Key takeaways
- You cannot govern AI you have not found. The inventory is the first real control, and everything else depends on it.
- Most plant AI is bought, not built: look in MES, CMMS, QMS, ERP, vision systems, OEM equipment and Microsoft 365.
- Capture 15 to 20 fields, not 60. Owner, purpose, data, decision type, integration point and risk tier matter most.
- Link the register to your OT asset inventory so an AI system maps to real assets, zones and conduits.
- Keep it alive through MOC, purchasing and an annual attestation, not through a once-a-year spreadsheet hunt.
Ask a plant manager how many AI systems run in the plant and you will usually hear "two pilots". Walk the site with the maintenance lead, the quality engineer and someone from IT and the number is closer to fifteen or twenty. The gap is vendor features nobody switched on deliberately, cameras bought by a line manager on a capex request, and GenAI tools people adopted on their own.
An AI use-case inventory, sometimes called an AI register, closes that gap. It is the first control that actually does something, because every other step (risk tiering, vendor checks, training, monitoring) needs a list to work from. This page covers where AI hides in a plant, what to capture, and how to stop the register going stale six months after you build it.
What it is, and does it apply to you
An AI inventory is a living list of AI systems and the use cases they serve, each with a named owner and a risk tier. It is not a project list. It includes things you did not build and might not have chosen.
It applies to you if you are subject to the EU AI Act as a deployer (most EU manufacturers using any AI), if you plan to align with ISO/IEC 42001 or the NIST AI RMF, or if a customer has started asking about AI in their supplier questionnaires. The AI Act does not prescribe an inventory format for deployers, but working out which obligations apply to which system is impossible without one. The NIST AI RMF's Govern function explicitly calls for mechanisms to inventory AI systems.
Where AI hides in a plant
Search in these places, roughly in order of yield:
| Where | What to look for | Who knows |
|---|---|---|
| MES and scheduling | "Optimized" sequencing, AI scheduling assistants, predicted cycle times | Production planning, MES admin |
| CMMS / EAM | Failure prediction, anomaly scores, AI work-order triage | Maintenance manager |
| QMS and SPC | Automated defect classification, AI root-cause suggestions, CAPA drafting | Quality engineers |
| ERP | Demand forecasting, inventory optimization, AI copilots in finance and purchasing | ERP key users |
| Vision systems | Deep-learning inspection, OCR on labels, presence checks with trained models | Line engineers, integrators |
| Historian and analytics | Anomaly detection, soft sensors, energy optimization add-ons | Process engineers, OT |
| OEM machines | "Adaptive", "smart" or "self-learning" features in CNC, welding, molding, robots | OEM service reports, maintenance |
| Microsoft 365 / Google | Copilot, Gemini, meeting transcription and summaries | IT admin console |
| Shadow GenAI | Public chatbots, browser extensions, coding assistants, AI in personal phones | Proxy logs, anonymous survey |
| Spreadsheets and scripts | Python models in a shared drive, Excel forecasting add-ins | Data-savvy engineers |
| People systems | Recruiting screening, productivity dashboards, camera analytics | HR, security, EHS |
Plant-floor reality: OEM equipment is the hardest category. The AI may run on an edge box inside the machine cabinet, update with firmware, and phone home through the OEM's remote access. If your OT asset inventory already lists that box, you are halfway there. If it does not, you have an OT security gap as well as an AI governance one.
Three practical techniques:
- Interviews. Thirty minutes each with maintenance, quality, production planning, process engineering, IT, HR and purchasing. Ask "what software or equipment do you use that predicts, classifies, recommends or optimizes?" rather than "do you use AI?". People under-report when you use the buzzword.
- Contract and license review. Search vendor release notes and contracts from the last two years for "AI", "machine learning", "predictive", "copilot", "assistant" and "neural".
- Technical scan. Pull SaaS discovery or proxy logs for known GenAI domains, check the Microsoft 365 admin center for Copilot licenses and usage, and ask the OT team to list devices that talk to OEM clouds.
Fields to capture
Resist the 60-column template. A register people abandon is worse than a short one they keep. These fields cover what you need for risk tiering, vendor follow-up and regulatory scoping:
| Field | Why it matters |
|---|---|
| ID and name | Stable reference for MOC and audits |
| Site, area, line or cell | Ties AI to physical location |
| Business owner (a named person) | Accountability |
| Technical owner (IT, OT or vendor) | Who fixes it |
| Purpose, in one sentence | Scope creep is visible later |
| Source: built, bought, embedded in equipment, GenAI tool | Drives vendor and contract questions |
| Vendor, product and version | Change tracking |
| Decision type: informs a human, recommends, acts automatically | Core risk driver |
| Control interaction: none, reads OT data, writes to OT | Core risk driver for safety and security |
| Purdue level and zone where it runs | Links to network segmentation |
| Linked OT assets | Joins to asset inventory |
| Data used, highest data class | IP and privacy exposure |
| Personal data or worker monitoring: yes/no | Privacy, works council, AI Act employment category |
| Affects product quality release or customer: yes/no | Quality and liability |
| Safety-relevant: yes/no | Functional safety boundary |
| Risk tier and date assessed | Output of triage |
| EU AI Act role and category (preliminary) | Deployer or provider, and likely category |
| Status: idea, pilot, production, retired | Lifecycle |
| Last review date | Keeps it honest |
Use our AI use-case register if you want these columns ready-made with drop-downs.
A word on Purdue levels and zones
Adding where the AI runs and what it touches is the field most IT-led registers skip. It is also the one that tells you whether a compromised model or vendor connection could reach a controller.
Level 4/5 ERP, M365 Copilot, cloud AI services
| (IT data, GenAI, forecasting)
Level 3.5 IT/OT DMZ <-- AI data brokers, historian replica
|
Level 3 MES, historian, AI analytics servers
|
Level 2 HMI, SCADA, edge AI boxes, vision PCs
|
Level 1 PLCs, drives, robot controllers
|
Level 0 Sensors, cameras, actuators
An anomaly model reading historian replicas in the DMZ is a very different risk from an edge box on Level 2 writing setpoints to a PLC. Our Purdue model and AI page covers this in more depth.
From inventory to risk tier
Once an entry exists, run a short triage. The two fields that move the tier most are decision type and control interaction. Personal data, quality release and safety relevance push it up further. The AI risk assessment page sets out a scheme and worked examples. For anything bought, send the OT AI vendor security questionnaire to fill the vendor fields you cannot answer yourself.
Keeping it alive
The inventory dies when it depends on someone remembering to update it. Wire it into processes that already run:
- MOC. Add one question to the MOC form: "Does this change add or change AI or machine learning, including vendor features or firmware?" A yes creates or updates a register entry before approval.
- Purchasing. Add an AI tick box to purchase requisitions and RFQs for software and equipment above a threshold. Buyers forward ticked requests to the register owner.
- FAT/SAT. Add "AI features present, enabled, data flows documented" to factory and site acceptance checklists.
- IT onboarding of SaaS. Any new SaaS goes through a quick AI question as part of the security review.
- Vendor release notes. Ask key vendors (MES, CMMS, QMS, ERP) to notify you of new AI features. Assign someone to skim release notes quarterly.
- Annual attestation. Once a year, each business owner confirms their entries are current. Ten minutes per owner.
Watch out: Model and firmware updates are changes. A vision model retrained by the integrator, or a CMMS vendor switching its prediction engine, can change behavior as much as a new system. If your MOC process exempts "software updates from the vendor", close that gap for AI-enabled systems.
What good looks like
- One register, owned by a named person, covering all sites.
- Every entry has a business owner who knows they are the owner.
- Entries link to OT assets and zones.
- New AI cannot reach production without an entry, because MOC and purchasing check.
- Shadow GenAI is known at category level and has an approved alternative.
- The board or executive team sees a one-slide summary by risk tier each quarter.
Effort
For a single plant of a few hundred people: two to four weeks of part-time work for the first pass, roughly 5 to 8 person-days in total, including interviews. Multi-site groups should pilot at one site, then send the template and interview guide to site leads. Ongoing upkeep, once wired into MOC and purchasing, is a few hours a month for the register owner.
Common mistakes
- Only counting projects. Vendor and OEM features are the majority.
- Asking "do you use AI?" People say no. Ask about predicting, classifying and optimizing.
- IT-only ownership. Without engineering and maintenance, OT AI stays invisible.
- Too many fields. Owners stop updating it.
- No link to MOC. The register is out of date within one quarter.
Where to start
- Name a register owner this week, usually in IT or operations excellence.
- Book seven 30-minute interviews: maintenance, quality, planning, process engineering, IT, HR, purchasing.
- Pull GenAI domain hits from proxy or SaaS discovery for the last 90 days.
- Load results into the register template and assign an owner to each row.
- Add the AI question to the MOC form and purchase requisition. If you are building governance from scratch, the 90-day plan puts this in weeks 1 to 4.
Sources and further reading
- NIST AI Risk Management Framework
- ISO/IEC 42001:2023 Artificial intelligence management system (ISO)
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), EUR-Lex
- 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
What is the difference between an AI inventory and an AI use-case register?
People use the terms loosely. A system inventory lists AI-enabled products and models. A use-case register lists what each is used for, by whom and with what decisions. One vision system can serve two use cases with different risk, for example advisory sorting on one line and automatic reject on another. Track use cases as the main record and link them to systems.
Does the EU AI Act require an AI inventory?
The AI Act does not use the word inventory for deployers. But you cannot work out which obligations apply, such as AI literacy, transparency or high-risk deployer duties, without knowing what AI you use. ISO/IEC 42001 and the NIST AI RMF both expect an organization to know its AI systems. In practice an inventory is the foundation for any compliance claim.
How long does it take to build a first AI inventory for one plant?
For a single site with a few hundred employees, a first pass usually takes two to four weeks of part-time effort: a few interviews per function, a review of the software and equipment list, and a scan of SaaS and Microsoft 365 usage. Expect it to be incomplete. The goal is a credible baseline you improve through MOC and purchasing.
Should shadow GenAI use go in the AI register?
Record it as a category, not individual chats. Note which tools are in use, roughly by which teams and for what kinds of tasks, then decide whether to approve an enterprise alternative or block the tool. Individual approved GenAI use cases that touch restricted data or produce controlled documents should be registered individually.
Related guides
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
AI Governance
AI vendor due diligence for industrial software and OEM equipment
Most AI in a plant arrives through vendors: MES, CMMS, QMS, vision integrators and OEM machines.
Guide · 7 min read
OT Security
OT asset inventory: the foundation for OT security and AI governance
An OT asset inventory is a maintained list of every device, system and connection in your control environment, with the attributes you need to make security decisions.
Guide · 6 min read
AI Governance
How to write an AI policy for a manufacturing company
A manufacturing AI policy has to cover more than office chatbots.
Guide · 8 min read