NIST AI RMF and ISO/IEC 42001
- Lesson 6
- 30 min
- Module 2: Laws and standards in plain words
Where we are
In the last lesson, we read the EU AI Act: it says what must be true for some systems, and from when. But it does not say how to run governance every week, for every system. In this lesson, we see the NIST framework and the ISO/IEC standards that organisations use for that, starting in a restaurant kitchen.
Lesson 6 of 19 · AI Governance: Policies, Risk and AccountabilityNIST AI RMF and ISO/IEC 42001
1. What governance is for
2. Laws and standards in plain words
3. Inventory, risk tiers and impact assessment
4. Policies people can follow
5. Controls, documentation and evidence
What you will be able to do
A law says what must be true; it does not say how to keep it true every week. You learn the NIST framework, the ISO/IEC 42001 management system and the ISO/IEC 42005 impact assessment. Then you see how the three fit together.
Before you startWhat to know and to have ready
What you need to know
- The six kinds of rule, from A map of the landscape. A framework is voluntary. A standard binds you only if you choose it or a contract asks for it.
- The categories and roles, from The EU AI Act.
What you need
Sources: the NIST AI Risk Management Framework (official page, checked 22 October 2026), the NIST Generative AI Profile, NIST AI 600-1 (checked 22 October 2026), the NIST crosswalks (checked 22 October 2026), and ISO’s pages for ISO/IEC 42001 and ISO/IEC 42005 (checked 22 October 2026). The ISO/IEC standards are sold by ISO. This course explains them from public information and its own templates, and does not copy their text.
A restaurant kitchen
Section titled “A restaurant kitchen”A good restaurant kitchen uses three kinds of help to keep food safe.
A food safety guide lists what to think about: storing food cold, washing hands, allergies, cleaning. It is good advice, and the kitchen chooses to use it. The kitchen also has a routine. Someone owns it. The fridge temperature is written down twice a day. The chef reviews the records every month. An inspector can come, read the records and confirm that the routine works. And before a new dish goes on the menu, the chef does a check for that dish. Who might be allergic? What could go wrong? What must the waiter say?
The guide says what matters. The routine makes it happen every day, in a way that someone else can check. The check for one dish looks at the people one change affects.
The problem: reviews only after something breaks
Section titled “The problem: reviews only after something breaks”Farida reads the “last review” line on Larkfield’s eight system cards (constructed for this course). The triage service had a review once: after an incident, in month 2. The support-resolution workflow and the support MCP server: never. The document assistant: at launch, and not since.
Each team did good work. But nobody owned a routine, so a review happened only when something broke. The EU AI Act gives some dates and duties. It does not tell Larkfield who checks the triage service each quarter, or what the check looks at. Larkfield needs the kitchen’s three kinds of help.
The idea: a guide, a routine and a check
Section titled “The idea: a guide, a routine and a check”The guide: the NIST AI Risk Management Framework
Section titled “The guide: the NIST AI Risk Management Framework”The NIST AI Risk Management Framework (AI RMF 1.0) comes from the US National Institute of Standards and Technology. NIST released it in January 2023, and it is voluntary. NIST says that it is being revised; the course gives no date for the revision.
Its core has four functions. GOVERN runs across the other three. In the course’s plain words:
Text description of the diagram
The diagram shows a large frame, GOVERN, which runs across everything: the rules, the roles and the culture. Inside the frame are three boxes in a loop. MAP: know the system, its context, who it affects and its risks. MEASURE: test and track those risks. MANAGE: act on them, by priority, and respond when something fails. An arrow labelled “new facts” goes from MANAGE back to MAP, so the loop starts again.
- Govern: who owns what, which rules apply, and how people learn them.
- Map: what each system does, who it affects, and what could go wrong.
- Measure: test those risks and keep tracking them.
- Manage: decide what to do about each risk, and respond when something fails.
For generative AI, NIST adds a companion, the Generative AI Profile (NIST AI 600-1, July 2024). It lists 12 risks, for example confabulation (a confident answer that is wrong), data privacy, information integrity and harmful bias. Use it as a checklist when you map a system that writes text.
The routine: ISO/IEC 42001
Section titled “The routine: ISO/IEC 42001”A management system is an organisation’s routine for one topic. It has policies, roles, regular reviews and records that show the routine runs. ISO/IEC 42001:2023 sets requirements for an AI management system. ISO published it in December 2023. Because it sets requirements, organisations can be audited against it by an outside certification body, and certified, like the kitchen inspector. What a certificate covers, and which body may give it, is a question for the certification body.
The check for one change: ISO/IEC 42005
Section titled “The check for one change: ISO/IEC 42005”An impact assessment asks, for one AI system, who could be affected, how, and what to do about it. ISO/IEC 42005:2025 gives guidance on how and when to assess the effects of an AI system on people and society. ISO published it in May 2025. As the course reads it, it is guidance, not a list of requirements to certify against. This course’s impact-assessment template follows its public outline in its own words.
A fourth ISO/IEC standard, ISO/IEC 23894:2023, gives guidance on AI risk management. The map lists it, but this course does not use it further.
How the three fit together
Section titled “How the three fit together”The three documents do different jobs, and they connect.
Text description of the diagram
The diagram shows three documents. The NIST AI RMF is a voluntary framework: it says what to think about, in four functions. ISO/IEC 42001 is a management system standard. It says how an organisation runs AI governance every time, so that an auditor can check it. ISO/IEC 42005 is a guidance standard: a check of one system’s effects on people. A solid arrow goes from NIST to 42005, because NIST lists an official crosswalk between them. A dotted two-way arrow links NIST and 42001: in the course’s view, they cover the same jobs from two sides. A dotted arrow goes from 42001 to 42005: in the course’s view, an impact assessment is one of the processes that a management system runs.
In this course, the templates follow the same jobs. The responsibility table, the policy, the board, the inventory and the risk tiers do the NIST jobs. The impact assessment follows ISO/IEC 42005’s public outline. The controls, the evidence and the incident records are the kind of routine that a management system keeps.
Only one link is official. NIST’s AI Resource Center lists a crosswalk from ISO/IEC 42005 to the AI RMF. NIST says that a crosswalk is not an endorsement. The other links are the course’s view. Use them as a way to remember the jobs, not as a mapping that an auditor will accept.
Sam asks the obvious question: “Do we need all three?” No. A small company can use the NIST functions as its guide, and the templates of this course as its routine, with no certificate. A company whose customers ask for ISO/IEC 42001 certification needs the routine in that standard’s form, checked by an auditor.
A tiny example: the document assistant, function by function
Section titled “A tiny example: the document assistant, function by function”Farida walks the document assistant through the four functions. She writes one line of output for each:
| Function | Question | Output for the document assistant |
|---|---|---|
| Govern | Who owns it? Which rules apply? | Mei is the accountable owner. The rules row from the last two lessons: PIPEDA, two contracts, minimal risk under the EU AI Act (most likely). |
| Map | Who does it affect, and what could go wrong? | Staff, and customers through staff’s answers. The partner shop’s staff too. Risks: a wrong answer; a partner sees Larkfield’s staff-only documents. |
| Measure | What do we know about those risks? | Recorded in the capstone course: 87.8% of 41 held-out questions correct; 430 tests that tried to read the other company’s documents (Larkfield’s or the partner shop’s), 0 leaks. |
| Manage | What do we do about them? | Last review: at launch. Set a review every six months, and a rule: a leak stops the partner’s access the same day. |
The recorded numbers come from an earlier course’s committed run. They show what evidence looks like; they do not say how good the assistant is today. Each line also uses only one function. The table is short because the system is simple and minimal risk. The triage service, which routes every ticket alone, would need more lines.
Discuss: “Let’s get certified this quarter”
Section titled “Discuss: “Let’s get certified this quarter””Mei’s large customer asked about ISO/IEC 42001.
- Omar: “Then we get certified this quarter. A certificate ends the discussion.”
- Kwame: “We do not have an inventory yet. The registry lists four of our eight systems.”
- Tomás: “Certification is the legal team’s job.”
Before you read on, answer these:
- What does the customer probably want: a certificate, or evidence?
- What must exist before an auditor can check a management system?
- Whose job is the routine?
The course’s answer. First ask the customer what it needs. Often a customer wants evidence: an owner per system, a risk process and records. A certificate is one way to show that, and it takes the longest. A management system can only be audited once it runs: an inventory, owners, reviews on a calendar and records of them. Larkfield has half an inventory and reviews that happen after incidents. So the honest plan is the routine first, with this course’s templates, and certification later, if a contract asks for it. And the routine is not the legal team’s job. The owners run it. The legal team answers the questions about the law.
What would change the answer: a customer contract that requires a certificate by a date. Then the standard binds Larkfield through the contract, and the plan needs a certification body and a budget.
When to use them, and what they cost
Section titled “When to use them, and what they cost”Use the NIST functions as a checklist when you set up governance, and when you review one system. They cost nothing to read. Use the Generative AI Profile when you map a system that writes text. Use an ISO/IEC 42005-style impact assessment for systems with a higher risk tier, not for every system. Module 3 sets the tiers, in Risk tiers and proportionate controls. Use ISO/IEC 42001 when you want a routine that an outside body can certify, or when customers ask for it.
The standards cost money to buy, and certification costs audit fees and staff time each year. Prices change: check the current page. The larger cost is the routine itself: owners who review, and records that show it. That cost is the same with or without a certificate.
Common mistake
A policy that copies the framework's words
Symptom: a 40-page policy has a chapter for each NIST function. No chapter names a person or a date. The triage service still has no review on the calendar.
Cause: the team copied the framework’s structure, but not its purpose. A function is a job that someone does, not a heading.
Fix: for each function, write who does it, for which systems and when, and where the record is kept. Then test it on one system, as Farida did with the document assistant.
Knowledge check
Section titled “Knowledge check”Check your understanding
Match each document to its job: (a) the NIST AI RMF; (b) ISO/IEC 42001; (c) ISO/IEC 42005. Jobs: (1) how the organisation runs AI governance every time, in a form an outside body can audit; (2) what to think about, in four functions; (3) how to assess one system's effects on people and society.
Show the answer
(a) and (2): the NIST AI RMF is a voluntary framework with four functions, govern, map, measure and manage. (b) and (1): ISO/IEC 42001 sets requirements for an AI management system, which an outside body can audit. (c) and (3): ISO/IEC 42005 gives guidance on AI system impact assessments.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
The triage service had one review, after an incident. Which NIST function failed, and what would you add?
Show the answer
GOVERN failed first: nobody owned a routine of reviews, so MANAGE happened only after something broke. Add a named owner for the review and a date on the calendar, for example each quarter. Write what the review looks at (the MAP and MEASURE outputs) and where the record is kept.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
Omar says: 'An ISO/IEC 42001 certificate would prove that we follow the EU AI Act.' Is that right? Explain in two sentences.
Show the answer
No. ISO/IEC 42001 is a standard that Larkfield may choose. A certificate shows that an auditor checked its management system, not that it meets each duty of a law. The EU AI Act is a law, and whether its duties are met is a question about each system, for the lawyer.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Recap
- The NIST AI RMF is a voluntary guide with four functions: govern (across everything), map, measure and manage. Its Generative AI Profile lists 12 risks for systems that write text.
- ISO/IEC 42001 sets requirements for an AI management system: a routine with owners, reviews and records, which an outside body can audit.
- ISO/IEC 42005 gives guidance on impact assessments for one system. NIST lists a crosswalk to it; the course’s other links are its own view.
- A framework or a certificate does not show that a law is met. Start with the routine; certify if a customer or a contract asks.
You can now use a guide, a routine and a check for one system. But Larkfield is a Canadian company, and the rules at home are different again. Next, you look at Canada, the US, the UK and other places, and set up a watch on the rules.