Short answer
NIST AI RMF is a voluntary reference framework — not a certification. US buyers and federal procurement questionnaires reference it, so startups produce a self-attested alignment statement that describes how their AI risk practices line up with four functions: Govern, Map, Measure, Manage.
For a 10-person startup that already runs SOC 2 or ISO 27001, the alignment statement takes 2–4 weeks to write and costs essentially nothing beyond internal time. 70% of the evidence is already there; AI RMF just reformats it into the Govern / Map / Measure / Manage shape buyers expect. Below is what each function means, a subcategory-level crosswalk to ISO 42001 and SOC 2, and a week-by-week plan. For how AI RMF fits among the three major AI governance frameworks, see the AI governance frameworks comparison.
The four functions
NIST AI RMF 1.0 was published in January 2023. The Generative AI Profile (NIST AI 600-1) followed in July 2024. Both are structured around four functions, each broken into categories and subcategories. For a startup, you do not need to implement every subcategory — you need to pick the ones relevant to your systems and provide evidence.
| Function | Outcomes | Startup evidence that typically works |
|---|---|---|
| Govern | Culture, policies, structures and processes for AI risk management | AI policy, roles & responsibilities RACI, board-level AI risk reporting cadence, vendor due diligence for LLM providers |
| Map | Context, categorisation, intended use, impacts and risks per AI system | AI system inventory, intended-use documentation, impact assessments per system, stakeholder mapping |
| Measure | Analysis, assessment, benchmarking and monitoring of identified AI risks | Model performance metrics, bias tests, safety evaluations, red-team results, incident metrics |
| Manage | Prioritisation, treatment and monitoring of AI risks in production | Risk treatment plans, human-in-the-loop controls, rollback procedures, incident response runbooks |
Crosswalk: NIST AI RMF ↔ ISO 42001 ↔ SOC 2
This is why running AI RMF alongside ISO 42001 is cheap: evidence collected once reports against both. The 2027 SOC 2 criteria add an AI points-of-focus layer that lines up with Govern and Manage; the column below shows where SOC 2 overlaps today.
| NIST AI RMF subcategory | ISO 42001 clause / Annex A | SOC 2 (current + 2027) |
|---|---|---|
| GV.1 (Context & governance) | Clause 5.1, 5.2 (Leadership); A.2 (AI policies) | CC1.1, CC1.2 (Control environment) |
| GV.3 (Roles & responsibilities) | Clause 5.3 (Roles); A.3 (Internal organisation) | CC1.3 (Authority & responsibility) |
| GV.4 (Culture) | Clause 7.3, 7.4 (Awareness & communication) | CC2 (Communication) |
| MP.1 (Context) | A.5 (Impact assessment of AI systems) | CC3.1 (Objectives) |
| MP.3 (Intended use) | A.5.2 (Intended use documentation) | (partial) CC3.2 (Risk analysis) |
| MP.5 (Impact analysis) | A.5.4 (AI system impact assessment) | CC3.2, CC3.3 (Fraud & change risk) |
| MS.2 (Measurement plans) | A.6.2 (Objectives for AI system) | CC4.1 (Monitoring) |
| MS.4 (Quality & integrity) | A.7 (Data for AI systems) | C1, PI1 (Confidentiality, Processing Integrity) |
| MG.1 (Risk treatment) | A.6.2.4 (Reporting of concerns); A.9 (Use of AI systems) | CC9 (Risk mitigation) |
| MG.4 (Continual improvement) | Clause 10 (Improvement) | CC4.2 (Evaluate deficiencies) |
Subcategory IDs follow the NIST AI RMF 1.0 taxonomy. See the full SOC 2 to ISO 42001 crosswalk for every CC-level control.
The four-week plan
Assumes one person at 50% capacity and an existing SOC 2 or ISO 27001 evidence base. Without that base, double the duration and add a scoping week upfront.
Week 1
Read NIST AI 100-1 Core (56 pages) and the Generative AI Profile (28 pages). Draft the AI system inventory — list every system, intended use, who owns it, and whether it is generative.
Week 2
For each system, complete the Map function worksheets: intended use, context, impacted stakeholders, high-level risks. Pull existing security risk register entries where they apply; do not duplicate.
Week 3
Document Govern evidence (policy, roles, board reporting) and Measure evidence (performance metrics, bias/safety evals, red team results). Link to your existing SOC 2 or ISO 27001 artefacts where they cover the subcategory.
Week 4
Draft the alignment statement, build the subcategory coverage table, review internally. Publish on the Trust Centre or share with the asking buyer. Set a 6-month refresh cadence — the AI inventory drifts fastest.
What goes in the alignment statement
Start with the structure buyers already know. Ten to twenty pages, divided into five parts:
- Scope. Which AI systems are covered. Point at the inventory. Call out carve-outs (internal productivity tools often sit outside).
- Govern. AI policy excerpts, roles RACI, board reporting cadence, training.
- Map. Intended use documentation, impact assessments, stakeholder analysis per system.
- Measure. Performance metrics, bias and safety eval results, red-team cadence, incident counts.
- Manage. Risk register treatments, human-in-the-loop points, rollback procedures, incident response.
End with a subcategory coverage table. Four columns: subcategory ID, implementation status (Implemented / Partial / Planned / Not applicable), evidence reference, owner. Buyers scan this table first — if it is thin, the statement does not pass procurement review.
One rule
Do not over-claim. "Implemented" means an auditor could find the evidence tomorrow. "Planned" is honest and often preferred by sophisticated buyers; it shows a roadmap. The fastest way to lose procurement trust is to mark every subcategory implemented and then have a buyer ask for the artefact on a specific one and find there is nothing behind it.
Frequently asked questions
Is NIST AI RMF a certification?
No. NIST publishes it as a voluntary reference framework. There is no accredited audit, no certificate. Companies produce an alignment statement — a self-attested document that describes how their AI risk management lines up with the framework's four functions. Buyers accept it as evidence the vendor thought through AI risk; they do not treat it as an attestation.
Why do buyers ask for it then?
Three reasons. First, US federal procurement questionnaires reference it directly (Executive Order on AI, OMB M-24-10). Second, enterprise security questionnaires in 2026 added NIST AI RMF alignment as a checklist question. Third, buyers want a starting point for AI risk conversations with vendors who are not yet ISO 42001 certified — the alignment statement fills that gap cheaply.
How long does an alignment statement take to produce?
2–4 weeks for a 10-person startup if you already run SOC 2 or ISO 27001. The reason is that most of the Govern function is reusable governance evidence, and Map/Measure/Manage overlaps heavily with your existing security risk register and monitoring. Writing a new one from zero (no prior compliance work) takes 6–10 weeks.
What is in the Generative AI Profile?
NIST released the Generative AI Profile (NIST AI 600-1) in July 2024 as a companion to the core RMF. It extends the four functions with 12 risks specific to generative systems: confabulation, dangerous content, data privacy leakage, environmental impact, harmful bias, human–AI interaction harms, information integrity, information security, intellectual property, obscene content, toxicity, and value chain. If you ship generative features, your alignment statement should address this profile, not just the core.
Does the AI RMF overlap with ISO 42001?
Heavily. Roughly 70% of Annex A controls in ISO 42001 map to Govern, Map, Measure or Manage subcategories. If you run an ISO 42001 AIMS, the alignment statement is largely a reformat of evidence you already produce. The crosswalk table below spells out the specific mappings.
What does a good alignment statement look like?
10–20 pages. Four sections (one per function). Each section cites the organisational policies, processes and controls that implement the subcategory outcomes, with links or references to the actual artefact (not just text claims). A table at the end lists every NIST AI RMF subcategory and marks each as implemented, partial, planned or not applicable. Buyers scan for the table first.
Produce an AI RMF alignment statement in weeks
LowerPlane generates the statement from your existing SOC 2 or ISO 27001 evidence, keeps the coverage table current, and refreshes it when the AI inventory changes.
Book a demo