Download: AI agent governance standards map for pharma R&D (one page) (PDF)
The eight documents with status, scope and primary-source link, as checked on 30 October 2026. The linked pages are the authority.
The standards map
The table below lists each document, what it covers, its status when last checked, and whether it reaches discovery research, GxP work or both. Each row names the issuing body's own page as its primary source. Statuses change: two of the eight are drafts, one is an industry initiative and one is a set of high-level principles, so re-check the linked page before relying on a row.
This page maps documents, not a governance programme. For what the programme itself contains, see What is AI agent governance in life sciences; for the first record to keep, see What should an AI agent inventory in pharma R&D record. This page is not legal or regulatory advice.
| Document (issuing body) | What it covers | Status | Discovery R&D | GxP work | Primary source |
|---|---|---|---|---|---|
| Pistoia Alliance agentic AI initiative (Pistoia Alliance) | A pre-competitive project to develop a life sciences and healthcare agent-to-agent communication protocol and an AI agent standard. Planned outputs named in the announcement: whitepapers, guidelines, reference implementations and scientific publications. | Initiative. Announced 4 September 2025 with seed funding from Genentech and a call for more sponsors. The announcement describes planned outputs; no published standard found on the Alliance's site (checked 30 October 2026). | Yes, by design: it sits under the Alliance's priority to harness AI to expedite R&D. | No regulatory standing. Useful as an emerging industry reference, not as a GxP requirement. | pistoiaalliance.org |
| Considerations for the Use of AI to Support Regulatory Decision-Making for Drug and Biological Products (US FDA) | AI models that produce information or data to support regulatory decisions on drug safety, effectiveness or quality. Proposes a risk-based credibility assessment framework tied to the model's context of use. | Draft, January 2025, docket FDA-2024-D-4689; marked draft and not for implementation (checked 30 October 2026). | No. The draft says it does not address AI used in drug discovery, or for operational efficiencies that do not affect patient safety, drug quality or study reliability. | Yes, where AI output supports a regulatory decision in the nonclinical, clinical, postmarketing or manufacturing phases. | fda.gov |
| Guiding principles of good AI practice in drug development (European Medicines Agency and US FDA, jointly) | Ten high-level principles for AI used to generate or analyse evidence across the drug product life cycle, including nonclinical, clinical, post-marketing and manufacturing phases: for example a risk-based approach, a clear context of use, adherence to standards including GxP, and data governance. | Published January 2026 (EMA page: first published 14 January 2026; FDA page current as of 14 January 2026). High-level principles, not a guideline or regulation (checked 30 October 2026). | Not addressed. Its scope is AI used to generate or analyse evidence across the drug product life cycle, naming nonclinical, clinical, post-marketing and manufacturing phases; it does not mention discovery (EMA's summary page refers to early research). | Yes. Principle 3 calls for AI to adhere to relevant standards, including Good Practices (GxP). | fda.gov |
| Reflection paper on the use of AI in the medicinal product lifecycle (European Medicines Agency) | Principles for AI and machine learning across the lifecycle, from drug discovery to post-authorisation, including data, model development, validation and regulatory interaction. | Final reflection paper, EMA/CHMP/CVMP/83833/2023. Adopted by CHMP on 9 September 2024 and by CVMP on 11 September 2024, published 30 September 2024 (checked 30 October 2026). | Partly. Discovery use can be low regulatory impact if poor performance only affects the developer; if results feed evidence for regulatory review, non-clinical principles apply. | Yes. EMA states it is the responsibility of the clinical trial sponsor, marketing authorisation applicant or holder, or manufacturer to ensure algorithms, models, datasets and pipelines are fit for purpose and in line with EU legislation, GxP standards and EMA guidelines. | ema.europa.eu |
| GAMP Guide: Artificial Intelligence (ISPE) | A risk-based life-cycle approach for AI-enabled computerised systems in GxP areas, covering the regulated company's and suppliers' responsibilities. Used alongside GAMP 5 (Second Edition). | Published industry guide, July 2025, 290 pages. Not a regulation (checked 30 October 2026). | Not its focus: it is written for GxP-regulated systems. | Yes. An industry reference for AI-enabled computerised systems in GxP areas, designed to be used alongside GAMP 5. | ispe.org |
| EU GMP Annex 22: Artificial Intelligence (European Commission) | AI models in critical GMP manufacturing applications for medicinal products and active substances. Covers static, deterministic machine-learning models; says dynamic, probabilistic, generative AI and large language models should not be used in critical GMP applications. | Draft. Consultation ran 7 July to 7 October 2025 and is closed; Annex 22 is not yet listed in EudraLex Volume 4 (checked 30 October 2026). The final text may differ from the draft. | No. Its scope is manufacturing. | GMP manufacturing only, and only once adopted. If generative AI or large language models are used in non-critical GMP uses, the draft says qualified, trained staff should always be responsible for their outputs. | health.ec.europa.eu |
| ISO/IEC 42001:2023, AI management system (ISO and IEC) | Requirements for establishing, implementing, maintaining and continually improving an AI management system within an organisation that develops, provides or uses AI. | Published international standard (ISO stage 60.60), Edition 1, December 2023, no amendment listed (ISO's page as archived 6 October 2026; iso.org blocks automated checks, so it could not be re-checked on 30 October 2026). | Yes, as an organisation-wide management system covering any AI use. | Not a GxP requirement. It can sit alongside a pharmaceutical quality system but does not replace computerised-system validation. | iso.org |
| AI Risk Management Framework, AI RMF 1.0, NIST AI 100-1 (US NIST) | Voluntary framework built on four functions: GOVERN, MAP, MEASURE and MANAGE. Companion Generative AI Profile, NIST AI 600-1 (July 2024). | Final, released 26 January 2023, intended for voluntary use. NIST states it is being revised as part of the White House AI Action Plan (checked 30 October 2026). | Yes. Voluntary and usable for any AI, including research agents. | Not a GxP requirement. Useful as shared vocabulary for an agent governance programme. | nist.gov |
Final, draft or initiative: what each status means
Final documents are published versions not marked as drafts. In this table that means NIST AI RMF 1.0, ISO/IEC 42001:2023, the EMA reflection paper and the ISPE GAMP AI guide. None of these is a law. NIST says the AI RMF is intended for voluntary use, ISO/IEC 42001 is a standard an organisation chooses to adopt, ISPE publishes industry guidance, and EMA describes its reflection paper as its current experience in a field where scientific knowledge is fast evolving.
Principles sit between the two. In January 2026 EMA and FDA jointly published ten guiding principles of good AI practice in drug development. The document says the principles identify areas where regulators, standards organisations and other bodies could work to advance good practice, which may help inform regulatory policies and guidelines in different jurisdictions. It is not itself a guideline with detailed requirements.
Drafts can still change. FDA's January 2025 guidance carries the label 'Draft — Not for Implementation' and states it contains nonbinding recommendations. The European Commission's consultation on Annex 22, together with revised Annex 11 and Chapter 4, closed on 7 October 2025; as of 30 October 2026 the EudraLex Volume 4 page still lists Annex 11 in its January 2011 revision and no Annex 22. Treat both drafts as signals of regulators' direction, not as requirements.
An initiative is a programme of work, not a document. The Pistoia Alliance announced its agentic AI work on 4 September 2025, with planned deliverables including an agent-to-agent communication protocol and an AI agent standard. A team can follow it or take part, but there is nothing in it yet to be audited against.
Separately, rules that predate these documents still apply where they cover the records or systems involved, whatever AI-specific guidance says. Where an agent's system creates, modifies, maintains or transmits electronic records required by FDA predicate rules, or electronic records submitted to FDA, 21 CFR Part 11 applies. Where it is a computerised system used in EU GMP activities, EU GMP Annex 11 applies; the current Annex 11 dates from 2011, and a revised version was consulted on alongside Annex 22.
Horizontal AI law is outside this map. The EU AI Act, Regulation (EU) 2024/1689, says in Article 2(6) that it does not apply to AI systems or models, including their output, specifically developed and put into service for the sole purpose of scientific research and development. Whether that exclusion covers a particular agent is a question for counsel; Atlas Bio's answer on whether the EU AI Act applies to AI in drug discovery discusses it.
Discovery R&D versus GxP work
The most useful question for a pharma R&D team is not which document is best but where the agent works. An agent that ranks targets, summarises literature or drafts hypotheses in discovery sits largely outside the regulatory drafts. FDA's draft says it does not address AI used in drug discovery. EMA's reflection paper says discovery use can be low regulatory impact if non-optimal performance only affects the developer.
The line moves when discovery output becomes evidence. EMA says that if discovery results contribute to the total body of evidence presented for regulatory review, the principles for non-clinical development should be followed. FDA's draft applies where an AI model produces information or data to support a regulatory decision about safety, effectiveness or quality, and the joint EMA-FDA principles cover AI used to generate or analyse evidence across the drug product life cycle, naming nonclinical, clinical, post-marketing and manufacturing phases. A team that may later submit agent-assisted work should keep records from the start that would let it show how the output was produced.
In GxP work the picture is denser. The ISPE GAMP AI guide is an industry reference for AI-enabled computerised systems in GxP areas. EMA states it is the responsibility of the clinical trial sponsor, marketing authorisation applicant or holder, or manufacturer to ensure models and pipelines are fit for purpose and in line with GxP standards, among others, and the joint EMA-FDA principles call for adherence to GxP. In manufacturing, the Annex 22 draft goes further than the others: its scope covers static, deterministic models in critical GMP uses, and it says generative AI and large language models should not be used in critical GMP applications. Agents built on large language models would, if the draft is adopted in that form, be limited to non-critical GMP uses, with qualified, trained people responsible for their outputs.
For governance that spans both worlds, ISO/IEC 42001 and the NIST AI RMF work at the organisation level. They do not tell you how to validate a model, but they give a structure for deciding which agents exist, who owns them and how risks are managed, which is where an inventory, named owners and incident handling come in.
How to use the map in practice
Start from your agent inventory. For each agent, record whether its output stays inside discovery or could become regulatory evidence or a GxP record, then read across the table to see which documents apply. The answer on what an AI agent inventory in pharma R&D should record lists the fields to capture.
Assign each agent a named owner who tracks the documents relevant to it. Who owns AI agent governance in a biotech covers how to split that accountability between the business, quality and IT.
Choose a first pilot in discovery, where regulatory exposure is lowest, and keep its records to the standard you would need if the work later fed a submission. How to run a first AI agent pilot in a biotech sets out the steps.
Plan for failure before it happens. Where an agent supports GxP work, its errors belong in the existing deviation and CAPA process; AI agent deviations, CAPA and incident handling explains how. NIST's MANAGE function also expects mechanisms to deactivate AI systems that behave inconsistently with intended use.
Re-check the drafts on a schedule. The FDA guidance and Annex 22 can be finalised, and NIST says AI RMF 1.0 is being revised. A quarterly check of each linked page is a reasonable default; update the inventory when a status changes.
Free one-page PDF of the table
A one-page PDF of the standards map, with each document's status, scope and primary-source link, is offered as a free download with this page. It carries the date it was last checked; the linked pages are the authority.
Atlas Bio publishes free governance downloads on its AI Trust & Security for Life Sciences page, including an AI vendor data-handling questionnaire, an agent inventory register, a validation plan outline and an audit-trail checklist; they are guidance documents, not software. Its AI risk assessment tooling, which produces a gap report against the NIST AI RMF, has been built and tested on synthetic data; it has not yet been used in a client assessment. Decisions on GxP scope and validation stay with your quality function.
Is there a single standard for AI agents in pharma?
No. As of 30 October 2026 none of the documents mapped here is a finished, agent-specific standard for pharma. The Pistoia Alliance's September 2025 announcement describes an AI agent standard as planned work; the rest cover AI generally or AI-enabled computerised systems.
Is FDA's AI guidance for drugs final?
No. FDA's January 2025 guidance on AI to support regulatory decision-making for drug and biological products is a draft marked not for implementation, and its scope excludes AI used in drug discovery. In January 2026 FDA and EMA also published ten joint guiding principles of good AI practice in drug development; they are high-level principles, not a final version of the draft.
Does EU GMP Annex 22 apply to AI agents?
Not as a validation route for most agents. Annex 22 is still a draft; as drafted, it does not apply to generative AI or large language models and says they should not be used in critical GMP applications. In non-critical GMP uses, it says qualified, trained staff should always be responsible for their outputs.
Do discovery research agents need GxP validation?
Usually not while the work stays in discovery: FDA's draft excludes discovery and EMA says discovery use can be low regulatory impact. If the output later becomes evidence for regulatory review, EMA expects non-clinical principles to be followed, so keep records from the start.
Should a biotech choose ISO/IEC 42001 or the NIST AI RMF?
They serve different purposes. ISO/IEC 42001 sets requirements for an AI management system that can be certified; the NIST AI RMF is a voluntary framework of outcomes. Teams can use NIST's functions as vocabulary and a management system to keep the work going.
Is the ISPE GAMP AI guide a regulation?
No. It is industry guidance published by ISPE in July 2025, used alongside GAMP 5 (Second Edition), to apply a risk-based approach to AI-enabled computerised systems in GxP work.
- Pistoia Alliance: Pistoia Alliance unveils agentic AI initiative and seeks industry funding to drive safe adoption (4 September 2025; initiative, checked 30 October 2026)
- US FDA: Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products (draft guidance, January 2025; draft as of 30 October 2026)
- US FDA: draft guidance PDF (January 2025)
- US FDA: Guiding Principles of Good AI Practice in Drug Development (joint with EMA; page current as of 14 January 2026; checked 30 October 2026)
- European Medicines Agency: Guiding principles of good AI practice in drug development (PDF, January 2026)
- European Medicines Agency: Artificial intelligence (lists the joint guiding principles, first published 14 January 2026)
- European Medicines Agency: Reflection paper on the use of AI in the medicinal product lifecycle, EMA/CHMP/CVMP/83833/2023 (adopted; published 30 September 2024; checked 30 October 2026)
- European Medicines Agency: Reflection paper PDF (final, adopted by CHMP 9 September 2024)
- ISPE: GAMP Guide: Artificial Intelligence (published July 2025; checked 30 October 2026)
- European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22 (closed; 7 July to 7 October 2025; checked 30 October 2026)
- European Commission: Annex 22 Artificial Intelligence - consultation draft (PDF)
- European Commission: EudraLex Volume 4 - Good Manufacturing Practice guidelines (no Annex 22 listed as of 30 October 2026)
- ISO: ISO/IEC 42001:2023 - Information technology - Artificial intelligence - Management system (published)
- ISO: ISO/IEC 42001:2023 catalogue page as archived 6 October 2026 (Published, stage 60.60, Edition 1, 2023-12)
- EU AI Act Service Desk (European Commission): Article 2 - Scope
- NIST: AI Risk Management Framework (AI RMF 1.0 released 26 January 2023; revision in progress per NIST, checked 30 October 2026)
- NIST AI 100-1: Artificial Intelligence Risk Management Framework (AI RMF 1.0)
- NIST AI 600-1: AI RMF Generative Artificial Intelligence Profile (July 2024)
- eCFR: 21 CFR Part 11 - Electronic Records; Electronic Signatures
- European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems (January 2011 revision)
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 →