SEC-01 / AI Governance

How to write an AI policy for a manufacturing company

A manufacturing AI policy has to cover more than office chatbots. It must reach vision systems, vendor features in MES and OEM machines, protect process know-how, keep AI out of safety functions without proper engineering, and route changes through your existing MOC.

8 min read Updated Plant managersCIOsCISOsQuality leadersEHS leaders

Key takeaways

  • Scope the policy to every AI system you use, buy or embed, including OT and OEM equipment, not only generative AI tools.
  • Classify data first: recipes, process parameters, CAD and PLC code are crown jewels and need explicit rules.
  • Exclude AI from safety instrumented functions unless it goes through full functional safety engineering.
  • Use your existing MOC and purchasing processes as the approval gates instead of building a new bureaucracy.
  • Give operators a one-page version they can actually read at the start of a shift.

Most AI policies we see in manufacturing were written by IT in a hurry after someone pasted a supplier contract into a public chatbot. They say "do not put confidential data into ChatGPT" and stop there. That covers maybe a fifth of the AI in a typical plant. The rest is sitting in vision systems on the line, predictive maintenance add-ons in the CMMS, scheduling optimizers in the MES and features your OEMs shipped in the last firmware update.

A useful AI policy for a manufacturer does four jobs. It defines what counts as AI and what is in scope. It sets rules for data, especially process know-how. It draws hard lines around safety and quality decisions. And it plugs approvals into processes you already run, mainly management of change (MOC) and purchasing. This page walks through each, gives an outline you can adapt, and ends with a one-page operator version.

Does this apply to you?

If you are a manufacturer with 100 or more employees, the answer is almost certainly yes, even if you have never launched an "AI project". Ask three questions:

  • Does anyone in the business use GenAI tools such as ChatGPT, Copilot or Gemini, approved or not?
  • Do any of your machines, cameras or software packages advertise "AI", "machine learning", "smart" or "self-optimizing" features?
  • Are you in the EU, or do you sell into it?

Two yeses and you need a policy. In the EU there is also a direct legal hook. The AI Act's AI literacy duty (Article 4) has applied since 2 February 2025. As of November 2026, after the Digital Omnibus on AI (Regulation (EU) 2026/1744), it requires providers and deployers to take measures to support AI literacy among their staff rather than guarantee a specific level. A policy plus basic training is the simplest evidence that you have done so. See our EU AI Act guide for the wider timeline.

Scope: what the policy must cover

Define AI broadly enough to catch the systems that matter, then use risk tiers to keep the overhead proportionate. A workable scope statement covers:

Category Plant examples Typical owner
Generative AI tools ChatGPT, Microsoft Copilot, coding assistants, AI in Office and ERP IT
AI in business software Demand forecasting in ERP, AI features in MES, CMMS, QMS IT with process owner
AI in OT and on the line Vision inspection, anomaly detection on historian data, AI-tuned control loops Engineering / OT
AI embedded in OEM equipment Smart welders, self-adjusting CNC, robot cells with learned paths Engineering with OEM
AI built in-house Data science models, Python scripts in production, Excel with ML add-ins Owner of the use case
AI used about people Productivity analytics, camera-based safety monitoring, HR screening HR with legal

The last row matters more than its size suggests. Monitoring workers with AI pulls in data protection law, works council rights in several EU countries, and in some cases the AI Act's high-risk employment category. Treat it as its own class.

Plant-floor reality: The OEM row is where you have least control. You may not even know a machine uses machine learning until a technician mentions it. The policy cannot fix that alone, which is why purchasing and FAT/SAT checklists need an AI question. More on that in building an AI use-case inventory.

Data classes: protect the crown jewels

The most expensive AI mistakes in manufacturing are IP leaks, not rogue robots. Your competitive advantage often lives in a few hundred kilobytes: a recipe, a cure profile, a set of weld parameters, a tooling drawing. Write the data rules first.

A four-class scheme works for most mid-market plants:

Class Examples GenAI public tools Approved enterprise AI AI vendors / cloud
Restricted (crown jewels) Recipes, formulations, process parameters and windows, CAD and tooling drawings, PLC and robot code, customer specs under NDA Never Only named, approved use cases Only with contract terms, named in register
Confidential Costs, supplier pricing, quality data, OEE and downtime data, network diagrams Never Yes, within tenant Case by case
Internal SOPs without parameters, training material, org charts No Yes Yes with standard terms
Public Marketing material, published datasheets Yes Yes Yes

Two points that often get missed:

  • Network diagrams, IP plans and firewall rules are restricted for security reasons, not just commercial ones. Pasting them into a chatbot to "tidy up the documentation" is a gift to an attacker.
  • Process data in aggregate can be a recipe. Six months of historian tags from a reactor tell a competent engineer most of what the recipe says. Treat bulk historian exports to AI vendors as restricted.

Safety-critical exclusions

Write this section with your EHS lead and controls engineers, not after them. The rule we recommend:

  1. AI must not be part of a safety instrumented function, safety PLC logic or machine safeguarding unless it has gone through the full functional safety lifecycle for that function (for example IEC 61508, IEC 62061 or ISO 13849 as applicable) and the safety authority in your company has signed it off. In practice, for most mid-market plants, this means "not at all".
  2. AI may advise operators in safety-relevant situations (for example, suggesting a cause for an alarm), but the procedure, the operator and the safety system stay in charge.
  3. AI that writes setpoints directly to the control system (closed loop) needs a documented risk assessment, defined operating limits enforced outside the AI, a manual fallback, and approval through MOC.

Watch out: The EU AI Act treats AI that is a safety component of a machine as potentially high-risk. As of November 2026, the Digital Omnibus has moved machinery-embedded AI toward a sectoral approach under the Machinery Regulation, with the relevant obligations applying from 2 August 2028. The detail is still settling. See AI Act and the Machinery Regulation before you rely on any exemption.

Approvals: use MOC and purchasing, not a new committee

A new AI approval board that meets monthly will be bypassed within a quarter. You already have two gates every change in a plant passes through: MOC for anything touching the process or equipment, and purchasing for anything you buy. Add AI to both.

  • MOC form: add one question. "Does this change introduce or modify AI or machine learning (including vendor features)?" If yes, attach the AI triage and add the AI owner as a reviewer.
  • Purchase requisitions and RFQs: add a tick box and a short vendor questionnaire for any software or equipment with AI features.
  • IT service requests: route requests for new SaaS or GenAI tools through the same triage.

Then tier the approvals by risk:

Tier Example Approver
Low GenAI to draft an email, AI in Office with internal data Line manager, logged in register
Medium Vision inspection advising an operator, PdM alerts, scheduling suggestions Process owner plus IT/OT review
High Closed-loop setpoint changes, automated quality release, worker monitoring, customer-facing outputs AI governance group, plus MOC and legal where relevant
Prohibited AI in safety functions without functional safety engineering; emotion recognition of workers Not approved

The AI risk assessment page explains the tiering in more depth. The roles behind these approvals are on roles and RACI.

A policy outline you can adapt

Keep the main policy to six to ten pages. Push detail into standards and checklists that can change without re-signing the policy.

  1. Purpose and principles. Why you use AI, and four or five principles (human accountability, safety first, protect know-how, proportionate controls, transparency with employees).
  2. Scope and definitions. The categories table above. A plain definition of AI that includes vendor and embedded features.
  3. Roles. Policy owner, AI governance group, use-case owners, IT, OT/engineering, quality, EHS, HR, legal.
  4. Data classification and handling. The data class table, with reference to your information security policy.
  5. Approved tools and acceptable use. Link to the approved tools list, maintained separately. Rules for GenAI, detailed in a shop-floor GenAI policy.
  6. Risk tiers and approval routes. Including the prohibited list and safety exclusions.
  7. Lifecycle requirements. Register entry, testing before go-live, monitoring, change control for model updates, decommissioning.
  8. People and workforce. Worker monitoring rules, consultation with employee representatives, training.
  9. Third parties. Vendor due diligence, contract minimums, OEM equipment.
  10. Incidents and exceptions. How to report AI-related incidents and how to request an exception.
  11. Review. Annual plus event-driven.

The one-page operator version

Operators, technicians and team leaders will not read ten pages. Give them one, laminated if you like, written at the reading level of a good SOP:

USING AI AT [SITE]: THE SHORT VERSION

1. Use only approved AI tools: [list]. Ask your
   supervisor if unsure.
2. Never type or upload recipes, set points, drawings,
   PLC code, customer specs or network info into any
   AI tool that is not on the approved list.
3. AI can suggest. You decide. Follow the SOP and
   the safety system, not the screen.
4. If an AI system (camera, alarm helper, chatbot)
   seems wrong, stop, use the manual method, tell
   your team leader and log it.
5. Do not bypass or retrain an AI system on the line.
   Changes go through MOC.
6. No AI tool may be used to record or rate
   colleagues unless the company has announced it.
7. Questions: [name, extension].

What it costs in effort

For a single-site manufacturer with a lean team, expect roughly:

  • Drafting the policy from a template: 3 to 5 days of one person's time, spread over two to three weeks.
  • Review cycle with ops, IT, OT, quality, EHS, HR and legal: two 90-minute workshops plus written comments.
  • Works council consultation where applicable: plan for several weeks, not days.
  • Operator briefing: 15 minutes in existing shift meetings, plus a short e-learning for office staff.

Multi-site groups should write one group policy and allow site appendices only for local law and site-specific tools.

Common mistakes

  • Writing a GenAI policy and calling it an AI policy. It leaves vision, PdM and OEM AI ungoverned.
  • Banning everything. Use goes underground onto personal phones.
  • No owner. A policy nobody maintains is out of date when the next vendor update ships.
  • Ignoring OT. Engineering is not consulted, so the policy has no teeth on the line.
  • Building a parallel approval process. Use MOC and purchasing.
  • Forgetting employees' representatives in countries where they have consultation or co-determination rights.

Monday-morning checklist

  • Pull your current IT acceptable use policy and see what it says about AI.
  • Ask engineering and maintenance to name every system they believe uses AI or machine learning. Start the AI use-case register.
  • Draft the data classification table with the people who own recipes and drawings.
  • Add the AI question to your MOC form and purchase requisition this month.
  • Start from the AI policy template rather than a blank page, and check where you stand with the AI governance readiness assessment.

This page is general guidance, not legal advice. Confirm regulatory obligations with your counsel.

Sources and further reading

  1. Regulation (EU) 2024/1689 (Artificial Intelligence Act), EUR-Lex
  2. Regulation (EU) 2026/1744 (Digital Omnibus on AI), EUR-Lex
  3. NIST AI Risk Management Framework
  4. ISO/IEC 42001:2023 Artificial intelligence management system (ISO)

This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.

FAQ

Frequently asked questions

Do manufacturers need a separate AI policy, or can it go in the IT acceptable use policy?

A short GenAI section in the IT acceptable use policy is a reasonable start, but it will not cover vision systems, predictive maintenance models or AI features in OEM machines. Most manufacturers end up with a short standalone AI policy that references the IT, OT security, quality and data protection policies rather than duplicating them. Keep it to six to ten pages plus a one-page operator version.

Should an AI policy ban ChatGPT and other public chatbots?

Outright bans tend to push use onto personal phones where you cannot see it. A better pattern is to approve one or two enterprise tools with contractual data protections, ban pasting restricted data such as recipes, drawings and PLC code into anything else, and explain why. Review the approved list every quarter because vendor terms change often.

Who should approve an AI policy in a mid-market manufacturer?

The executive team or managing director should sign it, after review by operations, IT, OT or engineering, quality, EHS, HR and legal. In EU countries with works councils, involve employee representatives early if the policy touches monitoring or performance data. Name an owner, usually the head of IT or operations excellence, who keeps it current.

How often should a manufacturing AI policy be reviewed?

Review it at least once a year and whenever something material changes: a new regulation applies, a major vendor changes its data terms, or you approve your first AI use case that influences process control or quality release. A short annual review plus event-driven updates is proportionate for most mid-market plants.

Related guides

Know where your plants stand

Score your AI governance, OT security and data readiness in about 10 minutes each. Results stay in your browser.