Status of the guidance this page relies on
No regulator document cited here sets supplier-qualification rules written for providers of large language models or AI agents. This page applies general computerised-system and outsourcing guidance to AI vendors and, where it does that, says it is our reading. Status as checked on 25 October 2026:
EU GMP Annex 11 (Computerised Systems, January 2011 revision, in operation since 30 June 2011) and EU GMP Chapter 7 (Outsourced Activities, revision 1, in operation since 31 January 2013) are 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, including in scope; this page quotes them as written. FDA's Contract Manufacturing Arrangements for Drugs: Quality Agreements (final guidance, November 2016) is non-binding and addresses commercial contract drug manufacturing, including contract testing laboratories; it says it does not explicitly apply to investigational or clinical trial materials, though many of its principles could be applied before commercial stages. It does not address software or AI suppliers, so this page borrows its structure only where a vendor's work touches CGMP activities. 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, Chapter 7 and draft Annex 22 are EU GMP texts for manufacturing. 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, cite the GLP framework that governs the study. In our reading, the supplier 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, supplier approval and the content of agreements belong to your quality function, with legal review of the contract.
What the guidance says about suppliers
Annex 11 section 3.1 (2011): when third parties such as suppliers or service providers are used to provide, install, configure, integrate, validate, maintain, modify or retain a computerised system or related service, or for data processing, 'formal agreements must exist between the manufacturer and any third parties, and these agreements should include clear statements of the responsibilities of the third party.' Section 3.2 says the competence and reliability of a supplier are key selection factors and 'the need for an audit should be based on a risk assessment.' Section 3.4 says quality system and audit information relating to suppliers or developers of software and implemented systems should be made available to inspectors on request. Section 4.5 says the regulated user should take all reasonable steps to ensure the system was developed under an appropriate quality management system, and that the supplier should be assessed appropriately.
The draft revised Annex 11 section 7 is more specific. 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 the regulated user 'remains fully responsible' according to the risk to product quality, patient safety and data integrity. Section 7.2 says the regulated user should, according to risk and system criticality, 'conduct an audit or a thorough assessment' of the vendor's or service provider's implemented procedures and documentation, and of the potential to leverage them rather than repeat the work. Section 7.3 asks for effective oversight against agreed service level agreements and key performance indicators. Section 7.4 asks that documentation be accessible and explainable from the regulated user's facility. Section 7.5 lists what a contract should cover when the regulated user relies on a service provider's qualification or operation of the system. Section 6.3 adds that where a system is purchased or provided as a service, a vendor-supplied requirements specification should be reviewed, approved and owned by the regulated user.
EU GMP Chapter 7 applies where a GMP activity itself is outsourced. It asks the contract giver to assess the legality, suitability and competence of the contract acceptor before outsourcing (7.5), to monitor and review its performance (7.7), and to have a written contract (7.1 and 7.14). The contract acceptor should not subcontract work without the contract giver's prior evaluation and approval (7.11) and should not make unauthorised changes outside the terms of the contract that may adversely affect quality (7.12); the contract should permit the contract giver to audit (7.17). In our reading, Chapter 7 is engaged when an AI vendor performs a GMP activity for you, which may include hosting or processing your GMP records, and not merely when you license software you run yourself; Annex 11 section 3 covers the system and service relationship in either case.
How deep the assessment should go
Neither Annex 11 nor its draft revision prescribes a format. Annex 11 says the need for an audit should be based on a risk assessment, and the draft revision offers 'an audit or a thorough assessment'. In our reading, the options range from a documentation or questionnaire-based assessment, sometimes called a postal audit, to an audit at the supplier's site or by remote session. GAMP 5 (Second Edition) has a section on supplier activities, including sub-supplier assessments, and an appendix on supplier assessment (M2), and the GAMP Guide: Artificial Intelligence has a supplier management appendix (M2), according to their public tables of contents; the detail is in the paid guides and is not summarised here.
In our reading, the factors that should set the depth for an AI vendor are: whether the AI output can reach a GxP record or a product decision; whether appropriately qualified and trained personnel check the output before it does; whether the vendor operates the system for you or only supplies software you run and control; how much you can test and monitor yourself; whether the vendor relies on another company's model; and what is already known about the vendor's quality system. A documentation assessment may suit a drafting aid whose every output a trained reviewer checks and rewrites. A static, deterministic model that classifies or predicts in a critical GMP step, the kind draft Annex 22 covers, would usually justify an audit. Record the rationale for the depth chosen, so the decision can be shown to an inspector.
Draft Annex 22 section 1 limits its scope to static models, which do not adapt during use, with deterministic output, used in critical applications with direct impact on patient safety, product quality or data integrity. It says dynamic models, models with probabilistic output, and generative AI and large language models 'should not be used in critical GMP applications', and that if generative AI or LLMs are used in non-critical GMP applications, which do not have direct impact on patient safety, product quality or data integrity, '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. So, in our reading, two early questions for any AI vendor are whether the model is fixed between released versions or changes during use, and whether the same input always gives the same output. If the draft is adopted as written, qualifying an LLM vendor for EU GMP manufacturing becomes a question about non-critical uses with trained personnel responsible for the output. The 2011 Annex 11 contains no such restriction.
What to ask about model versioning and change notification
Where the contract relies on the provider's qualification or operation of the system, the draft revised Annex 11 section 7.5(ix) expects it to agree 'on the process for release of new system versions and on the regulated user's possibility to test these prior to release.' At qualification, in our reading, the evidence to ask for is the vendor's written versioning and retirement policy and its record against it: how versions are named and pinned, how much notice past changes and retirements received, what release notes contain, whether customers can test a new version before it is applied, and whether the vendor can change prompts, filters or tools it controls without telling you. Retraining or fine-tuning that changes a model's behaviour should count as a change requiring notice.
Where the vendor's product is built on another company's model, ask who that model provider is, how its version changes reach you, and what the vendor can commit to on that provider's behalf. In our reading, if the contract cannot secure notice of model changes, the gap has to be closed by your own controls or by limiting the use. Our separate answer on change control covers what to do when a new version arrives; this section is about what to establish before you approve the vendor.
What to ask about training data and test evidence
Draft Annex 22 section 2.2, in its Principles section, says: 'Documentation for activities described in this section should be available and reviewed by the regulated user irrespective of whether a model is trained, validated and tested in-house or whether it is provided by a supplier or service provider.' In our reading, the same expectation extends in practice to the evidence the annex goes on to describe, so for a model in draft Annex 22 scope, ask the vendor for its evidence on intended use and input sample space (section 3), acceptance criteria (section 4), test data selection, labelling and pre-processing (section 5), independence of test data from training data (section 6) and the test plan and results (section 7). Section 5.6 says generation of test data or labels, 'e.g. by means of generative AI, is not recommended and any use hereof should be fully justified.'
Large language models are outside draft Annex 22 scope altogether; for them, these clauses are principles to borrow. In our reading, the practical questions are: what the vendor will disclose about the sources of training data, and at what level of detail; whether your inputs, outputs or files are used to train or improve any model, and whether that can be switched off by contract; what evaluation the vendor has run that bears on your intended use; and what you must test yourself because the vendor will not or cannot. If the vendor discloses training data only at a summary level, expect your own testing to carry most of the evidence. Your own validation remains the evidence that matters for your intended use.
What to ask about audit trails, data retention and exit
The draft revised Annex 11 section 12.9 says it should be possible to obtain a complete electronic copy of system data including audit trail data, and that 'Flat and locked files are not acceptable'. Section 17.1 says that after completion of a process, such as release of a product, GMP data and metadata, including audit trails, should be protected from deletion and changes throughout the retention period. The 2011 Annex 11 section 17 asks that archived data be checked for accessibility, readability and integrity, and that the ability to retrieve it be ensured and tested when the system changes. FDA's quality agreements guidance says the agreement should indicate that electronic records will be stored in accordance with CGMP and be immediately retrievable during the required record-keeping time frames; for US drug CGMP records, those time frames are set in 21 CFR 211.180.
In our reading, the questions for an AI vendor are: which records the system keeps of prompts, retrieved sources, outputs, tool calls and user actions, and which of those it can export in a searchable form; how long the vendor keeps them, and whether you control that period; where the data is stored and processed; and what happens to it if you leave. The draft revised Annex 11 section 7.5(viii) asks the contract to define 'an exit strategy by which the regulated user may retain control of system data'. Deletion of your data by the vendor is usually a confidentiality or data-protection term, but in a GxP context it must not destroy records you are required to keep. In our reading, the quality agreement should require that a complete copy, including metadata and audit trails, is returned in a searchable format and checked as readable before any deletion is authorised; the deletion commitment itself usually sits in the commercial contract or data processing agreement.
Quality-agreement clauses and where they come from
The table lists clauses we suggest for an agreement with an AI vendor whose system supports a GxP process. The first two columns are our illustrative summary, not regulator text; the third shows where each term usually sits; the fourth shows the text each clause rests on. FDA recommends that quality agreements be separate from, or at least severable from, commercial contracts such as master services agreements, and that they not cover general business terms such as confidentiality, pricing or limits on liability. That separation is FDA's recommendation; the EU texts cited here do not make it, but in our reading it is a sensible default. Your quality function decides which clauses apply, and your legal team drafts the wording.
| Clause (our illustrative summary) | What it should settle | Where it usually sits (our reading) | Text it rests on |
|---|---|---|---|
| Scope and responsibilities | Which system, models and services are covered, the intended use, and which party does each activity (configuration, testing, operation, monitoring, support) | Quality agreement | Annex 11 section 3.1 (2011); draft revised Annex 11 7.1 and 7.5(i)-(ii); Chapter 7 sections 7.14 and 7.15; FDA quality agreements guidance, section IV |
| Change notification and model versions | Advance notice of new model versions, retirements, behaviour-changing retraining and changes to vendor-controlled prompts, filters or tools; release information; your ability to test before release; no changes outside the agreed terms | Quality agreement | Draft revised Annex 11 7.5(ix); Chapter 7 section 7.12 where Chapter 7 applies; FDA quality agreements guidance, change control (borrowed structure); our reading for the AI-specific items |
| Incident and quality-issue reporting | What the vendor reports (outages, security events, behaviour problems, data errors), how fast, to whom, and how investigations and fixes are shared, so your own incident process can assess them | Quality agreement | Draft revised Annex 11 7.5(iii), (vi) and (vii); Annex 11 section 13 (2011) for your own obligation to report and assess all incidents |
| Right to audit and inspection support | Routine and for-cause audits, on site or remote, with access to relevant procedures and evidence; support during regulatory inspections | Quality agreement | Annex 11 sections 3.2 and 3.4 (2011); draft revised Annex 11 7.2, 7.5(iv) and (v); Chapter 7 section 7.17; FDA quality agreements guidance (routine and for-cause audits) |
| Subcontractors and model providers | Disclosure of subcontractors, including any underlying model provider and hosting provider, and notice of changes to them | Quality agreement | Chapter 7 section 7.11 where Chapter 7 applies; GAMP 5 (Second Edition) section 7.6, sub-supplier assessments (public table of contents only); otherwise our reading |
| Test and development evidence | What development, test and evaluation documentation the vendor will make available for review | Quality agreement | Draft Annex 22 section 2.2 (in scope) or as a borrowed principle; draft revised Annex 11 7.4 |
| Use of your data for training | Whether your inputs, outputs or files may be used to train or improve models | Commercial contract or data processing agreement | Our reading; FDA recommends keeping confidentiality terms out of the quality agreement |
| Audit trail, records and retention | Which records are kept, export of complete, searchable copies including audit trails, retention periods, and protection from change or deletion during retention | Quality agreement | Draft revised Annex 11 12.9 and 17.1; Annex 11 section 17 (2011); FDA quality agreements guidance (electronic records immediately retrievable); 21 CFR 211.180 for US drug CGMP records |
| Data return and exit | Return of a complete, searchable copy of your data and records, including metadata and audit trails, checked as readable before any deletion is authorised; transition support if the service ends | Quality agreement | Draft revised Annex 11 7.5(viii); Annex 11 section 17 (2011) |
| Deletion and confirmation | Deletion of your data after return and after any required retention period, with written confirmation | Commercial contract or data processing agreement | Our reading |
| Oversight and performance | Service levels and indicators you will monitor, and how the vendor reports against them | Quality agreement, with service levels often in the commercial contract | Draft revised Annex 11 7.3 and 7.5(iii); Chapter 7 section 7.7 |
After approval: keeping the supplier qualified
Qualification is not a one-time event. Chapter 7 section 7.7 asks the contract giver to monitor and review the contract acceptor's performance, and the draft revised Annex 11 section 7.3 asks for effective oversight against agreed indicators. The 2011 Annex 11 section 11 lists upgrade history and incidents among the things a periodic evaluation should cover where appropriate. In our reading, a change of model provider, a change in the vendor's ownership or data-handling terms, or a serious incident should trigger a reassessment rather than waiting for the next scheduled one. Our separate answers cover periodic review of an AI agent, how to validate one, and how to handle deviations; this page does not repeat them.
Qualifying a supplier starts with knowing which AI vendors and models 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, supplier approval and the terms of quality agreements stay with your quality function and legal counsel.
Does an AI vendor always need an on-site audit?
No. Annex 11 section 3.2 says the need for an audit should be based on a risk assessment, and the draft revised Annex 11 section 7.2 offers 'an audit or a thorough assessment' according to risk and system criticality. In our reading, a documentation assessment may suit a low-risk use where trained personnel review every output, while a system whose output can reach a GxP record or product decision would usually justify an audit, on site or remote. Record the rationale.
Is a vendor's SOC 2 report or ISO certificate enough to qualify it?
In our reading, not on its own. Such reports can be leveraged as evidence, and draft revised Annex 11 section 7.2 refers to the potential to leverage a supplier's work rather than repeat it. But they are not written against GMP expectations, so they may not show what you need on points such as notice of model changes, audit-trail export, record retention or support during inspections. Section 7.1 says the regulated user remains fully responsible.
Is draft Annex 22 in force?
No. As checked on 25 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.
Does FDA have a supplier-qualification rule for AI vendors?
Not specific to AI. FDA's 2016 quality agreements guidance is non-binding and addresses commercial contract drug manufacturing, so for an AI software vendor it is a structure to borrow, not a requirement. For finished drug products, the binding requirements include the CGMP regulations in 21 CFR Part 211, such as 211.22 on the quality control unit, which covers products made under contract by another company, and 211.68 on electronic equipment, plus 21 CFR Part 11 for required electronic records. Under 21 CFR 210.2(c), production of most phase 1 investigational drugs is exempt from Part 211.
What if the AI vendor uses another company's model?
In our reading, treat the model provider as a sub-supplier. Where the vendor performs an outsourced GMP activity, EU GMP Chapter 7 section 7.11 says it should not subcontract work without your prior evaluation and approval, and GAMP 5 (Second Edition) has a section on sub-supplier assessments. Either way, ask the vendor to disclose the model provider, explain how that provider's version changes reach you, and commit to notice of them.
- European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems (January 2011 revision, in operation since 30 June 2011)
- European Commission: EudraLex Volume 4, Chapter 7 - Outsourced Activities (revision 1, in operation since 31 January 2013)
- European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22 (closed; 7 July to 7 October 2025; checked 25 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 25 October 2026)
- FDA guidance: Contract Manufacturing Arrangements for Drugs: Quality Agreements (final, November 2016)
- FDA: Contract Manufacturing Arrangements for Drugs: Quality Agreements, full text (PDF)
- eCFR: 21 CFR 210.2 - Applicability of current good manufacturing practice regulations
- eCFR: 21 CFR 211.22 - Responsibilities of quality control unit
- eCFR: 21 CFR 211.68 - Automatic, mechanical, and electronic equipment
- eCFR: 21 CFR 211.180 - General requirements (records and reports)
- eCFR: 21 CFR Part 11 - Electronic Records; Electronic Signatures
- 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)
- GxP AI Agent Validation: 6 Risk-Based Steps
- GxP AI Model or Prompt Change: Retest by Risk
- Checking AI Agent Actions in GxP: Audit Trail Review, Access Review and Periodic Review
- AI Agent Errors in GxP and R&D: Deviations and CAPA
- What Does EU GMP Annex 22 (Artificial Intelligence) Mean for AI Agents?
- GxP Data Retention: AI Agent Logs, Audit Trails
- Private vs Public AI? Three Options for Pharma Data
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 →