Key takeaways
- Your internal tier and the EU AI Act category are different things. Most plant AI is low or medium internal risk and not high-risk under the AI Act.
- Decision authority (advise, recommend, act) and control interaction (read OT, write OT, safety) drive most of the risk.
- Worker monitoring and productivity analytics deserve a high tier by default because of privacy, works council and AI Act employment rules.
- Use the NIST AI RMF functions as the structure for what you do per tier, not as a scoring model.
- A one-page triage per use case is enough for most cases; reserve deep assessments for the high tier.
Risk assessment is where most AI governance programs either earn their keep or turn into paperwork. In a plant, you have a simple advantage: you already think in terms of hazards, consequences and layers of protection. An AI risk tier should feel like a lightweight HAZOP question, not a 40-page questionnaire.
This page gives you a tiering scheme built for industrial use cases, shows how it relates to the EU AI Act and the NIST AI RMF, and works through six common examples. The aim is to spend real effort only where it matters.
Two different questions
Keep these apart or you will confuse everyone:
- Internal risk tier. How much can this AI hurt our people, process, product, customers or know-how? This drives your controls.
- Regulatory category. Does the EU AI Act (or another law) put specific duties on us for this system? This drives compliance tasks.
A closed-loop optimizer on a furnace can be high internal risk and still not be "high-risk" under the AI Act. A recruiting screening tool can be modest internal risk and squarely high-risk under the Act. Do the internal tier first, then check the regulatory category as a separate step.
The tiering scheme
Score each use case on five factors. The highest factor usually sets the tier.
| Factor | Low | Medium | High |
|---|---|---|---|
| Decision authority | Informs a human who decides | Recommends, human approves each action | Acts automatically, human can override |
| Control interaction | No OT connection, or read-only from a DMZ replica | Reads live OT data; writes to MES or scheduling | Writes setpoints or commands to Level 1/2; any safety function |
| People | No personal data | Personal data incidental (e.g. names in work orders) | Monitors, rates or allocates work to individuals |
| Product and customer | Internal use only | Feeds quality decisions a human signs off | Releases product, or output goes directly to customers |
| Know-how exposure | Public or internal data | Confidential data within your tenant | Restricted data (recipes, CAD, PLC code) leaves your control |
Then apply three rules:
- Any safety function involvement goes to high tier at minimum, and if the AI is part of a safety instrumented function or machine safeguarding, it is excluded unless it passes your functional safety lifecycle.
- Worker monitoring or performance rating goes to high tier by default.
- Two or more medium factors go to high tier.
Plant-floor reality: "A human is in the loop" is only true if the human has the time, information and authority to disagree. An operator who has to accept 200 vision rejects a shift without seeing the image is not a control. Score decision authority on how it really works on nights, not how it reads in the URS.
What each tier requires
| Tier | Assessment | Approval | Controls before go-live | Review |
|---|---|---|---|---|
| Low | Register entry and one-page triage | Line manager | Data rules, user briefing | On change |
| Medium | Triage plus short risk note: failure modes, fallback, data flows | Process owner plus IT/OT | Test against known cases, manual fallback documented, monitoring of key metric | Every 2 years and on change |
| High | Full assessment: failure modes, security review, data protection impact where personal data, legal check | AI governance group, through MOC | Defined operating limits enforced outside the AI, validation plan, cyber review of zones and conduits, training, incident procedure | Annually and on change |
The AI risk triage worksheet puts the factors and rules on one page.
Mapping to the EU AI Act, carefully
As of November 2026, the AI Act categories most relevant to manufacturers are:
- Prohibited practices (Article 5). Applied since 2 February 2025. The one that bites in plants is emotion recognition in the workplace, which is banned except for medical or safety reasons. Fatigue or attention monitoring products need careful legal review on this point.
- High-risk, Annex III. Stand-alone uses in listed areas. For manufacturers the main one is employment and worker management: AI used for recruitment, decisions on promotion or termination, allocating tasks based on individual behavior or traits, or monitoring and evaluating performance. Following the Digital Omnibus on AI (Regulation (EU) 2026/1744), these obligations apply from 2 December 2027.
- High-risk, Annex I. AI that is, or is a safety component of, a product covered by listed EU product legislation. The Omnibus moved machinery to a sectoral approach under the Machinery Regulation, with obligations from 2 August 2028. If you build machines, or substantially modify AI in machines you operate, read our page on the AI Act and the Machinery Regulation.
- Transparency (Article 50). For example, people interacting with an AI chatbot should be told so. These obligations apply from 2 August 2026.
- Everything else. Most plant AI: vision inspection, PdM, energy optimization, scheduling. No specific AI Act duties beyond AI literacy, though product safety, GDPR and contract law still apply.
Watch out: You can become a provider rather than a deployer. If you put your name on an AI system, substantially modify a high-risk system, or change its intended purpose so it becomes high-risk, provider duties can shift to you (Article 25). Integrators who retrain vendor models for your line should read this too. See the EU AI Act guide and run the AI Act risk triage assessment if in doubt.
Mapping to the NIST AI RMF
The NIST AI RMF is voluntary and does not give you tiers. Use its four functions as the checklist of what to do for each use case, scaled by tier:
- Govern. Policy, roles, inventory. Same for all tiers.
- Map. Context and intended use, who is affected, what could go wrong. One paragraph for low, full workshop for high.
- Measure. Testing and monitoring metrics: false reject and false accept rates, drift, uptime. Proportionate to tier.
- Manage. Decide, treat, monitor, respond. Fallback procedures and incident handling.
For GenAI use cases, NIST's Generative AI Profile (NIST AI 600-1) lists risks such as confabulation and information security that map well onto shop-floor concerns. More on the framework at NIST AI RMF.
Worked examples
These are typical configurations. Your tier depends on how you actually deploy.
| Use case | Decision authority | Control interaction | People | Product / customer | Internal tier | Likely AI Act position (Nov 2026) |
|---|---|---|---|---|---|---|
| Vision defect detection, operator reviews rejects | Recommends | Reads camera; diverts parts via PLC | None | Feeds quality decision | Medium | Not high-risk in typical use; AI literacy applies |
| Vision with automatic reject and no human review, safety-critical part | Acts | Writes to PLC reject gate | None | Effectively releases product | High | Usually not high-risk unless a product safety component; check |
| Predictive maintenance alerts in CMMS | Informs | Reads historian via DMZ | Incidental | Internal | Low to medium | Not high-risk in typical use |
| Energy optimization writing setpoints to HVAC, compressors or furnaces | Acts | Writes to Level 2 | None | Indirect | High (medium if advisory only) | Usually not high-risk; machinery and safety review if setpoints affect safeguarding |
| AI scheduling in MES, planner approves | Recommends | Writes to MES | Can affect shift allocation | Delivery dates | Medium | Not high-risk unless it allocates work to individuals based on behavior or traits |
| GenAI SOP and troubleshooting assistant | Informs | None | Incidental | Internal; risk if SOPs released unchecked | Medium (restricted data and safety content) | Not high-risk; Article 50 transparency if it chats with people |
| Worker productivity analytics (cameras, scanners, wearables) | Informs or acts on people | Reads MES and badge data | Monitors individuals | Internal | High | Likely Annex III employment (from 2 Dec 2027); possibly prohibited if it infers emotions |
Notes on the less obvious rows:
- Vision with auto reject. The gate is "fail safe" for the customer only if the model's false-accept rate is well understood. Validate it like any other inspection method under your quality system, with a golden-sample set and periodic challenge parts.
- Energy optimization. The savings case is attractive, which is why people skip MOC. Enforce hard limits in the PLC or DCS, not in the AI, and give operators a one-touch revert.
- Worker analytics. Even in the US and UK, where the AI Act does not apply directly, data protection rules, labor law and employee trust make this a board-level decision. In Germany, works councils have co-determination rights over technical systems capable of monitoring employee behavior or performance.
Effort
- Low tier: 15 to 30 minutes per use case.
- Medium tier: half a day, mostly the owner plus one IT/OT reviewer.
- High tier: 3 to 10 person-days across functions, plus legal and works council time where relevant.
A plant with 20 registered use cases will typically have most in low or medium. Budget accordingly.
Common mistakes
- Treating the AI Act category as the risk assessment. Most dangerous plant AI is not "high-risk" in the Act's sense.
- Scoring the vendor's brochure instead of your deployment. Same product, different risk on different lines.
- No re-triage on change. A pilot that moves from advisory to automatic is a new use case.
- Leaving cyber out. An AI edge box with a vendor tunnel is a new conduit. Assess it.
- Over-engineering low-risk cases. Teams stop registering things.
Monday-morning checklist
- Take your AI use-case inventory and add the five factor columns.
- Triage your top ten use cases in one two-hour session with ops, IT, OT and quality.
- Flag anything that writes to control systems, touches safety, or monitors workers for a full assessment.
- Check high-tier items against the AI Act categories above and note open questions for counsel.
- Agree who approves each tier, using the roles and RACI guidance.
This page is general guidance, not legal advice. Confirm regulatory obligations with your counsel.
Sources and further reading
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), EUR-Lex
- NIST AI Risk Management Framework (AI RMF 1.0)
- CISA and partners: Principles for the Secure Integration of AI in Operational Technology (Dec 2025)
This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.
FAQ
Frequently asked questions
Is predictive maintenance high-risk under the EU AI Act?
Usually not. Predictive maintenance that flags likely failures for a planner to act on is not in the AI Act's Annex III list and is rarely a safety component. It can still matter internally if it drives decisions on safety-critical equipment. As of November 2026, check whether the asset falls under product safety law and confirm with counsel if the model affects a safety function.
What makes an industrial AI use case high risk?
In practice: the AI acts without a human checking, it writes to the control system, it touches a safety function, it decides on product quality release, it reaches customers directly, or it monitors or rates workers. Any one of these lifts a use case to at least medium. Two or more, or any safety involvement, should go to the high tier with a full assessment.
Do I need to follow the NIST AI RMF if I am a European manufacturer?
No, it is voluntary. It is still a useful structure because its four functions, Govern, Map, Measure and Manage, map neatly onto what a mid-market company needs to do. Many US customers ask about it in supplier questionnaires. You can use it alongside ISO/IEC 42001 and the EU AI Act without conflict.
How often should AI risk assessments be repeated?
Repeat the triage whenever the use case changes materially: a new model version, new data sources, expanded scope to another line, or a move from advisory to automatic action. Otherwise review high-tier use cases annually and medium-tier ones every two years. Tie re-triage to your MOC process so it happens automatically.
Related guides
AI Governance
How to build an AI use-case inventory in a manufacturing plant
An AI use-case inventory is a register of every AI system you build, buy or have embedded in equipment, with an owner, a risk tier and enough detail to act on. In a…
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. High-risk obligations now apply from 2…
8 min read
Frameworks
NIST AI RMF for manufacturers: Govern, Map, Measure, Manage on the plant floor
The NIST AI Risk Management Framework is a free, voluntary playbook for managing AI risk. For a manufacturer it becomes useful once you map its four functions onto real…
7 min read
Frameworks
EU AI Act and the Machinery Regulation: what machine builders and plants need to know
From 20 January 2027 the Machinery Regulation replaces the Machinery Directive and explicitly covers machine learning in safety functions. Since the July 2026 AI…
7 min read