Download: AI agent inventory register template (Excel)
One Excel workbook: the register, the field-to-framework mapping with sources, field guidance and terms of use. Free to copy and use inside your organization. General guidance, not regulatory advice.
Why an inventory comes first
Every other governance decision about an AI agent depends on knowing that the agent exists, what it is for and what it can touch. Atlas Bio's answer on what AI agent governance is in life sciences puts the inventory first of eight parts for that reason; this page sets out what each inventory entry should hold.
Two of the three frameworks mapped here ask for an inventory directly. EU GMP Annex 11, section 4.3, says 'An up to date listing of all relevant systems and their GMP functionality (inventory) should be available', and for critical systems asks for a description of data flows, interfaces with other systems and security measures. Annex 11 applies to computerised systems used as part of GMP regulated activities, so an agent working in a GMP process falls under it as part of that system. In the NIST AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1), GOVERN 1.6 reads: 'Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities.' The NIST AI RMF is voluntary.
The EU AI Act, Regulation (EU) 2024/1689, contains no general duty to keep an inventory. Its obligations depend on facts an inventory records: whether you are the provider or the deployer of a system, what its intended purpose is, and whether it is high-risk. Without the inventory, those questions cannot be answered system by system.
This page is general information, not legal or regulatory advice. Decisions on GxP scope, validation and AI Act classification belong with your quality, legal and regulatory functions.
The eight core fields and where each one maps
The table lists the eight fields every entry should have and the closest provision in each framework. Where a framework has no direct counterpart, the table says so. Most EU AI Act duties shown apply only to high-risk systems: under Article 113 as amended by the AI Omnibus, from 2 December 2027 for Annex III systems and from 2 August 2028 for systems that are high-risk under Article 6(1) and Annex I, which in pharma most often means AI in, or as, a medical device or in vitro diagnostic such as a companion diagnostic. Drug discovery is not an Annex III area, so for most discovery tools those duties are a reference rather than a requirement. One duty is not limited to high-risk systems: Article 4 requires providers and deployers of any AI system in scope to take measures to support the AI literacy of staff who operate or use it, and has applied since 2 February 2025. Record who is responsible for those measures for each agent in the owner field or the notes.
The Annex 11 column shows EU GMP provisions. Annex 11 is a GMP text; for agents used in GLP or GCP work, use the equivalent provision of the regime that applies, as explained in the GxP section below. The NIST AI RMF is voluntary, and NIST AI 600-1 entries are suggested actions.
| Field | What to record | EU GMP Annex 11 (and 21 CFR Part 11) | NIST AI RMF | EU AI Act |
|---|---|---|---|---|
| Owner | The named person accountable for the agent, usually the process owner in the business function it serves, plus the system owner who runs it | Section 2 names Process Owner, System Owner, Qualified Persons and IT; the glossary defines process owner as 'the person responsible for the business process' | GOVERN 2.1: roles and responsibilities documented and clear across the organization | Article 26(2): deployers of high-risk systems assign human oversight to natural persons with the necessary competence, training and authority; Article 4: AI literacy measures for staff, for any AI system in scope |
| Intended use | The task the agent performs, the workflow it serves, its users, and what it must not be used for | Section 4.3 inventory of systems and their GMP functionality; section 4.4 user requirements based on documented risk assessment and GMP impact | MAP 1.1: intended purposes, context-specific laws, norms and expectations understood and documented | Article 3(12) defines intended purpose as the use intended by the provider, as stated in its instructions and documentation, so record both the provider's stated purpose and your own use; under Article 25(1)(c) a deployer that changes a system's intended purpose so that it becomes high-risk is treated as its provider |
| Data and systems it can reach | The data classes it reads and writes (for example proprietary, personal or GxP records), the systems and tools it can call, and with what permissions | Section 4.3: data flows and interfaces for critical systems; section 12.1: access restricted to authorised persons; 21 CFR 11.10(d): limiting system access to authorized individuals | NIST AI 600-1 GV-1.6-003 suggests inventory entries consider data provenance and special rights for 'personal, privileged, proprietary or sensitive data'; MAP 4.1: risks of third-party data or software mapped | Article 26(4): where the deployer controls input data of a high-risk system, it must be relevant and sufficiently representative for the intended purpose |
| Model and version | The underlying model and its provider, the version in use, how it is accessed (vendor API, hosted, in-house), and any fine-tuning or retrieval sources | Section 10: changes, including configurations, made only under a defined procedure; draft Annex 22 sections 10.1-10.2: model under change and configuration control (consultation draft; applies only to static, deterministic ML models in critical GMP manufacturing uses, not to generative AI or LLMs) | NIST AI 600-1 GV-1.6-003: 'Underlying foundation models, versions of underlying models, and access modes'; MANAGE 3.2: pre-trained models monitored | Annex IV(1)(a) and (c): providers of high-risk systems document the system version and versions of relevant software; no equivalent deployer field |
| Autonomy level | How far the agent acts without a person: suggests only, drafts for approval, acts within set limits, or acts without routine review | No direct counterpart; the Principle says a system replacing a manual operation should not decrease product quality, process control or quality assurance, or increase overall risk | GOVERN 3.2: roles and responsibilities for human-AI configurations and oversight defined | Article 3(1): AI systems operate 'with varying levels of autonomy'; Article 14(3): oversight of high-risk systems commensurate with risks, level of autonomy and context of use |
| Human checkpoint | Which actions always need a person before they take effect, who that person is, and how the approval is recorded | Section 14: electronic signatures; section 15: where a computerised system records certification and batch release, it should allow only Qualified Persons to certify release, identify the person, and use an electronic signature; 21 CFR 11.10(g): authority checks so only authorized individuals sign or alter records | MAP 3.5: processes for human oversight defined, assessed and documented; NIST AI 600-1 GV-1.6-003 lists human oversight roles and responsibilities among items to consider in inventory entries | Article 14: high-risk systems designed for effective oversight by natural persons; Article 26(2) assigns that oversight |
| GxP or non-GxP | Whether the agent's output enters a GxP record, process or submission, which regime applies (for example GMP, GLP or GCP), with the reason, and the AI Act position: provider or deployer, any Article 2(6) or 2(8) exclusion claimed, and whether it is high-risk | The Principle limits Annex 11 to GMP regulated activities; section 1 bases validation extent on a documented risk assessment covering patient safety, data integrity and product quality; 21 CFR 11.1(b) sets Part 11 scope | MAP 1.1 (context-specific laws documented); GOVERN 1.1: legal and regulatory requirements involving AI understood, managed and documented | Article 2(6) excludes systems specifically developed and put into service for the sole purpose of scientific research and development; Article 2(8) excludes pre-market research, testing and development of AI systems, but not testing in real world conditions; Article 6 sets the high-risk rules; Article 3(3) and (4) define provider and deployer |
| Review date | Date of the last review and the next one, set by risk, plus the triggers that force an earlier review such as a model change or incident | Section 11: periodic evaluation to confirm systems remain in a valid state, including incidents, upgrade history, performance and security | GOVERN 1.5: ongoing monitoring and periodic review of the risk management process and its outcomes planned, 'including determining the frequency of periodic review'; MANAGE 4.1: post-deployment monitoring plans implemented | Closest is ongoing monitoring, not periodic review: Article 26(5) requires deployers to monitor the operation of high-risk systems on the basis of the instructions for use |
Fields worth adding once the core is in place
Several further fields make the register usable in an audit or an incident. Each has a direct anchor in the same frameworks, and the downloadable template includes them as optional columns.
Supplier and agreement: who provides the model or platform and which contract governs it. Annex 11 section 3.1 requires formal agreements with third parties used to provide, configure, maintain, modify or retain a computerised system or related service, or to process data, with clear statements of their responsibilities, and NIST GOVERN 6.1 asks for policies addressing AI risks from third-party entities.
How to switch it off: the mechanism, the person authorised to use it and the criteria. NIST MANAGE 2.4 asks for mechanisms, with assigned responsibilities, to supersede, disengage or deactivate AI systems whose performance is inconsistent with intended use, and NIST AI 600-1 GV-1.7-001 asks for protocols so generative AI systems can be deactivated when necessary. Annex 11 section 16 asks for alternative arrangements, such as a manual process, for systems supporting critical processes in the event of a system breakdown. Under Article 26(5) of the AI Act, deployers of high-risk systems that have reason to consider that use may present a risk within the meaning of Article 79(1) must inform the provider or distributor and the market surveillance authority without undue delay and suspend use.
Where its activity is logged and where its validation or test evidence sits. Annex 11 section 9 asks for consideration, based on risk, of an audit trail of GMP-relevant changes and deletions, and 21 CFR 11.10(e) lists secure, computer-generated, time-stamped audit trails among the controls for closed systems holding records that an FDA predicate rule requires. FDA's September 2003 guidance on Part 11 scope and application says FDA intends to exercise enforcement discretion on some Part 11 requirements, including audit trails, while predicate-rule requirements still apply. NIST MEASURE 2.1 asks that test sets, metrics and details of the tools used during testing be documented. For high-risk systems, Article 26(6) of the AI Act requires deployers to keep the logs the system generates automatically, to the extent they control them, for at least six months unless other law provides otherwise.
Status and change history: proposed, live, suspended or retired, with a link to the change record. NIST GOVERN 1.7 covers decommissioning AI systems safely, and Annex 11 section 4.2 asks that validation documentation include change control records where applicable.
Deciding GxP or non-GxP for an agent
This field changes how much the rest of the entry has to carry, so record the reasoning, not just the answer. Annex 11 applies to computerised systems used as part of GMP regulated activities. 21 CFR Part 11, section 11.1(b), applies to electronic records created, modified, maintained, archived, retrieved or transmitted under any records requirements in FDA regulations, and to electronic records submitted to FDA. An agent that drafts or changes such records, or whose output is relied on in a GxP decision, needs the GxP treatment of the system it is part of.
Annex 11 is a GMP text. It reaches R&D where the activity is GMP, for example manufacture of investigational medicinal products. For non-clinical studies under GLP, the EMA reflection paper says the OECD advisory documents on applying GLP principles to computerised systems (No. 17) and on GLP data integrity (No. 22) should be considered. For clinical trials, EMA's Guideline on computerised systems and electronic data in clinical trials (2023) applies. Part 11 applies wherever an FDA predicate rule requires the record, whether under GLP, GCP or GMP. Record which regime applies, not just 'GxP'.
Most discovery work starts outside GxP. The European Medicines Agency's Reflection paper on the use of AI in the medicinal product lifecycle (September 2024) 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. For high regulatory impact or high patient risk uses, it says GxP data and algorithm governance SOPs should be extended to cover the data, models and algorithms involved. An inventory entry marked non-GxP should therefore say what would make it GxP, so the classification is revisited when an agent's output moves towards a submission.
The AI Act position is a separate question from GxP and belongs in the same entry. The Article 2(6) research exclusion turns on why a system was developed and put into service: it covers systems specifically developed and put into service for the sole purpose of scientific research and development, so a general-purpose commercial tool is unlikely to qualify merely because a research team uses it. Article 2(8) separately excludes research, testing and development of AI systems before they are placed on the market or put into service, but not testing in real world conditions. Whether either applies is a question for your legal function. Record which role you hold for each system, whether you rely on the exclusion and why, and whether any route to high-risk status under Article 6 applies.
Where the draft Annex 22 fits
In July 2025 the European Commission consulted on a revised Annex 11, a revised Chapter 4 and a new Annex 22 on artificial intelligence, drafted by the EMA GMDP Inspectors Working Group with PIC/S; the consultation closed on 7 October 2025. As of October 2026 the Commission's EudraLex Volume 4 page still lists Annex 11 in its January 2011 revision and does not list an Annex 22, so the drafts are not in force and may change.
The consultation draft of Annex 22 applies to computerised systems used in the manufacture of medicinal products and active substances where AI models are used in critical applications with direct impact on patient safety, product quality or data integrity. It covers only static machine-learning models with deterministic output. It says dynamic models, models with probabilistic output, and generative AI and large language models should not be used in critical GMP applications, and it does not apply to generative AI or LLMs. Where such models are used in non-critical GMP applications, personnel with adequate qualification and training should always be responsible for checking that outputs suit the intended use (a human in the loop), and the draft's principles may be considered where applicable. For the models it covers, it asks for a detailed intended-use description, for which a process subject matter expert is responsible and which is documented and approved before acceptance testing, and for change and configuration control before deployment.
Most LLM-based agents therefore sit outside the draft's detailed requirements. The draft is a reason to record autonomy level, human checkpoint and model version precisely, not a requirement in force or one that applies to LLM agents.
Using the register template
The free Excel register on this page has one row per agent, the eight core fields and the optional ones as columns, drop-down lists for autonomy level, GxP classification and regime, AI Act position and status, and a second sheet repeating the field mappings above with links to their sources. A third sheet explains how to fill each column, and a fourth states the terms of use. It contains no example agents or data; the first row is a blank entry to copy.
Keep the register in a controlled location with a named owner. If you use it for GMP systems, treat it as a controlled document under your own procedures: the template does not validate anything and does not replace your computerised system inventory if you already have one; it can feed it. Start by listing every agent and AI tool in use, including those adopted without approval, give each a named owner, and fill the GxP field before anything else.
An inventory is easiest to build from a map of how AI is actually used. 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.
Does the EU AI Act require an AI inventory?
Not as such. The AI Act, Regulation (EU) 2024/1689, sets no general inventory duty. Its obligations depend on whether you are a provider or deployer of a system, its intended purpose and whether it is high-risk, which is the information an inventory records. High-risk duties apply from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I product-related systems such as AI in medical devices. Article 4, which requires measures to support the AI literacy of staff using any AI system in scope, has applied since 2 February 2025.
Does EU GMP Annex 11 require an inventory?
Yes, as EU GMP guidance for computerised systems used in GMP regulated activities. Section 4.3 says an up to date listing of all relevant systems and their GMP functionality (inventory) should be available, and asks for fuller system descriptions, including data flows, interfaces and security measures, for critical systems.
Should non-GxP research agents be in the inventory?
Yes. NIST AI RMF GOVERN 1.6 calls for inventorying AI systems generally, and recording non-GxP agents is how you notice when one starts feeding GxP work or a regulatory submission, at which point, the EMA reflection paper says, the principles for non-clinical development should be followed.
What counts as an autonomy level?
No regulation defines a scale. The AI Act notes that AI systems operate with varying levels of autonomy and asks that oversight of high-risk systems match it. A simple practical scale is suggest only, draft for approval, act within set limits, and act without routine review.
How often should an inventory entry be reviewed?
On a schedule set by risk, and whenever the model, tools, data access or intended use change. Annex 11 section 11 asks for periodic evaluation of computerised systems. NIST GOVERN 1.5 asks organisations to plan periodic review of their AI risk management process, including its frequency, and MANAGE 4.1 calls for post-deployment monitoring plans.
Who should own the inventory?
One named person or function, often quality or IT governance, should own the register itself, while each entry names its own agent owner in the business function the agent serves. Annex 11 section 2 expects cooperation between process owners, system owners, Qualified Persons and IT.
- European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems (revision 1, 2011)
- European Commission: EudraLex Volume 4 - Good Manufacturing Practice guidelines
- European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22 (closed 7 October 2025)
- European Commission: Draft Annex 22 - Artificial Intelligence (consultation document, July 2025)
- eCFR: 21 CFR Part 11 - Electronic Records; Electronic Signatures
- FDA guidance: Part 11, Electronic Records; Electronic Signatures - Scope and Application (September 2003)
- European Medicines Agency: Guideline on computerised systems and electronic data in clinical trials (EMA/INS/GCP/112288/2023)
- NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0)
- NIST AI 600-1, AI RMF: Generative Artificial Intelligence Profile
- AI Act Service Desk (European Commission): Article 2 - Scope
- AI Act Service Desk: Article 3 - Definitions
- AI Act Service Desk: Article 4 - AI literacy
- AI Act Service Desk: Article 6 - Classification rules for high-risk AI systems
- AI Act Service Desk: Article 14 - Human oversight
- 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: Annex IV - Technical documentation
- AI Act Service Desk: Article 113 - Entry into force and application
- European Medicines Agency: Reflection paper on the use of artificial intelligence in the medicinal product lifecycle (September 2024)
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 →