Short answer: the NIST AI Risk Management Framework (AI RMF 1.0) is a voluntary US framework that sorts AI risk work into four functions, Govern, Map, Measure and Manage. A SaaS company can put it to work by keeping five records: an AI inventory, a use-case record for each AI feature, a test and monitoring plan, a risk register with named owners, and an incident and rollback procedure. NIST does not certify anyone, so you cannot be "NIST certified". You can say you align to it, and show the records. US enterprise buyers and several questionnaires cite it, so one clear mapping saves repeat work.
This guide turns the four functions into things a product team can build, shows what the Generative AI Profile adds, and places the framework beside ISO/IEC 42001, SOC 2 and the EU AI Act. It is an engineering view, not legal advice.
What is the NIST AI RMF, and is it mandatory?
NIST published AI RMF 1.0 as NIST AI 100-1 in January 2023. NIST describes it as voluntary, rights-preserving, non-sector-specific and use-case agnostic. No law makes a SaaS company adopt it, and NIST runs no certification or audit scheme for it. NIST also says that AI RMF 1.0 is being updated and a revised version is in progress. Until that lands, cite it as "AI RMF 1.0".
The framework has three parts that matter in practice:
- The Core. Four functions, split into categories and subcategories. You map to this.
- The Playbook. NIST's voluntary online companion, with suggested actions.
- Profiles. Use-case or cross-sector versions of the Core, such as the Generative AI Profile (NIST AI 600-1), covered below.
The framework also names seven characteristics of a trustworthy AI system: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. Map your tests to these.
Why do enterprise buyers ask about it?
Because other tools they already use point to it. The Shared Assessments SIG has an Artificial Intelligence domain that maps to ISO/IEC 42001 and NIST AI 100-1. The Cloud Security Alliance AI Controls Matrix v1.1, released on 22 June 2026, maps to ISO 42001, ISO 27001, the EU AI Act and the NIST AI RMF and AI 600-1, and includes the AI-CAIQ questionnaire. If you answer these forms often, one NIST mapping feeds all of them. Our security questionnaire service covers those forms.
How do the four functions map to a SaaS company?
| Function | Categories | The question it answers | Records a SaaS team keeps |
|---|---|---|---|
| Govern | Categories: 6 | The question it answers: Who decides, and what are the rules? | Records a SaaS team keeps: AI policy, AI inventory, named owners, vendor rules, retirement steps |
| Map | Categories: 5 | The question it answers: What does each AI feature do, and what can go wrong? | Records a SaaS team keeps: Use-case record per feature, data flow, third-party components, impact notes |
| Measure | Categories: 4 | The question it answers: How do we know it works and keeps working? | Records a SaaS team keeps: Test sets and results, monitoring dashboards, security and privacy test reports |
| Manage | Categories: 4 | The question it answers: What do we do about the risks and the incidents? | Records a SaaS team keeps: Risk register, release approvals, kill switch, incident log |
Counting the tables in AI 100-1, the Core has 19 categories and 72 subcategories. NIST says these are not a checklist or a fixed order. It also calls Govern a cross-cutting function that runs through the other three. After Govern is in place, most teams start with Map, then move to Measure and Manage and loop back as things change. You pick the subcategories that fit your risk. A firm with one summarisation feature needs far less than one that makes decisions about people.
What does Govern mean for a SaaS company?

Govern is about decisions and accountability, not model code. The subcategories most useful to a SaaS company turn into documents and routines:
| NIST reference | What to build | Evidence to keep |
|---|---|---|
| GOVERN 1.1 Legal and regulatory requirements are understood, managed and documented | What to build: An obligations register: customer contracts, privacy law such as the DPDP Act, EU AI Act exposure, sector rules | Evidence to keep: The register, with an owner and a review date |
| GOVERN 1.2 and 1.4 Trustworthiness in policies and processes | What to build: An AI policy covering product use, staff use and data handling | Evidence to keep: A versioned policy and its approval record |
| GOVERN 1.6 Mechanisms to inventory AI systems | What to build: An inventory of every model, AI feature, vendor API and internal AI tool | Evidence to keep: A dated export, refreshed on a schedule |
| GOVERN 2.1 and 2.3 Roles, and executive responsibility | What to build: A named owner per AI system and an executive who signs off high-risk releases | Evidence to keep: A role matrix and sign-off records |
| GOVERN 6.1 and 6.2 Third-party AI risk and contingency | What to build: A review step before you adopt a model provider, and a fallback if it fails | Evidence to keep: A vendor file and a tested fallback |
Build the inventory first. You cannot assess or explain AI you have not listed, and every buyer questionnaire starts there. Include AI that staff use as well as AI in the product.
How do you map the risks of each AI feature?
Map is a short record per feature, written before launch and updated at each major change. Useful fields, each tied to a Map subcategory:
- Purpose and users (MAP 1.1). What the feature is for, who uses it, and what it must not be used for.
- Model and access mode (MAP 2.1, 4.1). Task type, provider, model name and version, and whether it runs through an API or on your own infrastructure.
- Data (MAP 4.1). What data the model sees, where it is stored, and whether the provider may train on it.
- Knowledge limits (MAP 2.2). Where the feature is known to fail and how a person should treat its output. A model card (Mitchell and colleagues, 2019) is one way to write this down.
- Human oversight (MAP 3.5). Who reviews outputs, when, and with what power to override.
- Impacts (MAP 5.1). Who could be harmed or helped, how likely it is and how large the effect would be.
Keep these records short and in version control next to the code, so a change to the model or prompt triggers a review of the record. A stale record tells a buyer something that is no longer true.
How do you measure AI risk in a live product?

Engineering teams already have testing habits. The framework asks you to make them deliberate and documented:
- Start with the biggest risks (MEASURE 1.1). Pick metrics for the most significant risks from your Map records, and write down any risk you will not or cannot measure, with the reason.
- Version your test sets (MEASURE 2.1). Record the test data, metrics and tools used, so a result can be repeated.
- Test under production-like conditions (MEASURE 2.3). Use realistic prompts, data and tenant settings.
- Monitor in production (MEASURE 2.4). Track quality, refusals, errors, latency and cost, and alert on drift.
- Test security and resilience (MEASURE 2.7). For LLM features, include prompt injection, data leakage between tenants and misuse of connected tools.
- Examine privacy and fairness where they apply (MEASURE 2.10, 2.11). Check what personal data appears in prompts, logs and outputs. Test for bias if the feature affects people.
- Use someone other than the builders (MEASURE 1.3), and let users report problems (MEASURE 3.3). Feed both back into your metrics.
Re-run the tests whenever the model, prompt, retrieval index or provider changes. Providers update models, and a feature that passed its tests before an update can fail after it.
What does Manage require when something goes wrong?
Manage turns findings into decisions and controls. The practical items are:
| NIST reference | What to build |
|---|---|
| MANAGE 1.2 and 1.3 Prioritise and respond to risks | What to build: A risk register that ranks risks by impact and likelihood. Record the response for each high risk: mitigate, transfer, avoid or accept. |
| MANAGE 1.4 Residual risk is documented for acquirers and users | What to build: A customer-facing note on known limits and remaining risks, which also answers many questionnaire items. |
| MANAGE 2.4 Ability to supersede, disengage or deactivate an AI system | What to build: A kill switch or feature flag per feature and, where customers want it, per tenant, with a named person allowed to use it. |
| MANAGE 3.1 and 3.2 Monitor third-party resources and pre-trained models | What to build: Pinned model versions, a watch on provider change notices, and re-testing before you adopt a new version. |
| MANAGE 4.1 and 4.3 Post-deployment monitoring, and incident communication | What to build: Monitoring plans that include appeal and override, incident response and decommissioning, plus a written incident log and customer communication steps. |
MANAGE 4.3 asks you to communicate incidents and keep a record. It is not a legal deadline. Legal clocks come from elsewhere. Examples are CERT-In's 6-hour reporting rule, the 72-hour detailed report to the Board under DPDP Rule 7 (from 13 May 2027), and, for providers of high-risk AI systems in the EU, Article 73 of the AI Act. As of November 2026, India has no AI-specific incident reporting rule. Our guide to AI incident response covers how to run one.
What does the Generative AI Profile (NIST AI 600-1) add?
NIST AI 600-1 was published on 26 July 2024. It is a voluntary, cross-sector profile of AI RMF 1.0 for generative AI, and it adds suggested actions organised under the same four functions. It names 12 risks that are new or made worse by generative AI: CBRN information or capabilities; confabulation; dangerous, violent or hateful content; data privacy; environmental impacts; harmful bias or homogenisation; human-AI configuration; information integrity; information security; intellectual property; obscene, degrading or abusive content; and value chain and component integration.
Four of these come up most often in SaaS reviews:
| Risk in AI 600-1 | What it looks like in a SaaS product | Typical control |
|---|---|---|
| Confabulation | What it looks like in a SaaS product: A feature states something false with confidence | Typical control: Grounding in retrieved sources, citations, evaluation sets, a visible limit notice |
| Data privacy | What it looks like in a SaaS product: Personal or customer data reaches a model, a log or another tenant | Typical control: Redaction, tenant isolation, retention limits, no-training terms with providers |
| Information security | What it looks like in a SaaS product: Prompt injection, data exfiltration, abuse of connected tools | Typical control: Input and output checks, least-privilege tool access, security testing |
| Value chain and component integration | What it looks like in a SaaS product: A provider changes a model or fails, or a dataset turns out to be tainted | Typical control: Version pinning, vendor review, a tested fallback |
Two actions in the profile are easy to adopt. For the inventory (GV-1.6), it suggests that entries for generative AI systems cover data provenance, known issues, human oversight roles, intellectual property and sensitive-data considerations, and the underlying foundation models, their versions and access modes. For third-party models (MG-3.1), it suggests re-assessing risk after fine-tuning or retrieval-augmented generation, and for any third-party model used in a use case that was not evaluated at the start.
How does the NIST AI RMF relate to ISO 42001, SOC 2 and the EU AI Act?
| What it is | Can you be certified? | Where it fits | |
|---|---|---|---|
| NIST AI RMF 1.0 | What it is: A voluntary risk management framework | Can you be certified?: No. NIST runs no certification | Where it fits: Your risk method and vocabulary |
| ISO/IEC 42001 | What it is: A management system standard for AI | Can you be certified?: Yes, through a certification body accredited for it, such as one accredited by NABCB, ANAB or UKAS | Where it fits: An auditable wrapper around your AI processes |
| SOC 2 | What it is: An attestation report by a licensed CPA firm, on controls against the Trust Services Criteria | Can you be certified?: No. It is a CPA opinion, not a certificate | Where it fits: Evidence of your general security controls |
| EU AI Act (Regulation (EU) 2024/1689) | What it is: A law with risk tiers and duties for providers and deployers | Can you be certified?: No. It sets legal duties, not a certificate you hold | Where it fits: A legal duty if you sell AI into the EU |
Many teams use the NIST AI RMF as their risk method and ISO 42001 as the part an auditor can check. Our guides on whether you need ISO 42001 and ISO 42001 versus SOC 2 cover that choice. Only a certification body accredited for ISO/IEC 42001 can issue an accredited certificate. Consultants, Mono included, provide readiness work, not certification.
On the EU AI Act, the AI Omnibus, published in the Official Journal on 24 July 2026, moved deadlines. Annex III high-risk obligations now apply from 2 December 2027, not 2 August 2026, and Annex I high-risk obligations from 2 August 2028. The Omnibus moved deadlines; obligations remain. The Act reaches providers outside the EU where the output of their system is used in the Union, so an Indian SaaS vendor selling an AI feature to EU customers is usually a provider, and the EU customer is the deployer. Whether your feature is high-risk is a question for counsel. Our guide to the EU AI Act for Indian SaaS companies covers that classification.
Do compliance platforms such as Vanta and Drata cover the NIST AI RMF?
Partly, and it depends on the plan. As of 20 November 2026, Vanta's pages say it supports ISO 42001, the EU AI Act and the NIST AI RMF with mapped controls, AI policies and AI impact assessments, and that features vary by plan. Drata's frameworks list includes the NIST AI RMF and ISO 42001, and does not list the EU AI Act. Vanta's pricing page lists plans without prices, and we could not load Drata's pricing page, so we quote no figures.
These tools collect evidence and monitor controls. They do not build the product features the evidence points to, such as per-tenant AI switches, audit logs, guardrails or a kill switch. Those live in your codebase.
What should you tell a buyer who asks if you follow the NIST AI RMF?
Say what is true and offer the records. A safe answer reads like this:
"We align our AI risk management to NIST AI RMF 1.0, and we use the Generative AI Profile (NIST AI 600-1) for our LLM features. Our AI inventory, use-case records, test results and incident procedure map to the Govern, Map, Measure and Manage functions. We can share the mapping under NDA. NIST does not certify against the framework, so we make no certification claim."
Avoid saying you are "NIST compliant" or "NIST certified", and avoid listing controls you have not built. A buyer who finds one false row will doubt the rest.
In what order should a small SaaS team start?
- Build the inventory. List every model, vendor and AI feature, and name an owner for each (GOVERN 1.6, 2.1).
- Write a short AI policy. Cover product use, staff use and customer data, and get an executive to approve it (GOVERN 1.2, 2.3).
- Write a use-case record for each customer-facing feature. Start with the one buyers ask about most (MAP 1.1 to 5.1).
- Set up tests and monitoring. Version a test set, run it on every model or prompt change, and watch production (MEASURE 2.1, 2.4).
- Add the controls. Feature flags, pinned model versions and tenant-level settings (MANAGE 2.4, 3.1).
- Write the incident and rollback runbook, and rehearse it once (MANAGE 2.3, 4.3).
- Publish the mapping. One page linking each function to the record that proves it.
If you want this built against your real product, with the evidence ready for buyer reviews, our AI governance work covers the policies, records and product controls. To see where you stand first, take the AI governance self-check. We are not a certification body, and legal wording goes to your counsel.
Sources: NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023 (doi.org/10.6028/NIST.AI.100-1). NIST AI Resource Center, AI RMF Core page. NIST AI 600-1, Generative Artificial Intelligence Profile, 26 July 2024 (doi.org/10.6028/NIST.AI.600-1). Mitchell et al., "Model Cards for Model Reporting", FAT 2019 (arXiv 1810.03993). Shared Assessments, SIG. Cloud Security Alliance, AI Controls Matrix v1.1. AICPA, SOC suite of services. European Commission, AI Act regulatory framework page; Regulation (EU) 2024/1689, Articles 2, 3 and 73. NABCB 220 (April 2026). Vanta and Drata product pages, checked 20 November 2026. Last reviewed 20 November 2026.*
FAQ
Is the NIST AI RMF mandatory for SaaS companies?
No. NIST describes AI RMF 1.0 as voluntary, and no law requires a SaaS company to adopt it. Buyers and questionnaires cite it, though, so aligning to it is a practical way to answer AI risk questions. NIST runs no certification for it, so you can claim alignment but not certification.
What are the four functions of the NIST AI RMF?
Govern, Map, Measure and Manage. Govern covers policies, accountability and culture, and runs through the other three. Map sets the context and finds risks. Measure assesses and tracks them. Manage prioritises, responds and recovers. Together they have 19 categories and 72 subcategories in AI RMF 1.0, and NIST says they are not a checklist.
Can a company be NIST AI RMF certified?
No. NIST does not certify organisations against the AI RMF, and it runs no audit scheme for it. You can say you align to the framework and share a mapping to your records. If a buyer wants an independent certificate, ISO/IEC 42001 is the AI standard that accredited certification bodies can certify against.
What is the NIST Generative AI Profile?
NIST AI 600-1, published on 26 July 2024, is a voluntary cross-sector profile of AI RMF 1.0 for generative AI. It lists 12 risks, including confabulation, data privacy, information security, intellectual property and value chain risks, and suggests actions under the Govern, Map, Measure and Manage functions.
How does the NIST AI RMF relate to ISO 42001?
The NIST AI RMF is a voluntary framework for thinking about and managing AI risk, with no certification. ISO/IEC 42001 is a management system standard that an accredited certification body can audit and certify. Many teams use the NIST AI RMF for their risk method and ISO 42001 for the auditable management system.
How should I answer a buyer who asks if we follow the NIST AI RMF?
State that you align your AI risk management to NIST AI RMF 1.0, name the Generative AI Profile if you use LLMs, and offer a mapping to your inventory, use-case records, test results and incident procedure. Do not say you are NIST compliant or certified, and do not list controls you have not built.
