NIST AI RMF for a small manufacturer or distributor

NIST’s AI Risk Management Framework is voluntary and free, and a small company can start with a one-page checklist.

Derik Lawlis Founder, ThriveAI · builds AI systems inside manufacturers

What the NIST AI RMF is

The NIST AI Risk Management Framework, or AI RMF, is a voluntary guide to managing the risks of AI systems. NIST, the US National Institute of Standards and Technology, released AI RMF 1.0 on January 26, 2023, as publication NIST AI 100-1.

Congress directed NIST to develop a voluntary framework, and the NIST FAQ says organizations are not required to use the framework. The framework document is written for organizations that design, develop, deploy, or use AI systems, of all sizes and in all sectors. That includes a plant whose only AI is a chat assistant on the office computers.

The framework is a free 48-page PDF in two parts. Part 1 explains how to frame AI risk and what makes an AI system trustworthy. Part 2 holds the core, which breaks four functions into 19 categories and 72 subcategories, and a short section on profiles.

NIST says AI RMF 1.0 is being revised as part of the White House AI Action Plan. On October 22, 2026, version 1.0 was still the edition that NIST offered for download.

What NIST means by trustworthy AI

The framework names seven characteristics of trustworthy AI. In the framework’s figure, valid and reliable sits at the base as a necessary condition of trustworthiness. Accountable and transparent relates to all the other characteristics.

• Valid and reliable. The output is right for the intended use and stays right over time, such as a drafted price that matches the cost records.
• Safe. The system does not endanger people, property, or the environment, such as a drafted work instruction that keeps every lockout step.
• Secure and resilient. The system resists attacks and unauthorized use, and keeps working or fails safely when something unexpected happens.
• Accountable and transparent. People can find out what the system did and who answers for each decision.
• Explainable and interpretable. A person can see how the system reached an output and what the output means for the job.
• Privacy-enhanced. The system protects personal information, such as employee files and the names of customer contacts.
• Fair, with harmful bias managed. The system does not treat people unfairly, such as a screening tool that ranks job applicants.

NIST adds that the characteristics involve tradeoffs, and that a single setting rarely calls for all seven. A quoting assistant raises questions of validity and privacy, and fairness moves to the front once AI touches decisions about people.

The four functions in plain words

The core sorts AI risk work into four functions: Govern, Map, Measure, and Manage. NIST calls Govern a cross-cutting function that runs through the other three. Once Govern is in place, NIST expects most users to start with Map and then move to Measure or Manage.

NIST also says the actions in the core “do not constitute a checklist.” The framework lets a small company pick the categories that fit the company’s resources instead of applying every category.

Govern: who decides, and under which rules

Govern sets the policies, the roles, and the culture for AI risk across the company. In NIST’s Govern table, each category reads as an outcome that holds once the work is done.

• Govern 1. “Policies, processes, procedures, and practices across the organization related to the mapping, measuring, and managing of AI risks are in place, transparent, and implemented effectively.”
• Govern 2. “Accountability structures are in place so that the appropriate teams and individuals are empowered, responsible, and trained for mapping, measuring, and managing AI risks.”
• Govern 3. “Workforce diversity, equity, inclusion, and accessibility processes are prioritized in the mapping, measuring, and managing of AI risks throughout the lifecycle.”
• Govern 4. “Organizational teams are committed to a culture that considers and communicates AI risk.”
• Govern 5. “Processes are in place for robust engagement with relevant AI actors.”
• Govern 6. “Policies and procedures are in place to address AI risks and benefits arising from third-party software and data and other supply chain issues.”

At a company with no compliance team, Govern starts with one named owner and a list of every AI tool in use. Govern 6 carries extra weight for a company that buys AI, because part of the risk sits with the vendor. The AI governance setup for a smaller company shows the register and the approval step in detail. AI security for a smaller company covers what each vendor keeps.

Map: where AI is used and what can go wrong

Map sets the context for each AI system: the purpose, the users, the data, and the possible harms. NIST’s Map section says the context gathered in Map should support an initial go or no-go decision on the system.

• Map 1. “Context is established and understood.”
• Map 2. “Categorization of the AI system is performed.”
• Map 3. “AI capabilities, targeted usage, goals, and expected benefits and costs compared with appropriate benchmarks are understood.”
• Map 4. “Risks and benefits are mapped for all components of the AI system including third-party software and data.”
• Map 5. “Impacts to individuals, groups, communities, organizations, and society are characterized.”

For a small company, Map means one row per AI use in the register. The row records the job, the data that goes in, the person who acts on the result, and the cost of a wrong answer.

Measure: test before go-live, then keep checking

Measure tests and tracks the risks found in Map. NIST’s Measure section says AI systems should be tested before deployment and regularly while in operation.

• Measure 1. “Appropriate methods and metrics are identified and applied.”
• Measure 2. “AI systems are evaluated for trustworthy characteristics.”
• Measure 3. “Mechanisms for tracking identified AI risks over time are in place.”
• Measure 4. “Feedback about efficacy of measurement is gathered and assessed.”

Measure 2 holds 13 subcategories, more than any other category. Some of them, such as Measure 2.12 on the environmental impact of model training, concern teams that train models more than companies that buy AI. For a company that buys AI, Measure means testing a tool on past work before go-live, then checking samples and counting reported errors.

Manage: act on the risks, and switch off what fails

Manage puts resources on the risks that Map and Measure found, and plans the response to incidents, according to NIST’s Manage section.

• Manage 1. “AI risks based on assessments and other analytical output from the MAP and MEASURE functions are prioritized, responded to, and managed.”
• Manage 2. “Strategies to maximize AI benefits and minimize negative impacts are planned, prepared, implemented, documented, and informed by input from relevant AI actors.”
• Manage 3. “AI risks and benefits from third-party entities are managed.”
• Manage 4. “Risk treatments, including response and recovery, and communication plans for the identified and measured AI risks are documented and monitored regularly.”

Manage 1.3 lists the options for responding to a risk: mitigate, transfer, avoid, or accept. Manage 2.4 asks for a way to supersede, disengage, or deactivate an AI system whose results stray from the intended use. At a small plant, the simplest version is a written manual process that staff can return to the same day.

A one-page NIST AI RMF checklist for a small manufacturer

This checklist is ThriveAI’s selection from the framework for a manufacturer or distributor with no compliance team. NIST says neither the core nor the Playbook is a checklist. Each item ends with the NIST subcategory behind the item, so you can check the source.

Govern

1. Name one person who owns AI risk and approves every AI tool before staff use the tool (Govern 2.1, 2.3).
2. List every AI tool staff use, including personal accounts and AI features inside software you already pay for (Govern 1.6).
3. Write a one-page AI policy that names the approved tools and the data each tool may see (Govern 1.2, 1.4).
4. Read each vendor’s terms on training, retention, and where data is stored, and keep a copy with the approval (Govern 6.1).
5. Train staff on the policy before they get an account, and again when a tool changes (Govern 2.2).

Map

6. For each AI use, record the job, the data that goes in, and the person who acts on the result (Map 1.1).
7. Rate each use by what a wrong answer would cost. A wrong draft email costs little, and a wrong price, tolerance, or hiring decision costs much more (Map 1.5, 5.1).
8. Check which customer contracts, confidentiality agreements, export rules, and privacy laws limit the data each tool may see (Govern 1.1, Map 4.1).
9. Decide which results a person must approve before anything is sent or saved, and name that person (Map 3.5).

Measure

10. Before go-live, run the tool on past work with known answers, and record how many results were right (Measure 2.3, 2.5).
11. Keep a log that shows what each tool read and drafted, and who approved each draft (Measure 2.4).
12. Each month, check a sample of approved results against the source records (Measure 1.2, 3.1).
13. Give staff one place to report a wrong answer, and count the reports for each tool (Measure 3.3).

Manage

14. Make a go or no-go decision for each use, and write down the risks the company accepts (Manage 1.1, 1.4).
15. Keep the manual process written down, so staff can switch a tool off and keep working (Manage 2.4).
16. Write a one-page incident plan for data that goes into the wrong tool, or a wrong answer that reaches a customer (Manage 4.3).
17. Review each tool on a set date, and again when the vendor changes the terms or the model behind the tool (Govern 1.5, Manage 3.1).

An illustrative example: a quote assistant on the ERP

This example is illustrative and describes no client. A small machine shop wants an AI assistant for quotes. The assistant reads an emailed request for quote, finds similar past jobs in the ERP, and drafts a price for an estimator.

Govern. The operations manager owns the assistant and adds a row to the AI register. The policy lets the assistant read past quotes, job costs, and drawings, and keeps employee files out. The operations manager reads the vendor’s terms on training and retention before anyone signs.

Map. The register row says the assistant drafts prices for repeat and similar parts, and an estimator acts on each draft. A wrong price can lose money on a job or lose the order, so the shop rates the use as high risk. Two customer contracts forbid sending drawings to outside services, so those customers’ drawings stay out of the assistant.

Measure. Before go-live, the estimators run the assistant on last quarter’s quotes and compare each draft with the price the shop sent. The estimators agree on a pass mark in advance and record the result. After go-live, an estimator flags each wrong line, and the operations manager reviews the flags every month.

Manage. The shop goes live on repeat parts only and keeps new parts manual. Every price still needs an estimator’s approval before the quote leaves. If the monthly flags rise, the operations manager switches the assistant off, and the estimators quote by hand as before.

The records for this example fit on one page: the register row, the test result, the monthly flag count, and the switch-off rule.

The Playbook and the Generative AI Profile

The AI RMF Playbook

The Playbook is NIST’s online companion to the framework. For each of the 72 subcategories, the Playbook suggests actions and lists questions an organization can answer to document the work.

The Playbook page says the Playbook is not a checklist and not a set of steps to follow from start to finish. According to the NIST FAQ, an organization can borrow as many or as few suggestions as apply to the organization’s use case.

NIST publishes the Playbook on the AI Resource Center as web pages and as PDF, CSV, Excel, and JSON files. The PDF runs to 147 pages, and NIST plans to update the Playbook after the revision of AI RMF 1.0.

A small company can use the Playbook one entry at a time. When you work on a row of the checklist, open the entry for the subcategory named at the end of the row. The Govern 1.6 entry, for example, suggests naming the person or team that maintains the AI inventory.

The Generative AI Profile

NIST released the Generative Artificial Intelligence Profile, publication NIST AI 600-1, on July 26, 2024. A profile applies the framework’s functions, categories, and subcategories to one setting or technology, in this case generative AI. The profile defines 12 risks that generative AI creates or makes worse.

• CBRN Information or Capabilities
• Confabulation
• Dangerous, Violent, or Hateful Content
• Data Privacy
• Environmental Impacts
• Harmful Bias or Homogenization
• Human-AI Configuration
• Information Integrity
• Information Security
• Intellectual Property
• Obscene, Degrading, and/or Abusive Content
• Value Chain and Component Integration

Confabulation is the profile’s word for confidently stated but false content, such as a wrong tolerance in a drafted quote. The intellectual property risk includes the “eased exposure of trade secrets,” a concern for any shop that pastes customer drawings into a chat tool. Human-AI configuration includes people who rely on the output too much, which the profile calls automation bias and over-reliance.

The profile lists more than 200 suggested actions, each tied to a subcategory. NIST notes that actions written for developers of generative AI may not apply to deployers. A company that only uses a chat assistant does not develop generative AI, so many developer actions do not apply.

Read the profile once staff use ChatGPT, Claude, or Copilot, or when a vendor’s product runs on a large language model. Start with GV-1.6-001, which adds generative AI systems to the AI inventory. GV-6.1-007 suggests an inventory of every third party with access to company content, and a list of approved generative AI providers. GV-1.6-002 suggests stating in policy any inventory exemptions for generative AI built into other software.

Is there a NIST AI RMF certification?

No. The framework is voluntary. The AI Resource Center says NIST “does not validate or endorse any individual organization or its approach to using the AI RMF.” On October 22, 2026, NIST’s AI RMF pages described no certification program.

A certificate that carries the framework’s name comes from the course or company that issued the certificate, not from NIST.

ISO/IEC 42001, published in December 2023, is a separate international standard for an AI management system. A company that wants a certificate for ISO/IEC 42001 is audited by a certification body. ANAB accredits certification bodies for ISO/IEC 42001, and the Standards Council of Canada does so in Canada.

The AI Resource Center lists a crosswalk from Microsoft, a table that matches AI RMF subcategories to clauses in a final draft of ISO/IEC 42001. NIST notes that a listing does not imply that either document fully covers the contents of the other.

When a customer, an insurer, or an auditor asks about AI risk, answer with records. Show the owner’s name, the AI register, the approval step, and the test results. Label each record with the matching NIST category. NIST designed the framework to give organizations a common language for AI risk, including with customers.

How ThriveAI helps

ThriveAI builds private AI systems on a company’s own data for manufacturers and distributors in Canada and the United States. Working sessions run by video call or on site.

The platform ThriveAI builds on is designed to keep each client’s data on a server dedicated to that client, in Canada or the United States. The client chooses where the model runs. One option is an open-weight model, a model whose files are published, running on that same server. The other is a frontier model under a written zero-data-retention control, with the model tier named. Under that control, the provider keeps no copy of prompts or answers beyond the exceptions the provider lists. A hosted model may process requests in another country.

A person approves every action before anything is sent or saved. That approval step is often called human in the loop. Map 3.5 asks a company to define and document a process for human oversight.

For the Govern and Map rows, start with the AI accounts staff already use. The free AI security review covers ChatGPT, Claude, or Copilot: the plan each person is on, training and retention, connectors and apps, and sharing. Derik works by email or a video call, needs no access and never a password, and sends a written list of fixes.

ThriveAI holds no audited certification, including ISO/IEC 42001, and does not certify companies against the AI RMF. This page is general information and is not legal advice.