Key takeaways
- NIST AI RMF 1.0 (January 2023) is voluntary and not certifiable. Its value is a shared vocabulary and a sensible order of work: Govern, Map, Measure, Manage.
- Govern sets roles and policy once. Map, Measure and Manage are done per use case, scaled to how much harm the system can cause.
- For GenAI copilots, use the Generative AI Profile (NIST AI 600-1, July 2024) as a checklist of extra risks such as confabulation and information security.
- A mid-market plant can run a credible first pass in about 90 days with an inventory, a triage worksheet and one owner per use case.
- NIST says the framework is being revised. Build on the four functions, which are stable, rather than on a specific document version.
The NIST AI Risk Management Framework (AI RMF 1.0, published January 26, 2023) is the most widely referenced voluntary method for managing AI risk. It is written for every sector, which makes it abstract. This page translates it for a mid-market plant: what each function means for a vision inspection cell, a predictive maintenance model, a scheduling optimizer and a GenAI copilot, and how much effort it takes.
What the AI RMF is, and what it is not
The AI RMF is a free framework from the US National Institute of Standards and Technology. It has two parts:
- Part 1 describes AI risk and the characteristics of trustworthy AI: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed.
- Part 2 is the Core: four functions (Govern, Map, Measure, Manage), each split into categories and subcategories. NIST also publishes a companion Playbook with suggested actions.
It is not a regulation, not a standard you can be certified against, and not an OT security framework. It will not tell you how to segment a vision PC from the PLC network. For that you still need IEC 62443 and your OT security program.
Note: As of November 2026, NIST's AI RMF page lists version 1.0 as current and states that it is being revised. NIST has also published a concept note (April 2026) for a critical infrastructure profile. The four functions are the durable part. Anchor your program on them, not on subcategory numbers that may change.
Does it apply to you?
Nobody forces a private manufacturer to use it. It applies in practice when:
- A US customer, especially in defense, automotive or medical supply chains, asks in a questionnaire how you manage AI risk.
- Your board or insurer wants a recognized reference for "reasonable" AI governance.
- You want one method that also feeds ISO/IEC 42001 and the EU AI Act, rather than three separate efforts.
If you run any AI that touches quality release, machine behavior, people decisions or customer data, it is worth using.
The four functions in plant language
+-------------------------------+
| GOVERN (once, site-wide) |
| policy, roles, risk appetite |
+---------------+---------------+
|
per use case v
+---------+ +-----------+ +----------+
| MAP | --> | MEASURE | --> | MANAGE |
| context | | test, KPI | | decide, |
| & harm | | & monitor | | respond |
+---------+ +-----------+ +----------+
^ |
+--------- change / drift ---------+
Govern: decide who owns AI risk
Govern is the cross-cutting function. It covers policy, accountability, workforce, culture and third-party risk. In a 400-person company this does not need a new department. It needs:
- A short AI policy that says what is allowed, what needs approval and what is banned (see AI policy for manufacturers).
- Named owners: one accountable executive, and a business owner for each use case. Use a simple RACI.
- A rule that AI changes go through your existing management of change (MOC) process, not around it.
- Supplier expectations for AI vendors and OEMs who ship AI inside machines.
Map: understand the use case before you trust it
Map is about context. For each use case, write down the intended purpose, who relies on the output, what happens when it is wrong, and what data it uses. This is where most manufacturers find their real risks, because the model is rarely the problem. The process around it is.
Measure: test it like you would test a gauge
Measure covers testing, metrics and ongoing monitoring. Think measurement system analysis, not data science theater. Define acceptance criteria before go-live, test on data from your lines, and keep monitoring after SAT.
Manage: act on what you found
Manage is prioritizing and treating risks, planning response and recovery, and deciding when to retire a model. In practice: a go/no-go decision with a named signer, a fallback mode, and a trigger for retraining or rollback.
Applying it to four common manufacturing use cases
| Use case | Map: main harm if wrong | Measure: what to test | Manage: typical control |
|---|---|---|---|
| Vision QC on a line | Escape of defective parts to customer, or false rejects driving scrap | Escape rate and false reject rate against a golden sample set, per SKU and lighting condition | Human review of rejects, periodic challenge parts, rollback to rule-based or manual inspection |
| Predictive maintenance | Missed failure causing downtime or a safety event; alarm fatigue | Lead time and precision on historical failures; alarm volume per shift | Model advises, planner decides; no automatic trips; review after every missed failure |
| Scheduling / sequencing optimizer | Infeasible plans, overtime spikes, unfair task allocation to individuals | Plan adherence, constraint violations, distribution of shifts and tasks | Planner override with reason codes; works council or HR review if it allocates work to named people |
| GenAI copilot (maintenance, SOPs, quotes) | Confident wrong answers (confabulation), leaked IP or recipes | Answer accuracy on a test set of real questions; data leakage checks | Approved tools only, retrieval from controlled documents, human sign-off for anything safety-related |
Plant-floor reality: The vision model passed FAT at the integrator with perfect lighting. Three months later a skylight, a new supplier's resin color and a dirty lens moved the false reject rate by a factor of three. Measure is not a one-time test. Plan a monitoring check at every changeover family and every material change.
A note on scheduling and worker-facing AI
A scheduler that only sequences orders is low risk. One that assigns tasks or shifts to named individuals based on their behavior or performance is different. In the EU, that kind of system can fall into the employment and worker management area of the AI Act's high-risk list. Map should flag it early and pull in HR and worker representatives.
The Generative AI Profile (NIST AI 600-1)
NIST published the Generative AI Profile as NIST AI 600-1 on July 26, 2024. It names twelve risks that are unique to or made worse by generative AI, and suggests actions under each AI RMF function. The ones that matter most on a plant floor:
- Confabulation: the copilot invents a torque value, a lockout step or a part number. Treat any safety-related answer as unverified until checked against the controlled document.
- Information security: prompt injection through documents, and copilots that can reach systems they should not. Keep copilots out of OT networks and give them read-only access to curated content.
- Data privacy and intellectual property: recipes, process parameters, customer drawings and employee data pasted into public tools.
- Value chain and component integration: you rarely control the underlying model. Your vendor's model update can change behavior overnight.
- Human-AI configuration: over-reliance by new technicians who trust the copilot more than the senior tech.
Use the profile as a checklist when you approve each GenAI tool, and pair it with a clear GenAI shop-floor policy.
What good looks like at a mid-market plant
You do not need every subcategory. A proportionate, defensible setup has:
- An AI use case register listing every AI system, including AI inside OEM machines and SaaS tools, with owner, purpose, data and risk tier.
- A triage rule that sorts use cases into low, medium and high. Low gets a one-page record. High gets full Map, Measure and Manage.
- Acceptance criteria written before go-live, signed by the business owner and quality.
- Monitoring tied to existing routines: daily quality meetings, maintenance KPIs, monthly management review.
- Incident handling that treats a bad AI output like any other nonconformance, with root cause and corrective action.
- Supplier terms that require AI vendors to notify model changes and share test evidence.
Effort and cost
| Scope | Typical effort | Who |
|---|---|---|
| Govern basics (policy, roles, register) | 10 to 20 person-days | IT/OT lead, quality, HR, one executive |
| Map + Measure + Manage for a low-risk use case | 1 to 2 days | Use case owner |
| Same for a high-impact use case (quality release, safety-adjacent) | 5 to 15 days plus testing | Owner, quality, controls engineer, vendor |
| Ongoing | A standing agenda item in monthly management review | AI owner group |
The framework is free. The cost is people time, and most of it lands on the few high-impact use cases.
Common mistakes
- Treating it as an IT project. The risks sit in quality release, maintenance decisions and people decisions. Owners must come from operations.
- Only covering what you built. AI inside a new palletizer, a vision system from an integrator or a SaaS MES feature is still your risk to Map.
- Measuring once. Drift on a line is normal. No monitoring means no Measure function.
- Copying all subcategories into a spreadsheet. Hundreds of rows nobody maintains is worse than twenty that are real.
- Ignoring security. The AI RMF does not replace OT security. A vision PC with remote access open to the vendor is a conduit. Treat it under your AI in OT networks controls.
Monday-morning checklist
- Name one accountable executive and one owner per known AI use case.
- Start the register. Ask maintenance, quality, planning and purchasing what tools they already use, including free GenAI tools.
- Agree a three-level triage rule and apply it to the register.
- Pick the highest-risk use case and run Map on it this month: purpose, users, failure modes, data.
- Write acceptance criteria and a fallback mode for that use case.
- Add an "AI" line to the MOC form and to monthly management review.
- Print the twelve GenAI risks from NIST AI 600-1 and use them when you approve the next copilot.
For a sequenced version of this, follow the 90-day AI governance plan. If you also need a certificate or EU compliance, see how the pieces fit in the framework crosswalk.
Sources and further reading
- NIST AI Risk Management Framework (NIST)
- NIST AI 600-1: AI RMF Generative Artificial Intelligence Profile (NIST)
This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.
FAQ
Frequently asked questions
Is the NIST AI RMF mandatory for US manufacturers?
No. The NIST AI RMF is voluntary guidance. No federal law requires a private manufacturer to adopt it. It does show up in customer questionnaires, procurement terms and some state-level AI discussions as a reference for reasonable practice, so aligning with it gives you a defensible answer when a customer, insurer or auditor asks how you manage AI risk.
Can we get certified against the NIST AI RMF?
No. NIST does not certify organizations and there is no AI RMF certificate. If you need a third-party certificate, ISO/IEC 42001 is the certifiable AI management system standard. Many manufacturers use the AI RMF to structure the risk work and ISO/IEC 42001 to package it into an auditable management system.
How does the NIST AI RMF relate to the EU AI Act?
They are complementary. The EU AI Act is law with legal duties for providers and deployers. The AI RMF is a voluntary method. Running Map and Measure well produces much of the evidence the AI Act expects, such as intended purpose, risk analysis, testing and human oversight records, but it does not by itself demonstrate legal compliance.
What is NIST AI 600-1?
NIST AI 600-1 is the Generative AI Profile of the AI RMF, published in July 2024. It lists risks that are unique to or made worse by generative AI, such as confabulation, information security and value chain risks, and suggests actions mapped to the Govern, Map, Measure and Manage functions. Use it when you deploy copilots or chat assistants.
Related guides
Frameworks
ISO/IEC 42001 for manufacturers: an AI management system that fits your existing ISO stack
ISO/IEC 42001:2023 is the certifiable management system standard for AI.
Guide · 7 min read
Frameworks
EU AI Act for manufacturers: roles, risk tiers and the 2026 timeline
Most manufacturers are deployers under the EU AI Act, with a few hard prohibitions and AI literacy duties already in force.
Guide · 8 min read
Frameworks
AI and OT security framework crosswalk for manufacturers
No single framework covers AI governance and OT security.
Guide · 8 min read
AI Governance
AI risk assessment for manufacturers: a practical tiering scheme
Tier each industrial AI use case by what it can do to the process, people and product: does it advise, act in closed loop or touch safety; does it use worker data; does it release product…
Guide · 7 min read