as written in the consultation versions. For US drug and biologic manufacturers the binding requirements are the CGMP regulations, including 21 CFR 211.68; FDA's Data Integrity and Compliance With Drug CGMP questions and answers (December 2018) is final guidance that contains nonbinding recommendations. ISPE's GAMP 5 (Second Edition, July 2022) and GAMP Guide: Artificial Intelligence (July 2025) are industry good practice, not regulation; both are paid, and this page uses only their public tables of contents.
Annex 11 and draft Annex 22 are EU GMP texts. For computerised systems in clinical trials, EMA's Guideline on computerised systems and electronic data in clinical trials (EMA/INS/GCP/112288/2023) is the closer reference; cite the framework that governs your process. This page is general information, not legal or regulatory advice. Decisions on GxP scope, criticality, review frequency and revalidation belong to your quality function.
What the guidance says about periodic review
Annex 11 section 11 (2011), Periodic evaluation: 'Computerised systems should be periodically evaluated to confirm that they remain in a valid state and are compliant with GMP. Such evaluations should include, where appropriate, the current range of functionality, deviation records, incidents, problems, upgrade history, performance, reliability, security and validation status reports.' The 2011 text sets no interval. Section 9 says audit trails need to be 'regularly reviewed', section 13 says all incidents, not only system failures and data errors, should be reported and assessed, and section 7.2 says the ability to restore backed-up data should be monitored periodically.
The draft revised Annex 11 section 14 is more detailed. Section 14.1 says periodic reviews should verify whether the system "remains 'fit for intended use' and in 'a validated state', or whether changes should be made and re-validation (complete or in parts) is required", and that findings should be analysed for consequences on product quality, patient safety and data integrity, and to prevent recurrence. Section 14.2 lists what reviews should include where applicable: changes since the previous review to hardware, software, configuration, platform, infrastructure and interfaces; changes to documentation, with a check that system changes are reflected in it; the combined effect of multiple changes, with undocumented changes identified, 'e.g. by means of configuration auditing'; actions from previous reviews, audits, inspections and CAPA; the conduct of, and actions from, audit trail reviews, access reviews and risk assessments; actions from incidents, problems, deviations and security incidents; support contracts, service level agreements and supplier key performance indicators; backup, restore and disaster recovery; archival; data integrity assessments; and changes to regulatory requirements.
On frequency, draft revised Annex 11 section 14.3 says reviews 'should be conducted, approved and closed according to plan', that their frequency 'should be established and justified based on the risk the system poses to product quality, patient safety and data integrity', and that a final review should be conducted when the system is taken out of use.
The draft Annex 22 covers only static machine-learning models with deterministic output in critical GMP uses. Its section 1 says it does not apply to generative AI and large language models, that such models should not be used in critical GMP applications (those with direct impact on patient safety, product quality or data integrity), and that in non-critical uses 'personnel with adequate qualification and training should always be responsible for ensuring that the outputs from such models are suitable for the intended use, i.e. a human-in-the-loop (HITL) and the principles described in this document may be considered where applicable'. For an LLM agent, its operation clauses are therefore principles to borrow, not requirements, and the draft expects appropriately qualified and trained personnel to be responsible for the agent's output. Section 10.3 says a model's performance, as defined by its metrics, 'should be regularly monitored to detect any changes in the computerised system'; section 10.4 says it should be regularly monitored whether input data are still within the model's sample space and intended use, with metrics defined for monitoring drift in the input data; and section 10.5 says that where a model gives input to a human decision and testing effort was reduced, records should be kept from that process and, depending on the criticality of the process and the level of testing of the model, this 'may imply a consistent review and/or test of every output from the model, according to a procedure'. Section 3.3 adds that, where a model supports a human decision and the effort to test it has been diminished, the operator's training and consistent performance should be monitored like any other manual process.
How often: setting a risk-based frequency
Neither the 2011 Annex 11 nor the draft revised Annex 11 gives a number of months. The draft ties frequency to risk and asks that it be justified. In US drug CGMP, 21 CFR 211.68(a) requires computers and related systems used in manufacture to be 'routinely calibrated, inspected, or checked according to a written program designed to assure proper performance', with written records kept, and 211.68(b) says 'the degree and frequency of input/output verification shall be based on the complexity and reliability of the computer or related system.' Section 211.68 does not use the term periodic review; it is the nearest US drug CGMP text on how often computerised systems are checked.
Review frequency does not make a use acceptable. In our reading, whether an AI agent may be used at all for a given GxP task, and what human check it needs, is a decision for the quality function before go-live; for an LLM agent, draft Annex 22 expects no critical GMP use and appropriately qualified and trained personnel responsible for the output in non-critical use. Under draft Annex 22, criticality depends on whether the use has direct impact on patient safety, product quality or data integrity, not on whether a person checks the output: an LLM agent in such a use falls in the critical category the draft says to avoid, whatever the review frequency, and the human-in-the-loop condition applies to non-critical uses. Whether a human check makes an AI's impact indirect is a criticality judgement for your quality function; the draft does not settle it. Periodic review finds problems after the fact; it does not replace that decision.
Within a use the quality function has accepted, in our reading, the factors that should push an AI agent's review frequency up are: the operation touches a required GxP record; the human check before its output takes effect is lighter, for example sampling rather than reviewing every output, where the quality function has accepted sampling for that use (draft Annex 22 section 10.5 notes that, depending on criticality and testing, review of every output may be needed); the model is hosted by a supplier who can change it; the agent can call tools or write to systems; it has a record of incidents, deviations or monitored drops in performance; and its inputs are changing, for example new document types or a new site. An agent that only drafts text a trained person rewrites, on a pinned model, with stable inputs and a clean record, sits at the other end. Write the chosen frequency and its justification into the validation plan or periodic review procedure before go-live, so the decision is on record and not improvised.
Calendar reviews are not the only trigger. In our reading, the procedure should also bring a review forward when monitoring shows performance outside the acceptance criteria, when a significant incident or deviation involves the agent, when the supplier announces or retires a model version, or when the intended use or the human review step changes. Some of these are change-control events in their own right; our separate answer on change control for AI models and prompts in GxP covers the retest decision.
What to check: performance against acceptance criteria
In our reading, the review should compare current performance with the acceptance criteria set at validation, not with a new, looser standard. Annex 11 section 11 lists performance and reliability among the items to evaluate. As a borrowed principle, draft Annex 22 sections 10.3 and 10.4 ask for regular monitoring of performance against the defined metrics and of whether inputs stay within the intended sample space.
In our reading, two practical points apply to an agent built on a large language model. First, output can differ between runs on the same input, so evidence from live use alone may not show a slow change; rerunning the validated test set, under the same rule for repeated runs and judging that the validation plan fixed, gives a like-for-like comparison. Record each rerun, and do not tune prompts or settings against that test set; if the review leads to a change, test the change on data the agent has not been adjusted to, so the check stays independent (draft Annex 22 sections 6.1 and 6.3 on test data independence and recording test data use, as a borrowed principle). Second, drift can come from the inputs as well as the model: new document types, a new site's records or a new kind of request can move the agent outside what it was validated for. Where a person reviews the agent's output, how often reviewers reject or correct it is a direct signal, and, as a borrowed principle, draft Annex 22 section 3.3 asks for the reviewer's own performance to be monitored where testing effort on the model was diminished.
What to check: audit trail review and access review
Audit trail review has its own frequency; in our reading it is usually shorter than the periodic review, because it is often tied to steps such as batch release. Annex 11 section 9 asks that audit trails be regularly reviewed. Draft revised Annex 11 section 12 asks for reviews under a documented procedure that says who reviews, what and when (12.5), by personnel not directly involved in the activities reviewed (12.6), targeted and based on risk, with a key element being to verify why a change was made (12.7), timely according to risk and before batch release unless a later review can be justified (12.8). FDA's 2018 data integrity guidance (question 8) says that where CGMP regulations specify a review frequency for the data, the audit trail review should follow it, for example review after each significant step under 21 CFR 211.188(b) and before batch release under 211.22, and otherwise the frequency should be set using process knowledge and risk assessment. The periodic review checks that these audit trail reviews happened and what they found (draft revised Annex 11 section 14.2(v)). In our reading, it should also check that the agent's own actions, as well as the people using it, are visible in the audit trail; draft revised Annex 11 section 12.1 frames audit trails around manual user interactions, but Annex 11 section 9 (2011) asks that, based on a risk assessment, consideration be given to a record of all GMP-relevant changes and deletions, with the reason documented, and draft section 12.2 says the audit trail should unambiguously capture who made a change, what was changed and when. In our reading, where an agent creates, changes or deletes GMP-relevant data, its actions should be logged and attributable to the agent, not recorded as manual actions of the person whose login it runs under.
Access review covers both people and the agent. Annex 11 section 12.3 (2011) says creation, change and cancellation of access authorisations should be recorded. Draft revised Annex 11 section 11.11 asks for recurrent reviews of user accounts and roles, documented, at a frequency 'commensurate with the risks and consequences of changes in systems and data made by unauthorised individuals', and section 11.10 sets least privilege and segregation of duties as guiding principles. In our reading, an agent's own service accounts, API keys and tool permissions belong in the same review: confirm the agent can still reach only the systems and records its intended use needs, and that no permission was added outside change control.
What to check: change history
Draft revised Annex 11 section 14.2(i) to (iii) asks the review to cover changes since the previous review, check that documentation reflects them, assess their combined effect, and identify undocumented changes, for example by configuration auditing; Annex 11 section 11 (2011) lists upgrade history. In our reading, for an agent the change history should cover the model and its version and settings, the system prompt and templates, the tools and permissions, the retrieval sources and how they are indexed, and the surrounding software and platform.
The useful test is reconciliation: does the configuration actually running match the configuration on record, and does every difference have an approved change behind it? Several small, individually assessed changes can together move the agent away from what was validated, which is why the draft asks for their combined effect to be assessed. Our answer on change control for AI models and prompts in GxP sets out which parts of an agent can change and how each change should be assessed.
What to check: incidents, deviations and CAPA trends
Annex 11 section 13 (2011) says all incidents should be reported and assessed and that the root cause of a critical incident should form the basis of corrective and preventive actions; section 11 lists deviation records, incidents and problems among the items for periodic evaluation. Draft revised Annex 11 section 14.2(iv) and (vi) ask the review to follow up actions from previous reviews, audits, inspections, CAPA, incidents, problems, deviations, security incidents and new security threats.
In our reading, the periodic review is the place to look across individual events for a pattern: the same kind of error recurring, errors clustered on one input type or site, reviewers repeatedly correcting the same field, or CAPA actions that were closed without the expected effect. A pattern that suggests the agent no longer performs as validated should lead to an investigation and a documented decision on retesting or revalidation.
What to check: the supplier and the underlying model
For a hosted model, the supplier can release, change and retire versions. Annex 11 section 3.1 (2011) expects formal agreements with third parties that state their responsibilities, and section 4.5 expects the supplier to be assessed appropriately. Draft revised Annex 11 section 7.1 says relying on a vendor or service provider does not change the requirements, section 7.3 asks for effective oversight under agreed service level agreements and key performance indicators, and section 7.5(ix) says the contract should agree 'on the process for release of new system versions and on the regulated user's possibility to test these prior to release'. Section 14.2(vii) and (viii) bring support contracts, service level agreements and supplier key performance indicators into the periodic review.
In our reading, the supplier part of the review should confirm which model versions were in use during the period and whether that matches the record; whether any version changes or retirement notices were received and how each was handled; whether the agreement still allows version pinning and testing before change; and whether the supplier's terms, data handling or service changed in ways that affect the intended use. Draft Annex 22 section 2.2 asks that documentation be available and reviewed by the regulated user whether a model is built in-house or provided by a supplier; as a borrowed principle, that applies here too.
Periodic review checklist for an AI agent (illustrative)
The table is our illustrative starting point, not regulator text and not a GAMP template. Your quality function sets the actual scope and frequency in your procedure, and higher-risk uses may need more than the table shows.
| Area | What to check (our illustrative reading) | Guidance it rests on |
|---|---|---|
| Fit for intended use and validated state | Is the intended use statement still accurate? Has the agent been used outside it? Conclude whether it remains validated or needs changes or revalidation, complete or in parts | Annex 11 section 11 (2011); draft revised Annex 11 14.1 |
| Performance against acceptance criteria | Rerun the validated test set under the same repeated-run rule and compare with the original criteria; review reviewer rejection and correction rates; check inputs against the validated range | Annex 11 section 11 (2011, performance, reliability); draft Annex 22 3.3, 10.3 and 10.4 (borrowed principles) |
| Audit trail reviews | Confirm the reviews ran at their own risk-based frequency, by independent reviewers, and that findings were investigated; confirm the agent's actions appear in the audit trail | Annex 11 section 9 (2011); draft revised Annex 11 12.5 to 12.8 and 14.2(v); FDA 2018 data integrity guidance, question 8 |
| Access review | Confirm people's access and the agent's own accounts, keys and tool permissions still match least privilege and the intended use; no permission added outside change control | Annex 11 section 12.3 (2011); draft revised Annex 11 11.10, 11.11 and 14.2(v) |
| Change history | Reconcile the running configuration (model version and settings, prompts, tools, retrieval sources, platform) with approved changes; assess the combined effect; identify undocumented changes | Annex 11 sections 10 and 11 (2011, upgrade history); draft revised Annex 11 14.2(i) to (iii) |
| Incidents, deviations and CAPA | Trend events since the last review; check CAPA effectiveness; follow up open actions from previous reviews, audits and inspections | Annex 11 sections 11 and 13 (2011); draft revised Annex 11 14.2(iv) and (vi) |
| Supplier and underlying model | Model versions in use versus the record; version change or retirement notices and how they were handled; agreement terms on notice and testing; supplier KPIs | Annex 11 sections 3.1 and 4.5 (2011); draft revised Annex 11 7.1, 7.3, 7.5(ix) and 14.2(vii), (viii) |
| Backup, restore and archival | Including the agent's logs and records of what it produced and who reviewed it, where these are GxP records | Annex 11 section 7.2 (2011); draft revised Annex 11 14.2(ix), (x) |
| Regulatory changes | Whether new or adopted guidance, such as a final Annex 11 or Annex 22, changes what the agent must meet | Draft revised Annex 11 14.2(xii) |
| Frequency and next review | Record the justified frequency, any events that bring the next review forward, and the date; plan a final review at retirement | Draft revised Annex 11 14.3 (21 CFR 211.68(a) and (b) are the nearest US text, on routine checks and input/output verification, not periodic review) |
Where ISPE's GAMP material fits
GAMP 5 (Second Edition) has operation appendices on system monitoring (O3), operational change and configuration management (O6) and periodic review (O8), according to its public table of contents. The GAMP Guide: Artificial Intelligence has an operation-phase appendix (P3) whose public table of contents lists system monitoring, incident and problem management, CAPA, operational change and configuration management, and periodic review, and an IT infrastructure appendix with a section on model monitoring and drift detection. The detailed content of these sections is in the paid guides, not in ISPE's public material, and nothing on this page summarises it.
Periodic review starts with knowing which AI models, prompts and data sources already touch regulated work. 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, review frequency, validation and compliance stay with your quality function.
How often should an AI agent in a GxP process be periodically reviewed?
None of the regulatory texts cited here sets a fixed interval. The draft revised Annex 11 section 14.3 says frequency should be established and justified based on the risk to product quality, patient safety and data integrity. In our reading, within a use your quality function has accepted, an agent that touches GxP records, runs on a supplier-hosted model or has its output checked by sampling rather than in full warrants more frequent review, plus event-triggered reviews when monitoring, incidents or supplier changes call for one. Review frequency does not make an otherwise unsuitable use acceptable.
Is periodic review the same as audit trail review?
No. Audit trail review is a recurring check of the records themselves, at a frequency set by the data it covers; FDA's 2018 data integrity guidance ties it to the review frequency CGMP sets for the data, such as before batch release, where CGMP specifies one, and otherwise to process knowledge and risk assessment. Periodic review is a wider evaluation of whether the system remains validated, and it checks that audit trail reviews were done and acted on.
Does the draft Annex 22 require performance monitoring for LLM agents?
No. The draft Annex 22 excludes generative AI and large language models from its scope, says they should not be used in critical GMP applications, and in non-critical uses expects appropriately qualified and trained personnel to be responsible for the output (a human in the loop). Its monitoring clauses, sections 10.3 and 10.4, can be borrowed as principles for non-critical uses. It is a consultation draft, not adopted guidance. That does not remove monitoring: the current Annex 11 still applies to the computerised system, and its section 11 includes performance and reliability in periodic evaluation.
What should we do if a periodic review finds the agent no longer meets its acceptance criteria?
Treat it as a finding to investigate. The draft revised Annex 11 section 14.1 asks reviews to decide whether changes or revalidation, complete or in parts, are required, and to analyse findings for consequences on product quality, patient safety and data integrity. Your quality function decides whether use continues while the cause is found.
Should the supplier's model updates be part of the periodic review?
Yes, in our reading. The draft revised Annex 11 brings supplier contracts, service level agreements and KPIs into periodic review, and says contracts should agree how new versions are released and whether you can test them first. The review should confirm which model versions ran during the period and that each change was assessed.
- European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems (January 2011 revision, in operation since 30 June 2011)
- European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22 (closed; 7 July to 7 October 2025; checked 20 October 2026)
- European Commission: Revised Annex 11 Computerised Systems - consultation draft (PDF)
- European Commission: Annex 22 Artificial Intelligence - consultation draft (PDF)
- European Commission: EudraLex Volume 4 - Good Manufacturing Practice guidelines (lists Annex 11, January 2011 revision; no Annex 22 as of 20 October 2026)
- eCFR: 21 CFR 211.68 - Automatic, mechanical, and electronic equipment
- FDA guidance: Data Integrity and Compliance With Drug CGMP - Questions and Answers (final, December 2018), full text (PDF)
- European Medicines Agency: Guideline on computerised systems and electronic data in clinical trials (EMA/INS/GCP/112288/2023)
- ISPE: GAMP 5 - A Risk-Based Approach to Compliant GxP Computerized Systems (Second Edition, July 2022)
- ISPE: GAMP 5 (Second Edition) public table of contents (PDF)
- ISPE: GAMP Guide: Artificial Intelligence (product page, published July 2025)
- ISPE: GAMP Guide: Artificial Intelligence, public table of contents (PDF)
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 →