Status of the guidance this page relies on
No regulator document cited here mentions prompts, prompt templates or retrieval sources by name, and none sets revalidation rules for an AI agent built on a large language model. This page applies general computerised-system change-control guidance to those components. 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. FDA's Computer Software Assurance for Production and Quality Management System Software (CSA) is final, non-binding guidance; the current version, issued on 3 February 2026, supersedes the version issued on 24 September 2025. Its stated scope is medical device production and quality system software, so for drug and biologic CGMP it is context, not a requirement. For US drug and biologic manufacturers the binding requirements are the CGMP regulations, including 21 CFR 211.68(b), which requires appropriate controls so that changes to master production and control records or other records are made only by authorized personnel, and, for required electronic records, 21 CFR Part 11, whose section 11.10(k) calls for controls over systems documentation, including revision and change control procedures. FDA's 2003 Part 11 scope and application guidance says FDA intends to exercise enforcement discretion over the audit-trail requirement in 11.10(k)(2), while it continues to enforce appropriate controls over systems documentation, and predicate rules apply in full. 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; final version adopted by the GCP Inspectors Working Group on 7 March 2023) is the closer reference; for non-clinical safety studies, OECD Series on Principles of Good Laboratory Practice and Compliance Monitoring No. 17, Application of GLP Principles to Computerised Systems (an advisory document, 2016). In our reading, the change-control principles are similar, but cite the framework that governs your process. This page is general information, not legal or regulatory advice. Decisions on GxP scope, criticality, change classification and revalidation belong to your quality function.
What the guidance says about changes
Annex 11 section 10 (2011): 'Any changes to a computerised system including system configurations should only be made in a controlled manner in accordance with a defined procedure.' Section 1 says decisions on the extent of validation should be based on a justified and documented risk assessment, taking into account patient safety, data integrity and product quality. Section 11 asks for periodic evaluation to confirm that systems remain in a valid state, and lists upgrade history among the things such an evaluation should cover where appropriate. Section 13 says all incidents, not only system failures and data errors, should be reported and assessed.
The draft revised Annex 11 is more specific. Section 3.1(ii) says any change to a computerised system, 'including but not limited to its configuration, its hardware and software components, and its platform and operating system', should be made in a controlled manner under defined procedures, and that 'any significant change which may impact product quality, patient safety or data integrity, should be subject to re-qualification and validation.' Section 6.6 asks that the chosen configuration be documented in a controlled configuration specification. Section 14.1 asks periodic reviews to 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 section 14.2(iii) asks them to identify undocumented changes, for example by configuration auditing.
The draft Annex 22 covers only static, deterministic machine-learning models in critical GMP uses and excludes generative AI and large language models. For their non-critical use, its section 1 says 'the principles described in this document may be considered where applicable', so for an LLM agent its clauses are principles to borrow, not requirements. Section 10.1 says the model, the system it runs in and the whole process should be under change control before deployment, that any change to the model, the system or the process 'should be documented and evaluated to determine if the model needs to be retested', and that 'any decision not to conduct such retest should be fully justified.' Section 10.2 asks for configuration control with effective measures to detect unauthorised change. Sections 10.3 and 10.4 ask for regular monitoring of performance against the defined metrics and of whether inputs stay within the model's intended sample space.
Section 1 of the same draft says 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. In our reading, if the draft is adopted as written, change control for an LLM agent in EU GMP manufacturing becomes a question for non-critical uses only, with a qualified, trained person responsible for the output. The current 2011 Annex 11 contains no such restriction. Draft Annex 22 applies to the manufacturing of medicinal products and active substances, which in our reading includes GMP quality control testing as part of manufacture; it does not cover clinical trials or non-clinical safety studies.
Which parts of an AI system can change
An AI agent has more moving parts than conventional software, and several can change without a code release. In our reading, each of these belongs in the controlled configuration specification that the draft revised Annex 11 describes:
The model: its provider, name and exact version, and settings that change its behaviour, such as sampling settings. The instructions: the system prompt, prompt templates and any examples given to the model. The tools: which functions, systems and records the agent can call, and with what permissions. The retrieval sources: which document sets the agent searches, how they are indexed, and any embedding model used to index them, which is itself a model. The surrounding software: the agent code, platform, infrastructure and operating system, which the draft revised Annex 11 lists by name. And the process: the intended use and the human review step around the agent.
A practical test of completeness, in our reading: if one of these changed overnight, would your records show it, and would anyone be told? If the answer is no, the configuration is not under control, whatever the change procedure says.
Vendor-pushed model updates
For a hosted model, the provider decides when new versions are released and old ones retired. The guidance already treats this as a supplier question. Annex 11 section 3.1 (2011) expects formal agreements with third parties that set out clear responsibilities, and section 4.5 expects the supplier to be assessed appropriately. The draft revised Annex 11 section 7.1 says relying on a vendor's qualification or a service provider's operation of a system does not change the requirements, 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.'
CSA gives a worked example for a software-as-a-service system with automatic updates. Under the service agreement, the vendor documents the changes, testing and test results for updates to the functions the manufacturer has identified. The manufacturer assesses the changes and their possible effect on the intended use, performs risk-based assurance testing appropriate to the impact, and keeps a record of the risk assessment and assurance activities. The example is a software-as-a-service product life cycle management system that a medical device manufacturer uses for project management, which the manufacturer judged not to be high process risk; it is not an AI model. In our reading, the pattern transfers, but the depth of testing should rise with the risk of the operation.
In our reading, the practical questions for an AI provider are: can you pin a specific model version; how much notice is given before a version changes or is retired; whether release notes describe behaviour changes; and whether you can test a new version before it reaches your process. If updates can arrive unannounced, detect them, for example by logging the model identifier the provider returns with each call where it returns one, and retest on detection. An identifier may not change when behaviour does, so pair it with scheduled reruns of a fixed test set and treat results outside the variation your validation plan allows as a detected change. In our reading, such an endpoint should not be used for operations whose output can reach a GxP record or a product decision without independent human verification. That assessment belongs to your quality function.
Prompt and template edits
Prompts are not named in Annex 11, draft Annex 22 or CSA. In our reading, a system prompt or prompt template is configuration: it is set by the regulated user, it changes what the system does, and Annex 11 section 10 brings 'system configurations' within change control. Small wording changes can alter an LLM's output in ways that are hard to predict from the text of the edit, so a prompt change is worth a documented impact assessment even when it looks cosmetic.
That does not mean every edit needs full revalidation. Following the draft Annex 22 section 10.1 principle, each change should be evaluated and the decision recorded, with a justification if no retest is done. A common, proportionate approach is to keep the acceptance test set from the original validation and rerun the affected cases, judging the results against the same written criteria, before the new prompt is released. Prompts kept in a controlled repository with versions and approvals make this traceable; prompts edited in a live console do not.
Retrieval-source changes
Where an agent answers from a document set, changing that set changes the agent's inputs. Draft Annex 22 section 10.1 brings changes to the process, and to 'physical objects the model is using as input', within the change evaluation; in our reading, a document set an agent retrieves from is the closest equivalent for an LLM agent. Section 10.4 asks for monitoring of whether inputs stay within the intended sample space. Read as principles, they suggest three cases.
Routine updates to documents that are already under document control, within the scope the agent was validated for, are usually covered by that document control plus monitoring, provided the change procedure justifies this in advance and the index is shown to return the current effective version and no longer return the superseded one. Document control alone does not check the index. Adding a new kind of source, a new site's documents or an uncontrolled collection widens the inputs, and in our reading it is closer to a change in intended use. Changing how the sources are processed, such as the embedding model, the indexing method or how documents are split, changes what the agent retrieves for the same question, so it needs a retest of retrieval on the validated test cases.
Risk-based triggers
Annex 11 section 1 bases the extent of validation on a justified and documented risk assessment, and the draft revised Annex 11 section 4.3 bases validation strategy and effort on intended use and potential risks to product quality, patient safety and data integrity. CSA asks manufacturers to consider failures that are reasonably foreseeable, not only likely. Applied to a change, in our reading, the questions that set the level of retest are these.
What does the affected operation touch: a GxP record, a decision about product, or only a draft a person rewrites? Is there an independent human check before the output can affect product or a required record, and is that check real? How much of the system does the change reach: one template, one tool, or the model underneath every operation? Can the change be detected, and reversed, if it goes wrong? Has monitoring already shown drift, a fall in performance or an incident? Write these triggers into the validation plan before go-live, so that each change gets the same questions and a recorded answer.
Change control decides more than testing. Separately from the retest decision, your change procedure should check whether the change affects a process, method or control described in a marketing authorisation or application, which may need a variation or supplement; your regulatory affairs function makes that call.
Decision table: does this change need revalidation?
The table is our illustrative starting point, not regulator text and not a CSA or GAMP classification. 'Retest affected operations' means rerunning the relevant validated test cases against the original acceptance criteria and recording the result. Because LLM output can differ between runs on the same input, in our reading the validation plan should fix how many times each case is run and how variation is judged, and reruns after a change should use the same rule. 'Revalidate' means repeating validation, complete or in parts, for the intended use. Your quality function sets the actual rules in your change procedure, and higher-risk operations may need more than the table shows.
| Change | Our suggested starting point (illustrative) | Guidance it rests on |
|---|---|---|
| New model or model version, chosen by you | Treat as significant. Update the risk assessment and test set for the new model's failure modes, then retest every operation the model serves against the original acceptance criteria; revalidate in parts or in full if the model is replaced by a different one or results change materially | Annex 11 sections 1 and 10 (2011); draft revised Annex 11 3.1(ii): significant changes need re-qualification and validation; draft Annex 22 10.1 (borrowed principle) |
| Vendor-pushed update to a hosted model | As above, plus a supplier question: obtain notice and release information, and test before use where the agreement allows. If updates can arrive unannounced, detect them (a returned model identifier alone may not show a change, so also rerun a fixed test set on a schedule) and retest on detection; in our reading, do not use that endpoint where its output can reach a GxP record or product decision without independent human verification | Annex 11 3.1, 4.5 and 10 (2011); draft revised Annex 11 7.1, 7.5(ix); CSA SaaS automatic-update example |
| Edit to the system prompt or a prompt template | Configuration change: documented impact assessment, then rerun the affected test cases before release. A justified no-retest decision may fit an edit shown not to change the text sent to the model, such as comments or metadata in a template, confirmed by comparing the rendered prompt before and after | Annex 11 section 10 (2011, configurations); draft Annex 22 10.1 (justify any decision not to retest) |
| Change to model settings, such as sampling settings or output length limits | Configuration change: rerun the affected test cases | Annex 11 section 10 (2011); draft revised Annex 11 6.6 (controlled configuration specification) |
| New or changed tool, permission or connected system | Retest affected operations, including tests that prohibited actions are still refused; revalidate in parts if the agent can now write to records it could not before, and reassess whether the use is now critical, since draft Annex 22 section 1 says LLMs should not be used in critical GMP applications | Annex 11 sections 1 and 10 (2011); draft revised Annex 11 3.1(ii) and 4.3; draft Annex 22 section 1 |
| Routine update to controlled documents the agent already searches | Usually covered by document control and monitoring, if the change procedure justifies this category in advance and the index is confirmed to return the current version and not the superseded one; no retest unless that check fails or monitoring shows a change in performance | Annex 11 sections 1 and 10 (2011); draft Annex 22 10.1 (justified no-retest), 10.3 and 10.4 (borrowed principles) |
| New kind of retrieval source, or a change to the embedding model, indexing or document splitting | Assess as a change to inputs or intended use; retest retrieval and answers on the validated test cases | Annex 11 section 10 (2011); draft Annex 22 10.1 and 10.4 (borrowed principles) |
| Platform, infrastructure or security patch outside the model | Risk-based regression testing of the affected functions, in line with your normal patch procedure | Annex 11 Principle (IT infrastructure qualified) and section 10 (2011); draft revised Annex 11 3.1(ii) (platform and operating system), 15.10 to 15.13 (updated platforms, timely patching) |
| New intended use or a change to the human review step | Revalidate for the new or changed use; removing or weakening review usually raises the risk rating | Annex 11 section 1 (2011); draft revised Annex 11 4.3 |
| No change made, but monitoring shows drift, lower performance or an incident | Investigate and assess; retest or revalidate in parts as the root cause requires | Annex 11 sections 11 and 13 (2011); draft revised Annex 11 section 14; draft Annex 22 10.3 and 10.4 (borrowed) |
Where ISPE's GAMP material fits
GAMP 5 (Second Edition) has appendices on operational change and configuration management (O6), periodic review (O8), patch and update management (S4) and artificial intelligence and machine learning (D11), according to its public table of contents. The GAMP Guide: Artificial Intelligence has an operation-phase appendix whose public table of contents lists system monitoring, incident and problem management, CAPA, operational change and configuration management, and periodic review. CSA itself names GAMP 5 (Second Edition) as an industry resource. The detailed content of these sections is in the paid guides, not in ISPE's public material, and nothing on this page summarises it. Our separate answer covers what ISPE has published about the AI guide.
Change control 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, change classification, validation and compliance stay with your quality function.
Does a change to an AI prompt need revalidation?
It needs change control and a documented decision, and usually a rerun of the affected test cases before release. No regulator text names prompts, but Annex 11 section 10 brings system configurations within change control, and in our reading a prompt is configuration. As a borrowed principle, the draft Annex 22 asks that any decision not to retest be fully justified.
Is draft Annex 22 in force?
No. As checked on 20 October 2026, the Commission's consultation page shows the consultation closed after running from 7 July to 7 October 2025, and the EudraLex Volume 4 page lists no Annex 22. The revised Annex 11 consulted on at the same time is also a draft. The 2011 Annex 11 remains current.
What if our AI vendor updates the model without telling us?
Treat it as a supplier and change-control issue. Annex 11 expects formal agreements and supplier assessment, and the draft revised Annex 11 says the contract should agree the process for releasing new versions and whether you can test them first. If updates can arrive unannounced, detect them, using scheduled reruns of a fixed test set as well as any model identifier the provider returns, and retest on detection; in our reading, do not rely on such an endpoint where its output reaches a GxP record or product decision without independent human verification.
Does FDA's CSA guidance apply to AI model changes at a drug manufacturer?
Its stated scope is medical device production and quality system software, so for drug and biologic CGMP it is useful context, not a requirement; the binding US requirements are the CGMP regulations, including 21 CFR 211.68(b), and 21 CFR Part 11 for required electronic records, subject to the enforcement discretion FDA describes in its 2003 Part 11 scope and application guidance. CSA's software-as-a-service example shows a risk-based pattern for automatic updates: vendor documentation, an assessment of the effect on intended use, testing proportionate to the impact, and a record of the risk assessment and the assurance activities.
Does changing the documents an AI agent searches count as a change?
It can. Routine updates to documents already under document control are usually covered by that control and by monitoring, provided the index is confirmed to return the current version. Adding a new kind of source, or changing how documents are indexed or the embedding model, changes what the agent retrieves and in our reading needs an assessment and a retest of retrieval.
- 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)
- FDA guidance: Computer Software Assurance for Production and Quality Management System Software (final; current version issued 3 February 2026, superseding the version of 24 September 2025)
- FDA: Computer Software Assurance guidance, full text including the SaaS automatic-update example (PDF)
- eCFR: 21 CFR 211.68 - Automatic, mechanical, and electronic equipment
- eCFR: 21 CFR 11.10 - Controls for closed systems
- FDA guidance: Part 11, Electronic Records; Electronic Signatures - Scope and Application (final, August 2003)
- European Medicines Agency: Guideline on computerised systems and electronic data in clinical trials (EMA/INS/GCP/112288/2023; final, adopted 7 March 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 →