Key takeaways
- Approve a usable enterprise tool. Bans without an alternative push GenAI onto personal devices.
- Recipes, process windows, drawings, PLC and robot code, customer specs and network details never go into unapproved tools.
- GenAI can draft SOPs and suggest troubleshooting steps, but released documents go through your normal document control.
- Generated PLC code is untrusted code: review, simulate, test offline, and never use it in safety logic.
- LLMs produce confident wrong answers. Anything that touches safety, lockout or energy isolation must be checked against the controlled source.
Walk through any plant and you will find generative AI in use. A maintenance tech asks a chatbot what an E-047 fault on a servo drive means. A process engineer pastes a batch report into a chatbot to summarize a deviation. A controls engineer asks for structured text to debounce a sensor. A quality tech drafts a work instruction. Some of this is on approved tools. A lot of it is on personal phones.
None of these uses is inherently bad. Several are genuinely helpful. The risks are specific: your know-how leaking into tools you do not control, and confident wrong answers reaching places where wrong answers hurt people or product. This page sets out a GenAI policy for the shop floor and engineering that addresses both without killing the benefit. It sits under your main AI policy.
Does this apply to you?
If any employee has a smartphone, yes. The question is whether the use is visible and governed. A quick test:
- Do you have an approved GenAI tool with enterprise data terms?
- Do operators and technicians know which data they must never paste into a chatbot?
- Is there a rule for AI-generated PLC or robot code?
If the answer to any is no, write this policy now.
Protecting IP: what never goes in
Most manufacturing know-how is compact and easy to copy and paste. Write the list in plain words and repeat it in training:
- Recipes, formulations, bills of process, cure and heat-treat profiles
- Process parameters and windows, setpoints, golden-batch data, historian exports
- CAD models, drawings, tooling and fixture designs
- PLC, robot, CNC and HMI code and project files
- Customer specifications, drawings and anything under NDA
- Network diagrams, IP addresses, firewall rules, remote access details, passwords
- Personal data about colleagues
These "restricted" items may go into an approved enterprise tool only for a named, approved use case recorded in your AI use-case inventory. They never go into public or personal tools.
Plant-floor reality: The leak usually is not malicious. It is a night-shift engineer trying to fix a faulting line at 3 a.m. who pastes a code block and an error message into whatever chatbot is on their phone. Give that engineer a better option: an approved tool on the engineering laptop, and a reference library of your own manuals it can search.
Approved tools
Keep the approved list short and visible:
| Tool category | Example status | Allowed data | Notes |
|---|---|---|---|
| Enterprise chat assistant with tenant data protections | Approved | Up to confidential; restricted only for registered use cases | Default tool for office and engineering |
| AI in Office / ERP within your tenant | Approved | As per the underlying system's permissions | Check permissions hygiene before rollout |
| Coding assistant in engineering IDE | Approved with conditions | Non-safety code; no customer code under NDA | See PLC rules below |
| Retrieval assistant over your manuals and SOPs | Approved per use case | Internal, controlled documents | Must cite source document |
| Public consumer chatbots | Not approved for company data | Public only | Blocked on company devices where practical |
| Browser extensions with AI | Not approved | None | Often read every page you open |
Review the list quarterly. Vendor data terms and features change often. Note that enterprise AI that can search your file shares will surface anything users can already access, so tidy up over-shared folders before you switch it on.
Approved uses, with guardrails
SOPs and work instructions
GenAI is good at turning a messy draft, a video transcript or an engineer's notes into a clear, consistently formatted work instruction. Rules:
- The author remains the author. They check every step against the actual process.
- AI-drafted documents go through normal document control: review, approval, version, training record.
- Safety-critical content (lockout/tagout, confined space, hot work, chemical handling, guarding) is reviewed by EHS before release.
- Do not include restricted parameters unless the tool is approved for that use case.
Maintenance troubleshooting
Useful for explaining fault codes, suggesting likely causes, summarizing long manuals and drafting work-order notes. Better still when the assistant is grounded in your own OEM manuals and past work orders and cites the source. Rules:
- The OEM manual and your procedures remain the authority. The assistant can point you to them.
- Any step involving stored energy, pressure, heat, chemicals, guarding or bypassing an interlock is checked against the controlled procedure before anyone acts.
- Never use an assistant's suggestion to justify bypassing a safety device, even "temporarily".
Data analysis and reporting
Summarizing shift logs, drafting 8D reports, explaining SPC charts. Rules: use only approved tools for confidential data, and check numbers against the source. LLMs are prone to miscopying figures. The quality of what you get back also depends on the quality of what you put in; see industrial data quality.
PLC and robot code generation
This deserves its own section because the failure modes are physical.
What can go wrong:
- Plausible but wrong logic. Code that compiles and looks idiomatic but mishandles an edge case: a sensor that fails high, a race between two rungs, a timer that resets on first scan.
- Missing interlocks and permissives. The assistant does not know your machine's hazards or your interlock philosophy.
- Dialect confusion. Mixing IEC 61131-3 structured text conventions with vendor-specific instructions, or inventing function blocks that do not exist in your platform.
- Unsafe defaults. Outputs that energize on startup, missing fault states, no handling of communications loss.
- IP leakage. Pasting a full project or customer-owned code into an unapproved tool.
- Supply chain. Suggestions that pull in unvetted libraries or snippets on PC-based control and edge devices.
Rules we recommend:
- Treat generated code like code from an unknown contractor. It is untrusted until reviewed.
- Peer review by a second qualified engineer, against your coding standard.
- Test in simulation or on a test rig before any download to a live controller. Include fault and edge cases.
- Changes to production controllers go through MOC, with the usual backup and rollback.
- Never use AI-generated code in safety PLCs, safety functions or safety-rated configurations.
- Record in the MOC or commit message that AI assisted, so reviewers know to look harder.
Watch out: "You can't patch a PLC on a Tuesday" cuts both ways. If an AI-assisted change goes wrong on a running line, you may not get another window for weeks. Keep the last known-good project archived and tested before any change.
Hallucination in safety contexts
LLMs generate text that sounds right. NIST's Generative AI Profile calls this risk "confabulation": confidently stated false content. In an office, it costs time. On a shop floor, it can hurt someone.
Verification rules by context:
| Context | Verification required |
|---|---|
| Email, meeting notes, general drafting | Author reads before sending |
| Reports with numbers | Numbers checked against source |
| Work instructions and SOPs | Normal document control; EHS review if safety content |
| Troubleshooting suggestions | Checked against OEM manual and procedure before acting on energy, pressure, chemicals, guarding |
| PLC, robot, CNC code | Peer review, simulation or test rig, MOC |
| Anything in a safety function | Not permitted |
| Customer-facing technical answers | Reviewed by a qualified engineer before sending |
Training and transparency
A 20-minute session covers it for most staff: what tools are approved, what never goes in, why outputs must be checked, and who to ask. In the EU, this also contributes to the AI Act's AI literacy obligation, which as of November 2026 (after the Digital Omnibus) requires deployers to take measures to support staff AI literacy. If you deploy a chatbot that employees or customers interact with, Article 50 transparency obligations that apply from 2 August 2026 may require telling them it is AI.
Effort
- Selecting and configuring an enterprise tool: depends heavily on your existing licenses; often a few weeks of IT time.
- Writing the policy: 2 to 3 days using the AI policy template.
- Training: 20 minutes per person, integrated into shift meetings and toolbox talks.
- PLC code rules: a half-day workshop with controls engineers to agree the review and test steps.
Common mistakes
- Ban without alternative. Use moves to phones.
- Treating Copilot as safe by default. It surfaces whatever permissions allow, including over-shared folders.
- No rule for code. Engineers make their own, inconsistently.
- Trusting fluent answers. The better the writing, the less people check.
- Leaving GenAI out of the register. Approved uses that touch restricted data or controlled documents should be risk-tiered like any other; see AI risk assessment.
Monday-morning checklist
- Confirm which GenAI tool, if any, has enterprise data terms in your contracts.
- Publish the "never goes in" list in plain words, on one page.
- Agree the PLC code review and test rule with your controls lead.
- Add a GenAI segment to the next round of shift meetings.
- Check that any AI tool reaching OT data follows the guidance in securing AI in OT networks.
This page is general guidance, not legal advice. Confirm regulatory obligations with your counsel.
Sources and further reading
- NIST AI 600-1: Generative AI Profile (NIST)
- NIST AI Risk Management Framework
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), EUR-Lex
This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.
FAQ
Frequently asked questions
Can engineers use ChatGPT or Copilot to write PLC code?
They can use an approved tool to draft or explain standard, non-safety logic, provided restricted code is not pasted into unapproved tools and the output is treated like code from an unknown contractor: peer reviewed, tested in simulation or on a test rig, and changed through MOC. Generated code must never be used in safety PLCs or safety functions.
Is it safe to use generative AI for maintenance troubleshooting?
It is useful for suggesting likely causes, explaining fault codes and summarizing manuals, especially when grounded in your own documents. It is not safe as a substitute for the OEM manual, your lockout/tagout procedure or a qualified technician. Require technicians to verify any step involving energy, pressure, chemicals or guarding against the controlled procedure.
Should we block ChatGPT on the plant network?
Blocking public tools on company devices is reasonable once you offer an approved alternative with contractual data protections. Blocking alone usually moves use to personal phones, where you have no visibility. Pair the block with a clear one-page rule, a short training session and an easy way to request new tools.
Does the EU AI Act apply to our use of ChatGPT or Copilot?
As a user of general-purpose AI tools, you are typically a deployer. As of November 2026, the AI literacy duty applies, and if you put a chatbot in front of employees or customers, transparency obligations that applied from 2 August 2026 may require telling people they are interacting with AI. Most internal GenAI use is not high-risk. Confirm specifics with counsel.
Related guides
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
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
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.
Guide · 7 min read
OT Security
Securing AI in OT networks: a practical guide for manufacturers
Treat every AI system that touches OT as a new conduit and a new piece of software you have to patch.
Guide · 8 min read