Why ownership is split, and who holds each part
An AI agent sits across a business process, a computerised system, a quality system and a set of contracts. Each of those already has an owner in a regulated company, and the regulators' texts keep those owners rather than inventing one new AI owner. EU GMP Annex 11, section 2, says 'There should be close cooperation between all relevant personnel such as Process Owner, System Owner, Qualified Persons and IT', and that all personnel should have appropriate qualifications, level of access and defined responsibilities. Its glossary defines the process owner as 'the person responsible for the business process' and the system owner as 'the person responsible for the availability, and maintenance of a computerised system and for the security of the data residing on that system'. The system owner is defined by that responsibility, not by department, and in many companies sits outside IT.
The European Commission's July 2025 consultation draft of a new Annex 22 on artificial intelligence takes the same line for AI models: it asks for close cooperation during algorithm selection, model training, validation, testing and operation between parties including process subject matter experts (SMEs), QA, data scientists, IT and consultants. ISPE's public material on its GAMP Guide: Artificial Intelligence (July 2025) says the guide augments established roles and introduces new ones, including the activities of the quality assurance unit in supporting decisions and compliance. Under knowledge management, ISPE adds that it covers how specialist expertise such as data scientists and machine learning engineers fits with established roles like data and system owners.
Responsibility toward regulators stays with the company. The European Medicines Agency's reflection paper on AI in the medicinal product lifecycle (September 2024) says it is the responsibility of the clinical trial sponsor, marketing authorisation applicant or holder, or manufacturer to ensure that all algorithms, models, datasets and data processing pipelines used are fit for purpose and in line with legal, ethical, technical, scientific and regulatory standards, including GxP standards. A vendor can run the model; it does not take that responsibility off you.
This page is general information, not legal or regulatory advice. The role split below is one common way to divide the work; your own quality system and procedures decide the final allocation.
A RACI for an AI agent used in GMP work
The table assigns each lifecycle decision for an agent whose output enters a GMP record or process. An agent built on a generative or large language model should be kept to non-critical GMP work with a person accepting each output; see the draft Annex 22 note below. R is responsible for doing the work, A is accountable, C is consulted and I is informed. 'Approver' marks a required quality sign-off: where QA is shown as approver, the R&D owner stays accountable for the decision and QA's approval is a condition of it, so no row has more than one A. QA is shown as A where the decision itself is a quality decision: the extent of validation and release for live GMP use. 'R&D owner' means the process owner in the function the agent serves; 'System owner / IT' means the person responsible for the agent's availability, maintenance and data security, often but not always in IT, with the IT team that operates it. Where an agent touches batch certification records, consult the Qualified Person: Annex 11 section 15 lets only Qualified Persons certify batch release.
Neither Annex 11 nor any FDA or EMA text prescribes this matrix: it is a practical allocation, with the closest public anchor for each decision in the last column. The anchors are GMP texts. In a GLP study the quality assurance unit must stay independent of the study, so it monitors rather than approves, and the approver role passes to testing facility management under its own procedures, with the study director accountable for the study. In a clinical trial the sponsor or investigator, as the responsible party, signs off release. The section on discovery and GxP below explains both. The draft Annex 22 is a consultation text, not in force. It does not apply to generative AI or large language models, and says such models should not be used in critical GMP applications, those with direct impact on patient safety, product quality or data integrity. It is cited here as the regulators' clearest public statement of who should own intended use.
| Decision | R&D owner | QA | System owner / IT | Legal | Public anchor |
|---|---|---|---|---|---|
| Propose the agent and record it in the inventory | A/R | C | R | I | Annex 11 section 4.3: an up to date inventory of systems and their GMP functionality |
| Define the intended use and acceptance criteria | A/R | Approver | C | C | Draft Annex 22 sections 3.1 and 4.2: a process SME is responsible for the intended-use description and acceptance criteria, approved before acceptance testing |
| Decide GMP scope and validation extent | R | A | R | I | Annex 11 section 1: extent of validation based on a justified, documented risk assessment |
| Set the data and systems the agent can reach | A | C | R | C for personal data | Annex 11 section 12: access restricted to authorised persons, with access authorisations recorded; EMA reflection paper section 2.7 on personal data |
| Select the model supplier and sign the agreement | R | R for supplier assessment; approver of the quality agreement | R | A for the contract | Annex 11 section 3.1: formal agreements with clear statements of the third party's responsibilities; sections 3.2 and 4.5: supplier competence assessed, audit need based on risk |
| Approve the agent for live GMP use | R | A | R | I | Annex 11 Principle: the application should be validated; 21 CFR 211.22(c) and (d): the quality control unit approves procedures affecting product quality, under written responsibilities |
| Review and accept the agent's outputs | A/R | C | I | I | Draft Annex 22 section 1: for generative AI or LLMs in non-critical GMP use, qualified and trained personnel are responsible for output suitability; section 3.3: where a human decides and model testing has been reduced, the operator's responsibility is part of the intended use |
| Approve changes to model, prompts, tools or access | A | Approver for GMP impact | R | C if the AI Act role changes | Annex 11 section 10: changes, including configurations, made only under a defined procedure |
| Report and handle incidents | A | Approver for GMP impact and CAPA | R | C for regulatory reporting | Annex 11 section 13: all incidents reported and assessed; root cause of critical incidents drives corrective and preventive action |
| Periodic review | A | R | R | I | Annex 11 section 11: periodic evaluation to confirm systems remain valid and GMP compliant |
| Suspend or retire the agent | A | Approver of the retirement plan and record retention | R | C | Annex 11 section 17: archived data checked and retrievable after system changes; EMA reflection paper on how to suspend or decommission a model |
Who approves an agent and who owns its intended use
Approval comes in two steps that different people sign. The R&D owner decides that the agent should exist and what job it does, because only the person responsible for the business process can say what the agent is for and what it must not be used for. For GMP use, QA then approves that the agent has been shown fit for that intended use before it goes live. Annex 11 bases the extent of validation on a justified and documented risk assessment, and under 21 CFR 211.22 the quality control unit approves or rejects procedures and specifications affecting the identity, strength, quality and purity of the drug product, with its responsibilities and procedures in writing. The GMP sign-off should therefore sit in QA's written procedures rather than be decided agent by agent.
Intended use belongs to the process side, not to IT or the vendor. The draft Annex 22 says the intended use of a model should be described in detail based on in-depth knowledge of the process, and that a process subject matter expert should be responsible for the adequacy of that description, documented and approved before the start of acceptance testing; the same applies to the acceptance criteria for the model's test metrics (section 4.2). The process SME and the process owner may be different people; name both. The draft covers only static, deterministic machine-learning models in critical GMP applications and excludes generative AI and LLMs, but the allocation is a sensible default for any agent: the person who knows the process writes the intended use and acceptance criteria, and the quality role for the applicable regime approves them where the work is GxP.
ISPE's public table of contents for the GAMP AI guide places context of use and scope definition in the concept phase, and lists sections on governance for achieving compliance and on roles and responsibilities in its quality risk management and data and model governance appendices. The content of those sections is in the paid guide, so use them to check your allocation rather than as a source of specific role wording.
Who reviews outputs, handles incidents and retires the agent
The people who use the agent's output answer for it. The draft Annex 22 says that where generative AI or large language models are used in non-critical GMP applications, personnel with adequate qualification and training should always be responsible for ensuring the outputs are suitable for the intended use. It also says that where a model gives an input to a decision made by a human operator and the effort to test the model has been reduced, the description of the intended use should include the operator's responsibility, and the operator's training and consistent performance should be monitored like any other manual process. Where the reviewer signs a record, 21 CFR 11.10(j) requires written policies that hold individuals accountable and responsible for actions initiated under their electronic signatures, and 11.10(i) requires that people who develop, maintain or use those systems have the education, training and experience for their tasks. An agent can draft, route and flag; a named, trained person accepts.
Incidents need one route that everyone uses. Annex 11 section 13 says all incidents, not only system failures and data errors, should be reported and assessed, and that the root cause of a critical incident should form the basis of corrective and preventive actions. In practice the system owner and IT contain the agent, the R&D owner assesses the effect on the work, QA decides GMP impact and any CAPA, and legal advises on any regulatory or contractual notice. Under the EU AI Act, Article 26(5) will add a duty for deployers of high-risk systems to inform the provider or distributor and the market surveillance authority, and to suspend use, where they have reason to consider the use may present a risk. As amended by the Digital Omnibus on AI, Article 113 applies these high-risk obligations from 2 December 2027 for Annex III systems and 2 August 2028 for systems that are high-risk under Article 6(1) and Annex I. Whether a system is high-risk is a legal question.
Retirement is a decision, not just switching the agent off. The EMA reflection paper says that for all models, especially those with no human in the loop, a system risk management plan should cover how to trigger a suspension or decommissioning of a model and how to carry it out. ISPE's public table of contents gives the GAMP AI guide a retirement appendix covering retirement planning, execution and reporting and lessons learned. Annex 11 section 17 adds that when relevant changes are made to a system, the ability to retrieve archived data should be ensured and tested, so records the agent created stay readable after it is gone.
How the roles differ between discovery R&D and GxP work
Most discovery work starts outside GxP, and the role split can be lighter there. The EMA reflection paper says AI in drug discovery can be of low regulatory impact if non-optimal performance only affects the developer, but that if results contribute to the body of evidence presented for regulatory review, the principles for non-clinical development should be followed. It also says that SOPs implementing GxP principles on data and algorithm governance should be extended to all data, models and algorithms used in cases of high regulatory impact or high patient risk. ISPE says the GAMP AI guide is also relevant to early research outside traditional GxP-regulated areas. On the FDA side, the January 2025 draft guidance on AI to support regulatory decision-making gives recommendations to sponsors on AI used to produce information or data intended to support regulatory decisions about the safety, effectiveness or quality of drugs. It states that it does not address AI used in drug discovery, or for operational efficiencies such as internal workflows or drafting a submission that do not affect patient safety, drug quality or the reliability of study results. It is a draft, not for implementation, and contains non-binding recommendations.
The quality role also changes by regime. In EU GMP, Annex 11 lists Qualified Persons among the people who must cooperate; the Qualified Person is an EU and EEA role, and Annex 11 section 15 gives it the certification of batch release, not the approval of computerised systems. US GMP has no Qualified Person: 21 CFR 211.22 gives the quality control unit the authority to approve or reject drug products and review production records, and responsibility for approving procedures or specifications that affect the identity, strength, quality and purity of the drug product. In non-clinical studies under GLP, 21 CFR 58.31 requires testing facility management to designate a study director and assure there is a quality assurance unit, and 58.33 makes the study director the single point of study control. 21 CFR 58.35 gives the quality assurance unit a monitoring role that must be entirely separate from and independent of the people directing and conducting the study, so it inspects rather than approves; the EMA reflection paper points to the OECD advisory documents on computerised systems (No. 17) and data integrity (No. 22). In clinical trials, EMA's 2023 guideline on computerised systems uses 'responsible party' for the sponsor or principal investigator, and says the responsible party is ultimately responsible for the validation and operation of a computerised system, even when a service provider performs the work; it notes that AI-specific requirements are not covered in that guideline initially. Record which regime applies to each agent, not just 'GxP'.
| Decision | Discovery R&D (non-GxP) | GxP use |
|---|---|---|
| Approval to use | R&D owner approves, with IT security and legal review of data and contract terms | R&D owner proposes; QA approves for GMP use under written procedures; in GLP, testing facility management approves and the QAU monitors; in GCP, the responsible party signs off release |
| Intended use | Written by the R&D owner, including what would make the work GxP | Written by the process SME or owner; approved before acceptance testing |
| Output review | Scientist using the output, by the team's own rules | Trained, named person; accountable under the e-signature policy where a record is signed |
| Changes | R&D owner and IT, logged | Under the defined change procedure with an assessment of GxP impact by the regime's quality role |
| Incidents | R&D owner and IT; legal if data or contract terms are affected | Reported and assessed under the incident procedure; the regime's quality role decides GxP impact and CAPA |
| Moving into GxP | Trigger: output starts to feed a GxP record, decision or regulatory submission | Reclassify, record the regime (GMP, GLP or GCP), assess validation extent, and add the quality role that regime requires to every row above |
What legal owns
Legal's part is narrower than QA's, but several decisions need its sign-off, usually with regulatory affairs. Under the EU AI Act, Regulation (EU) 2024/1689, whether the company is a provider or a deployer of a system, whether a research exclusion in Article 2(6) or 2(8) applies, and whether a system is high-risk are legal classifications. Once the high-risk rules apply, from 2 December 2027 or 2 August 2028 depending on the category, Article 25 can turn a deployer into a provider if it changes a system's intended purpose so that it becomes high-risk. Article 4, as amended by the Digital Omnibus on AI, already requires providers and deployers to take measures to support AI literacy among staff and others operating or using AI systems on their behalf; it does not require them to guarantee any individual's level of AI literacy.
The EMA reflection paper says it is the applicant's or marketing authorisation holder's responsibility to ensure that all personal data, including data indirectly held within AI models, is stored and processed in line with EU data protection law. Supplier agreements are the third legal anchor: Annex 11 section 3.1 says formal agreements must exist with third parties that provide, configure, maintain or modify a computerised system or process data, with clear statements of their responsibilities. Legal owns the contract; QA and IT decide what it must say about quality and security.
A RACI is easiest to agree once you can see which agents and AI tools are actually in use. Atlas Bio's Secure AI for Life Sciences starts with an AI risk assessment: a map of how AI is used across your organisation today, including tools adopted without approval, where research data flows to external models and under which terms, gaps against the NIST AI RMF, and a prioritised plan. 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 AI Act classification stay with your quality, legal and regulatory functions.
Should one person own all AI agents in a biotech?
One person or function can own the governance programme and the inventory, but each agent needs its own named owner in the function it serves. Annex 11 section 2 expects cooperation between process owner, system owner, Qualified Persons and IT rather than a single owner of every system.
Is QA always involved in approving an AI agent?
For GMP use, yes: QA approves the agent as fit for its intended use under its written procedures. In GLP studies the quality assurance unit must stay independent of the study and monitors rather than approves, so in a common allocation approval sits with testing facility management, which 21 CFR 58.31 makes responsible for designating the study director and assuring there is a quality assurance unit, with the study director accountable for the study. In clinical trials the sponsor or investigator, as responsible party, signs off. For discovery work with no GxP or regulatory-submission impact, QA is usually informed rather than approving until the agent's output starts to feed GxP work.
Can the AI vendor own the agent's validation?
A vendor can do much of the work, but responsibility stays with the regulated company. The EMA reflection paper places it on the sponsor, applicant, holder or manufacturer, Annex 11 section 3.1 requires formal agreements setting out the third party's responsibilities, and EMA's clinical-trials guideline says the responsible party remains ultimately responsible for validation and operation.
Who decides whether the EU AI Act applies to an agent?
Legal, with input from the R&D owner on what the system is for and how it was developed. Whether the company is a provider or a deployer, whether a research exclusion applies and whether the system is high-risk are legal classifications that should be recorded in the agent's inventory entry.
What does the ISPE GAMP AI guide say about roles?
ISPE's public material says the guide augments established roles and introduces new ones, including the activities of the quality assurance unit in decisions and compliance, and that it covers how data scientists and machine learning engineers fit with established roles like data and system owners. The detailed role descriptions are in the paid guide.
Who can switch an agent off?
Name that person in the inventory entry before the agent goes live. Usually the system owner or IT executes the shutdown, the R&D owner or QA can order it, and the criteria are written down in advance. The EMA reflection paper asks that a system risk management plan define how to trigger a suspension or decommissioning of a model and how to carry it out.
- European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems (revision 1, 2011)
- European Commission: Draft Annex 22 - Artificial Intelligence (consultation document, July 2025)
- European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22 (closed 7 October 2025)
- ISPE: GAMP Guide: Artificial Intelligence (product page, published July 2025)
- ISPE: GAMP Guide: Artificial Intelligence, public table of contents (PDF)
- ISPE iSpeak: New GAMP Guide Provides Framework to Achieve High-Quality AI-Enabled Computerized Systems for Life Sciences
- Pharmaceutical Engineering: New GAMP Guide Addresses Challenges Posed by AI-Enabled Computerized Systems (September/October 2025)
- European Medicines Agency: Reflection paper on the use of artificial intelligence in the medicinal product lifecycle (September 2024)
- European Medicines Agency: Guideline on computerised systems and electronic data in clinical trials (EMA/INS/GCP/112288/2023)
- FDA draft guidance: Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (January 2025)
- eCFR: 21 CFR Part 11 - Electronic Records; Electronic Signatures
- eCFR: 21 CFR 211.22 - Responsibilities of quality control unit
- eCFR: 21 CFR 58.31 - Testing facility management
- eCFR: 21 CFR 58.33 - Study director
- eCFR: 21 CFR 58.35 - Quality assurance unit
- AI Act Service Desk (European Commission): Article 2 - Scope
- AI Act Service Desk: Article 4 - AI literacy
- AI Act Service Desk: Article 25 - Responsibilities along the AI value chain
- AI Act Service Desk: Article 26 - Obligations of deployers of high-risk AI systems
- AI Act Service Desk: Article 113 - Entry into force and application
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 →