Topic
AI
AI in PLC and automation work, with the failures as well as the uses: why chat models get ladder wrong, local models, safety of generated code, and documentation.
7 articles
Best AI for PLC programming in 2026: how to choose and test one
There is no single best tool; the right one works inside your platform, can see your project, and produces code you can compile and test. Siemens, Rockwell and Beckhoff ship assistants in their own engineering software, CODESYS sells an MCP server, and general models need a compile-and-test loop around them, especially for ladder logic.
9 min readTIA Portal and CODESYS MCP servers: what they do, and the risks
An MCP server is a small program that exposes an engineering tool's functions, such as reading code, writing blocks or compiling, as tools an AI client can call through the open Model Context Protocol. CODESYS sells an official one for Structured Text; the TIA Portal ones are third-party. Use one only with version control, approval of each change and no route to a running controller.
10 min readWhy ChatGPT writes ladder logic that does not run
It can write PLC code that looks correct and frequently does not compile. General models are strong on structure and weak on the four things that decide whether PLC code works: dialect consistency, timer semantics, scan order and addressing. None of those are visible by reading, and all of them are visible to a compiler, which is why a validation loop matters more than a better prompt.
8 min readAgentic AI on the plant floor: what is actually running
Engineering assistance is genuinely deployed and runtime autonomy is not. Siemens, Rockwell and Beckhoff all ship code assistants inside their engineering tools, with reported productivity gains of twenty to thirty percent concentrated in onboarding and documentation. None of them writes to a running controller, and the constraint is not technical but regulatory: safety cases are built on determinism and traceability.
10 min readRunning a local model for PLC work
Yes, and on an OT network it is often the only permitted option. A machine with 16 GB of RAM runs a capable 7B to 14B parameter model through Ollama or LM Studio, exposing an OpenAI-compatible endpoint that most tools can point at. Smaller models are noticeably weaker at code, which is why a validation loop around the output matters more locally than it does in the cloud.
8 min readWhere AI-written code stops being acceptable
For non-safety logic reviewed by a competent engineer, it is a drafting tool like any other. For safety functions it is not acceptable, because a safety case requires a documented, deterministic development process and a probabilistic generator cannot provide one. The practical line runs between code that can stop production and code that can hurt somebody.
7 min readThe one AI task in automation with no downside
Yes, and it is the strongest current use case because the risk is close to zero. A model reading a program can produce an I/O schedule, a control narrative and rung comments faster than a person, and errors are visible on the page rather than in the machine. It cannot decide what the machine should do, so it cannot get that wrong.
6 min read