Status: voluntary, under revision and not written for agents
NIST released AI RMF 1.0 (NIST AI 100-1) on 26 January 2023 and describes it as intended for voluntary use. Its FAQ asks whether private or public sector organizations will be required to use the framework, and answers no. NIST's framework page also says AI RMF 1.0 is being revised as part of the White House AI Action Plan (checked 30 October 2026), so the subcategory numbers on this page are those of version 1.0.
NIST offers no certificate to earn. Neither the framework nor NIST's AI RMF FAQ describes a certification or conformity-assessment scheme, so a claim to be 'NIST AI RMF certified' has no basis in NIST's own documents. A company can say it uses the framework and show the records behind that. ISO/IEC 42001 works differently: it is a management-system standard, and ISO/IEC 42006:2025 sets requirements for bodies that audit and certify AI management systems according to ISO/IEC 42001.
The Generative AI Profile, NIST AI 600-1 (July 2024), is a cross-sectoral profile of, and companion resource for, the AI RMF. It lists suggested actions under the AI RMF subcategories, each with an ID: GV for GOVERN, MP for MAP, MS for MEASURE and MG for MANAGE, as in GV-1.6-001. Neither document is written for agents. The AI RMF 1.0 text does not use the word agent, and AI 600-1 mentions autonomous agents once, in a list of security threats under MS-2.7-001. NIST's separate AI Agent Standards Initiative (page created 17 February 2026) says NIST conducts gap analyses to produce voluntary guidelines to inform industry-led standardization for AI agents; it does not change the AI RMF 1.0 subcategories used here. Applying the AI RMF and AI 600-1 to agents is therefore an interpretation: the mapping below is Atlas Bio's, and each action ID points to NIST's own wording.
This page is narrower than its neighbours. What is AI agent governance in life sciences describes the whole operating model. What should an AI agent inventory in pharma R&D record lists the fields of one record, which is a single row here (GOVERN 1.6). Here the subject is one framework, taken function by function, against the work agents do in a pharma or biotech R&D organization. It is not legal or regulatory advice.
The four functions mapped to agent activities
NIST describes GOVERN as a cross-cutting function, meant to inform and be infused throughout the other three. It says that, once GOVERN outcomes are in place, most users start with MAP and continue to MEASURE or MANAGE, that the functions may be performed in any order, and that the actions do not constitute a checklist. Read the table in that spirit: for each agent activity, the function and subcategory it falls under, the AI 600-1 actions that apply most directly, and the record a pharma R&D team would keep.
| Agent activity | AI RMF function and subcategory | NIST AI 600-1 actions (ID: what NIST suggests) | Record to keep in pharma R&D |
|---|---|---|---|
| Inventory of agents | GOVERN 1.6: mechanisms are in place to inventory AI systems. | GV-1.6-001: enumerate generative AI systems for the AI system inventory. GV-1.6-003: consider in each entry data provenance, known issues, human oversight roles, special rights for proprietary or sensitive data, and the underlying foundation models, their versions and access modes. | One entry per agent: owner, purpose, model and version, data it reads, tools it calls and who approves its output. |
| Acceptable use and go/no-go | GOVERN 1.3: the level of risk management is set by risk tolerance. GOVERN 3.2: roles for human-AI configurations and oversight are defined. | GV-1.3-002: set minimum performance or assurance thresholds as part of go/no-go deployment approval. GV-3.2-003: define acceptable use policies for interfaces and human-AI configurations, including the kinds of queries the application should refuse. | Approval criteria per agent tier, and the outputs that need a named person's sign-off before anyone relies on them. |
| Third-party models and tools | GOVERN 6.1 and 6.2: third-party risks are addressed and contingency processes exist for failures or incidents in third-party data or AI systems deemed high-risk. | GV-6.1-007: inventory third parties with access to organizational content and keep approved provider lists. GV-6.1-009: update vendor due diligence to cover intellectual property, data privacy, security and other risks. GV-6.2-003: set incident response plans for third-party generative AI technologies. | Approved model and vendor list, contract terms on data use, and a supplier assessment where the vendor supports GxP work. |
| Context of use, data and permissions | MAP 1.1: intended purposes and settings are understood and documented. MAP 2.2: knowledge limits and human oversight of output are documented. | MP-1.1-001: consider internal versus external use, scope, fine-tuning and data sources such as retrieval-augmented generation. MP-1.1-002: document the expected and acceptable context of use with domain experts. MP-2.2-001 and MP-2.2-002: document upstream data dependencies and how the system interacts with external networks. | A context-of-use statement: discovery only, or possible regulatory evidence; the data classes and systems, such as an ELN or LIMS, the agent may read or write; the external sites it may reach. |
| Change of use | MAP 4.1: approaches for mapping technology and legal risks, including third-party data or software, are in place. | MP-4.1-008: re-evaluate risks when adapting a model to a new domain, and set warning systems for use in a new domain where earlier assumptions may no longer hold. | A trigger to re-review the agent when its output starts to feed a regulatory submission or a GxP record. |
| Testing before use | MEASURE 2.3: performance is measured for conditions similar to deployment. MEASURE 2.5: the system is shown to be valid and reliable, and the limits of generalization are documented. | MS-2.5-001: avoid extrapolating performance from narrow, anecdotal assessments. MS-2.5-003: review and verify sources and citations in outputs, before deployment and in ongoing monitoring. MS-2.3-003: share pre-deployment test results with those who hold release approval authority. | A test set drawn from the agent's real tasks (for a literature agent, a check of each citation against the cited paper), signed off by the approver. |
| Security testing | MEASURE 2.7: security and resilience are evaluated and documented. | MS-2.7-001: assess threats including compromised dependencies, data breaches and autonomous agents. MS-2.7-007: red-team for resilience against attacks including prompt injection and data poisoning. | A red-team record covering instructions hidden in documents, web pages or tool output the agent reads. |
| Monitoring in use | MEASURE 2.4: functionality and behaviour are monitored in production. MANAGE 4.1: post-deployment monitoring plans are implemented. | MG-4.1-002: establish and evaluate post-deployment monitoring, particularly for confabulation, CBRN or cyber risks. | Sampled review of outputs, errors by task type and logs of the tools the agent called. |
| Incidents and switching off | GOVERN 1.7: decommissioning is planned. GOVERN 4.3: practices enable AI testing, identification of incidents and information sharing. MANAGE 2.4: responsibility is assigned to supersede, disengage or deactivate. MANAGE 4.3: incidents and errors are tracked and communicated. | GV-1.7-001: protocols so systems can be deactivated when necessary. MG-2.4-002 and MG-2.4-004: escalation procedures, and reviewed criteria that warrant deactivation. GV-4.3-002: a minimum set of incident reporting fields. MG-4.3-002: record and track errors, near-misses and negative impacts. | Named owner of the off switch, written deactivation criteria and an incident log; for GxP work, a route into deviation and CAPA. |
GOVERN: the inventory, owners and rules come first
GOVERN is where an agent programme starts, because the other functions need something to point at. GOVERN 1.6 asks for mechanisms to inventory AI systems, and AI 600-1 suggests enumerating generative AI systems for that inventory (GV-1.6-001). GV-1.6-003 lists what an entry may hold, including the underlying foundation models, their versions and access modes. For an agent that matters in practice: a model update from the vendor can change the agent's behaviour without any change to the agent itself, so the version belongs in the record.
The same action names special rights and considerations for proprietary or sensitive data. In pharma R&D that is the field that shows which agents can see unpublished assay results, compound structures or sequence data, and under which vendor terms.
GOVERN 2.1 asks for roles and responsibilities to be documented and clear across the organization, and GOVERN 3.2 for roles in human-AI configurations and oversight. Each agent needs a named owner and a named approver for its output; Who owns AI agent governance in a biotech sets out how to split that between R&D, QA, IT and legal. GV-1.3-002 adds the go/no-go step: minimum performance or assurance thresholds, agreed before the agent is approved for use.
Third parties sit in GOVERN too. GV-6.1-007 suggests an inventory of third parties with access to organizational content and approved lists of providers. Where an agent supports GxP work, the AI vendor may also need qualifying as a supplier, which is a quality-system task the AI RMF does not replace.
MAP: context of use decides how much of the rest applies
MAP 1.1 asks that intended purposes, settings and context-specific laws and norms be understood and documented. MP-1.1-002 suggests documenting the expected and acceptable context of use with domain experts. For an R&D agent the most useful single fact is whether its output stays inside discovery or could become evidence: a literature-triage agent and an agent that drafts sections of a study report carry different risks even if they use the same model.
Permissions are part of context. MP-1.1-001 points to the data sources an application draws on, such as retrieval-augmented generation, and MP-2.2-002 to how the system interacts with external networks. For an agent the practical record is which systems it can read, which it can write to, which external sites it can reach and which actions need a person first. What to check before giving an AI agent access to your ELN or LIMS covers that step in detail.
Context also changes. MP-4.1-008 suggests re-evaluating risks when a model is adapted to a new domain and setting warnings for when earlier assumptions may no longer hold. The common case in pharma is an agent built for discovery whose outputs start appearing in regulated work; that move should trigger a fresh MAP review, not happen by drift.
MEASURE: test what the agent does, not only the model
NIST says AI systems should be tested before deployment and regularly while in operation. MEASURE 2.3 asks for performance to be shown in conditions similar to deployment, and MEASURE 2.5 for the system to be shown valid and reliable, with the limits of generalization documented. For agents that means testing whole tasks, including the tool calls, not only the answers of the underlying model.
Two AI 600-1 actions translate directly. MS-2.5-001 warns against extrapolating performance from narrow, anecdotal assessments, which rules out approving an agent on a handful of good demonstrations. MS-2.5-003 asks for sources and citations in outputs to be reviewed and verified before deployment and during monitoring; for an agent that summarizes literature or patents, checking each citation against the cited document is the core test.
Security belongs in MEASURE as well. MS-2.7-007 suggests red-teaming for resilience against attacks including prompt injection and data poisoning. An agent that reads external papers, web pages or supplier documents can meet instructions hidden in that content, so the red-team set should include them; Prompt injection risk for research AI agents explains the attack.
Who tests matters. MEASURE 1.3 asks that internal experts who did not serve as front-line developers of the system, or independent assessors, be involved in regular assessments, and MS-2.3-003 that pre-deployment results be shared with those who hold release approval authority. In a small biotech that can be as simple as a scientist outside the build team reviewing the test results before the R&D lead signs the go decision.
MANAGE: monitoring, incidents and the off switch
MANAGE 4.1 asks for post-deployment monitoring plans, and MG-4.1-002 for monitoring processes particularly for confabulation, CBRN or cyber risks. For a research agent the realistic monitor is a sampled review of outputs by a qualified person plus the agent's own logs of what it read and which tools it called.
Changes are a MANAGE question too. MG-3.1-003 suggests re-assessing model risks after fine-tuning or retrieval-augmented generation is added, and for third-party models used in ways not covered by the initial testing. In GxP work that maps onto change control: Change control for AI models and prompts in GxP covers when an update needs revalidation.
MANAGE 2.4 asks for mechanisms, and assigned responsibilities, to supersede, disengage or deactivate AI systems whose performance or outcomes are inconsistent with intended use. GV-1.7-001 suggests protocols so systems can be deactivated when necessary, and MG-2.4-004 specific, regularly reviewed criteria that warrant deactivation. Writing those criteria down before launch, with a named person able to act on them, is what turns an off switch from an idea into a control.
MANAGE 4.3 asks that incidents and errors be communicated and that processes for tracking, responding to and recovering from them be followed and documented. MG-4.3-002 suggests recording reported errors and near-misses as well as harms, and MG-4.3-001 after-action assessments that include communicating to relevant legal and regulatory bodies where applicable. Where an agent supports GxP work, errors that reach, or could affect, a GxP record, decision or product should be assessed under the existing deviation and CAPA procedures, while errors caught by the designed human review are tracked as monitoring data and trended; When an AI agent gets it wrong: deviations, CAPA and incident handling explains how.
Where the Annex 22 draft and ISO/IEC 42001 overlap
EU GMP Annex 22 is a draft. The European Commission consulted on it from 7 July to 7 October 2025 and the consultation is closed; as of 30 October 2026 the EudraLex Volume 4 page does not list an Annex 22. The draft applies to computerised systems used in manufacturing 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 static models with deterministic output, does not apply to generative AI and large language models, and says they should not be used in critical GMP applications. In non-critical GMP uses, it says personnel with adequate qualification and training should always be responsible for ensuring their outputs are suitable for the intended use.
So for agents built on large language models, the draft is not a validation route: it says such models should not be used in critical GMP applications, and in non-critical GMP uses it expects qualified personnel to be responsible for the outputs, with its principles to be considered where applicable. Where an agent calls a static machine-learning model with deterministic output inside a critical manufacturing step, the draft could apply to that model directly once adopted, and the final text may differ from the draft. Because Annex 22 is a GMP manufacturing annex, agent work in discovery, nonclinical or clinical research falls outside its scope.
For ISO/IEC 42001, NIST publishes its own crosswalk from the AI RMF. It was prepared against the final draft (FDIS) text of ISO/IEC 42001, so check the clause numbers below against the published 2023 standard before relying on them. The ISO/IEC 42001 column lists selected items that NIST's crosswalk pairs with each AI RMF subcategory, not the full mapping. The Annex 22 draft does not mention the AI RMF, so the Annex 22 column is Atlas Bio's alignment of the draft's sections, and says so where there is no matching section.
| AI RMF area | EU GMP Annex 22 draft (GMP manufacturing; detailed requirements for critical uses) | Selected ISO/IEC 42001 items in NIST's crosswalk |
|---|---|---|
| Inventory (GOVERN 1.6) | No inventory section. Section 2.2 says documentation should be available and reviewed by the regulated user whether a model is trained in-house or supplied. | B.4.2 Resource documentation; B.4.3 to B.4.6 data, tooling, system and computing, and human resources. |
| Roles and oversight (GOVERN 2.1, GOVERN 3.2) | Section 2.1: close cooperation between process subject matter experts, QA, data scientists, IT and consultants, all with adequate qualifications, defined responsibilities and an appropriate level of access. | GOVERN 2.1: 5.3 Roles, responsibilities and authorities, and B.3.2 AI roles and responsibilities. GOVERN 3.2: B.6.1.3, B.9.3 and B.4.6, among others. |
| Context of use (MAP 1.1) | Section 3.1: intended use described in detail, including the input sample space, with a process subject matter expert responsible and the description approved before acceptance testing. | 6.1.4 AI system impact assessment; B.5.2 AI system impact assessment process. |
| Thresholds and go/no-go (GOVERN 1.3, GV-1.3-002) | Section 4.2: acceptance criteria defined by a subject matter expert and approved before testing. Section 4.3: performance at least as high as the process the model replaces. | GOVERN 1.3: 6.1.1 General, 6.1.2 AI risk assessment, 6.1.3 AI risk treatment. |
| Testing (MEASURE 2.3, MEASURE 2.5) | Sections 5 to 7: representative test data, independence of test data from training data, an approved test plan, and documented, investigated deviations. | MEASURE 2.3: B.7.4 Quality of data for AI systems. MEASURE 2.5: B.6.2.4 AI system verification and validation. |
| Monitoring (MEASURE 2.4, MANAGE 4.1) | Section 10.3: regular monitoring of model performance. Section 10.4: monitoring whether input data stay within the sample space, with metrics for drift. | MEASURE 2.4: 9.1 Monitoring, measurement, analysis and evaluation; B.6.2.6 AI system operation and monitoring; B.6.2.8 AI system recording of event logs. |
| Human review (GOVERN 3.2, MAP 3.5) | Sections 3.3 and 10.5: where a model gives input to a human operator's decision and the effort to test the model has been reduced, the intended use should include the operator's responsibility, the operator's performance should be monitored, and records of the review should be kept (depending on criticality, review of every output). Section 1: for non-critical use of generative AI, qualified staff are responsible for the outputs. | MAP 3.5: B.6.1.3, B.6.2.7 AI system technical documentation and B.8.2. |
| Changes (MG-3.1-003) | Section 10.1: model, system and process under change control, with any change evaluated for retest. Section 10.2: configuration control and detection of unauthorized change. | NIST's crosswalk has no single item for change management, so none is listed here. |
| Incidents and deactivation (MANAGE 2.3, 2.4, 4.3) | No incident or deactivation section in the draft; section 7.3 covers deviations during testing only. | MANAGE 2.3 (response to newly identified risks): 10.2 Nonconformity and corrective action. MANAGE 2.4: B.9.4 Intended use of the AI system. MANAGE 4.3: 9.3.2 Management review inputs; B.8.5 Information for interested parties. |
A starting order for a small biotech
First, build the inventory (GOVERN 1.6) and name an owner and an approver for each agent (GOVERN 2.1). Second, write a one-paragraph context of use for each, saying whether its output could become regulatory evidence (MAP 1.1). Third, agree go/no-go thresholds and test the agent on its real tasks, including a citation check and a prompt-injection test (GV-1.3-002, MS-2.5-003, MS-2.7-007). Fourth, write the deactivation criteria and the incident route before launch (MG-2.4-004, MG-4.3-002).
Keep the evidence as you go. Because the framework is voluntary, the value to a partner, sponsor or auditor is in the records: the inventory, the context-of-use statements, the test results and the incident log. Re-check NIST's framework page each quarter while AI RMF 1.0 is being revised, and the European Commission's EudraLex pages for the final Annex 22.
Atlas Bio publishes free governance downloads on its AI Trust & Security for Life Sciences page, including an AI vendor data-handling questionnaire mapped to the NIST AI RMF and AI 600-1, and an agent inventory register; they are guidance documents, not software. Its AI risk assessment tooling, which produces a gap report against the NIST AI RMF subcategories, 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 the NIST AI RMF mandatory for pharma companies?
No. NIST's FAQ says organizations will not be required to use the AI RMF and that NIST produced it as a voluntary framework. It does not replace GxP rules such as 21 CFR Part 11 or EU GMP Annex 11 where those apply to the records or systems an agent touches.
Can a company be NIST AI RMF certified?
Neither the framework nor NIST's AI RMF FAQ describes a certification scheme, so a 'NIST AI RMF certified' claim has no basis in NIST's documents. A company can describe how it uses the framework. Certification of an AI management system is an ISO/IEC 42001 matter: ISO/IEC 42006 sets requirements for the bodies that audit and certify against it.
Does NIST AI 600-1 cover AI agents?
Not specifically. AI 600-1 is a profile for generative AI, and it mentions autonomous agents once, among security threats in MS-2.7-001. Its actions on inventory, testing, monitoring and incidents still apply to agents built on generative models, which is how this page uses them. NIST's separate AI Agent Standards Initiative says it will produce voluntary guidelines for AI agents.
Does following the AI RMF satisfy GxP validation?
No. The AI RMF is a risk-management framework, not a validation method. Agents that support GxP work still need validation under the company's quality system, proportionate to risk and intended use; the AI RMF records can feed that work but do not replace it.
How does the AI RMF relate to EU GMP Annex 22?
They overlap on intended use, acceptance criteria, testing, monitoring and human review. Annex 22 is still a draft, covers critical GMP manufacturing uses only, and says generative AI and large language models should not be used in critical GMP applications.
Where should a biotech start with the AI RMF for agents?
With GOVERN 1.6, the inventory. NIST AI 600-1 suggests enumerating generative AI systems for it (GV-1.6-001) and lists fields to consider (GV-1.6-003). Then name owners and write each agent's context of use.
- NIST: Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023 (PDF)
- NIST: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024 (PDF; source of every action ID on this page)
- NIST: AI Risk Management Framework (voluntary use; AI RMF 1.0 being revised; checked 30 October 2026)
- NIST: AI Risk Management Framework FAQs (organizations not required to use the framework)
- NIST: AI RMF to ISO/IEC FDIS 42001 AI management system crosswalk (PDF)
- NIST Trustworthy and Responsible AI Resource Center: AI RMF crosswalks
- European Commission: Annex 22 Artificial Intelligence, consultation draft (PDF)
- European Commission: Consultation on EudraLex Volume 4 Chapter 4, Annex 11 and new Annex 22 (closed; 7 July to 7 October 2025)
- European Commission: EudraLex Volume 4, Good Manufacturing Practice guidelines (no Annex 22 listed, checked 30 October 2026)
- eCFR: 21 CFR Part 11, Electronic Records; Electronic Signatures
- European Commission: EU GMP Annex 11, Computerised Systems (revision January 2011, PDF)
- NIST: AI Agent Standards Initiative (created 17 February 2026, updated 14 August 2026)
- ISO: ISO/IEC 42001:2023, AI management system (published December 2023)
- ISO: ISO/IEC 42006:2025, requirements for bodies providing audit and certification of AI management systems
- What Is AI Agent Governance in Pharma R&D and GxP?
- AI Agent Inventory for Pharma R&D: 8 Core Fields
- What Is ISO/IEC 42001 and Does a Biotech Need It?
- What Does EU GMP Annex 22 (Artificial Intelligence) Mean for AI Agents?
- AI Agent Errors in GxP and R&D: Deviations and CAPA
- AI Trust & Security for Life Sciences
- Technical evidence: what Atlas Bio has built
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 →