Generative AI for engineering
Large language models are entering automation engineering: generate IEC 61131-3 code, draft specifications, translate and document, assist diagnosis. The productivity gain is real — but a model that "makes things up" in a safety context is a hazard. The rule: AI proposes, the engineer validates.
The use cases
Automation copilots
Assisted IEC 61131-3 code generation (ST, LD), function-block suggestions, cross-platform conversion, explanation and review of existing programs.
Requirements & I/O lists
Draft user requirements (URS), I/O lists, P&ID sketches and interlock matrices from a natural-language description.
Manuals, minutes, translation
Operator manuals, HAZOP minutes, consistent multilingual translation of standards and procedures — a strength of language models.
Knowledge assistants (RAG)
Retrieval-augmented generation (RAG) grounds answers in real documentation: manuals, alarm history, lessons learned — for a reliable troubleshooting copilot.
AI proposes, the engineer validates
A generative model produces plausible text, not guaranteed truth. In automation, suggested code must be reviewed, tested and — for a safety function — validated to the applicable lifecycle (IEC 61511, IEC 61508). The generator speeds the first draft; it replaces neither review nor testing. It is a copilot, never the pilot.
| Use case | Maturity | Human validation |
|---|---|---|
| Documentation & translation | mature | editorial review |
| Code explanation & review | mature | engineer's judgment |
| Specification generation | emerging | review + test |
| Automation-code generation | emerging | review + test; if safety, 61511/61508 lifecycle |
| Diagnostic assistant (RAG) | emerging | check against cited sources |
The proportionality rule: the closer the deliverable is to safety or production, the heavier the human validation. A manual is proofread; code driving a safety instrumented function goes through the full lifecycle and its independent assessment.
Containing hallucination
A language model optimises plausibility, not accuracy — it can assert a wrong answer with confidence. In engineering, that risk is managed by systematic safeguards, not by trust.
| Risk | Safeguard |
|---|---|
| Plausible but wrong answer | grounding (RAG) on internal documentation |
| Invented block, address or standard | systematic verification, citations required |
| Code or data leakage | EU / on-premise hosting, usage rules |
| Over-confidence in the output | AI proposes, human validates (EU AI Act Art. 14) |
Key challenges
- Hallucination & verification — the model can invent a block, an address or a non-existent standard; grounding (RAG) and systematic verification are non-negotiable.
- Confidentiality & IP — code, recipes and drawings are sensitive; on-premise deployment or EU hosting, and clear rules on what may be sent to an external model.
- Tool integration — value comes from the copilot embedded in the IDE (TIA Portal, Codesys, AVEVA) rather than a separate chat with copy-paste.
- Liability & validation — who is accountable for generated code? Decision traceability and final validation stay human; the EU AI Act formalizes human oversight.
Frame & references
- EU AI Act — General-purpose AI (GPAI) models carry their own obligations; use in a "high-risk" function imposes documentation, human oversight and risk management.
- ISO/IEC 42001 — Frame the use of generative tools with a management system: policy, roles, traceability of assisted decisions.
Related standard pages
Major players
Industrial copilots
Siemens Industrial Copilot, Codesys AI, Rockwell, AVEVA.
General code copilots
GitHub Copilot, Cursor, Claude, ChatGPT.
RAG & private platforms
Azure OpenAI, AWS Bedrock, Mistral, modèles open-weight on-premise.