Eighteen months of talk. Now things ship.
Since the first Industrial Copilot previews in late 2023, every major automation vendor has announced something "AI" at the PLC layer. Code-generation copilots, natural-language ladder assistants, agentic HMI builders, on-controller ML runtimes, edge AI retrofits for legacy gear. At Automation Fair 2024 in November, Rockwell and Microsoft jointly demoed a FactoryTalk Design Studio copilot. Siemens shipped the Engineering Copilot TIA to the Xcelerator marketplace in July 2024. Beckhoff renamed TwinCAT Chat to TwinCAT CoAgent and expanded it from "assistant" to "agent." Edge Impulse added first-class PLC support via the Arduino Opta.
I spent the last two weeks putting five of these in front of real code, real alarms, and real production logic. Here is the honest grade sheet.
The four categories of "AI at the PLC"
Before the scorecard, the taxonomy. Anyone who tells you "AI on the PLC" without specifying which of the four things they mean is selling a demo, not a product.
- Code-generation copilots — AI that writes ladder logic, structured text, or HMI scripts from natural-language prompts. The Siemens and Rockwell camps both play here.
- Agentic engineering assistants — same as #1, but with multi-step planning and I/O-configuration tool use. Beckhoff's CoAgent is the clearest example.
- On-controller ML runtimes — neural networks running on the PLC or a directly-attached coprocessor. Siemens' TM NPU module, Beckhoff's TwinCAT Machine Learning, Rockwell's Logix AI.
- Edge ML retrofits — AI bolted onto legacy control hardware you already own. Edge Impulse + Arduino Opta is the canonical cheap entry point; Viam, Opto 22, and several embedded-MCU vendors crowd in behind it.
Each category has a different failure mode and a different justification for the spend. Conflating them is the easiest way to burn a pilot budget.
Scorecard: five products, graded
1 · Siemens Engineering Copilot TIA — B+
What it is. A managed-service copilot that plugs into TIA Portal V19, V20, or V21. Connects local projects to Azure OpenAI GPT via Siemens' Xcelerator cloud layer.
What ships. Structured Control Language (SCL) generation from prompts. Machine-visualization generation that hands off to WinCC Unified. Code explanation and documentation on existing blocks.
Where it holds up. SCL generation on well-scoped function blocks. Documentation and explanation, which is genuinely useful when you inherit a project from someone else.
Where it breaks. Ladder is not its strength — the model generates SCL because that's what the training corpus looks like. Safety logic is explicitly out of scope. The managed-service pricing model means every team that wants access has to negotiate with Siemens; it is not a click-to-enable add-on.
Production traction. thyssenkrupp announced in 2025 it would roll the Copilot across its global engineering locations — the single biggest named reference in the category.
Verdict. The most mature thing in this category today. Plan for a six-week onboarding, not a six-day trial.
2 · Rockwell FactoryTalk Design Studio Copilot — Incomplete
What it is. An Azure-OpenAI-powered generative AI assistant inside FactoryTalk Design Studio, Rockwell's cloud-native controls-design SaaS. Announced jointly with Microsoft at Automation Fair 2024. Natural-language prompts for code generation, troubleshooting, and code explanation.
What ships. At the November 2024 announcement, the copilot was still in prototype state and not available to end users. As of this writing, Rockwell has not published a GA date.
Verdict. I cannot grade what I cannot run. What I will say: the Rockwell footprint in North American process and packaging plants is large enough that when this actually ships, it will matter more than the press-release schedule suggests. The wait is annoying, not disqualifying.
3 · Siemens SIMATIC S7-1500 TM NPU — A-
What it is. A neural-processing-unit module for the S7-1500 and ET 200MP, equipped with an Intel Movidius Myriad X VPU. Trained models drop onto an SD card; the module presents inference results to the controller through the standard I/O scan.
What ships. Has been in production since 2020. TM NPU 2.0 firmware is current. Proven use cases: visual quality inspection, image-guided robot systems, time-series anomaly detection on process data.
Where it holds up. Integrates cleanly into the SIMATIC automation system without a bolt-on PC. Deterministic. Understandable to a controls engineer who is used to I/O cards — because that's what it looks like.
Where it breaks. You need a trained model. If your data-science bench is empty or if your data collection is ad hoc, the module sits on the backplane doing nothing. The tool chain is also Siemens-centric; if your shop is mixed-vendor, the integration cost doubles.
Verdict. Still the single most useful piece of "AI on the PLC" hardware on the market. It is not new. That is exactly why it is reliable.
4 · Edge Impulse on Arduino Opta PLC — A (the hack)
What it is. The Edge Impulse platform, running on the Arduino Opta PLC. GMM and K-means anomaly-detection algorithms. You can stand one up on legacy equipment for a BOM that lands in the low hundreds of dollars.
What ships. Working production deployments. Edge Impulse reports inference latencies as low as 19 ms for anomaly-detection models running on the Opta — inside the control-loop window for anything that is not motion-critical.
Where it holds up. Retrofits. This is the category-killer play if you have a twenty-year-old line and a specific failure mode you want to catch. You do not rip out the Allen-Bradley or Siemens logic; you add a parallel inference path that signals into existing I/O.
Where it breaks. Anomaly detection is not root-cause. It tells you something is off; it does not tell you why. And the Edge Impulse model lifecycle — data capture, labeling, training, redeploy — is a discipline your plant either has or it doesn't. Without it, models drift silently and the warning becomes noise.
Verdict. If your Monday-morning question is "what do I pilot first to get real AI value on a real line without a rip-and-replace," the answer is this. The ceiling on what it can do is lower than the vendor copilots. The floor is higher.
5 · Beckhoff TwinCAT CoAgent — B
What it is. Beckhoff's in-engineering-environment AI agent, renamed from TwinCAT Chat to TwinCAT CoAgent to reflect the shift from assistant to agent. Uses OpenAI, Anthropic, or local models. Generates PLC function blocks, configures I/O, builds HMI pages from sketches.
What ships. CoAgent is a real product inside current TwinCAT. The HMI-from-sketch demo is the one that turns heads — you can hand it a whiteboard photo and get a plausible-if-not-perfect HMI page back.
Where it holds up. Greenfield builds inside a Beckhoff shop. The multi-model choice (bring-your-own-LLM) is the right architecture; not everyone wants the same model or the same privacy posture, and letting shops use local models is a real advantage for defense, pharma, and anyone else on an air gap.
Where it breaks. The agentic planning is still shallow. Ask it to modify three interacting function blocks and it will cheerfully do the first one correctly and break the other two. As with every agent in the consumer space, the gap between "one-shot demo works" and "three-shot production task works" is the gap that matters. Treat it as a first-draft generator.
Verdict. Real product. Real progress since the TwinCAT Chat preview. Still not ready to be trusted unsupervised on safety-adjacent logic.
The three tests to run before you pilot any of these
Whatever you pilot, run all three of these in the first two days. Each one has failed at least one product on this list.
- The safety-rated I/O test. Ask the copilot to generate logic involving a safety input. Either it refuses — correct — or it treats the safety input like normal I/O. The second outcome is disqualifying. No exceptions.
- The change-management test. After generation, can you diff the AI-written output against the prior revision in the project's version-control tool of choice? Can you annotate, approve, or roll back the change exactly the same way you would a human-written one? If the answer is "not quite the same," your audit posture is broken the first time something fails at 3 a.m.
- The hallucination-frequency test. Run twenty realistic prompts drawn from your last ten projects. Count how many produce (a) hallucinated variable names that don't exist in the project, (b) impossible function block references, or (c) code that compiles but is logically wrong. Your internal threshold for a green light should be no higher than 1 in 20 for any of the three failure modes. Most products on the market today, run against a real project, are closer to 1 in 5.
What I would install Monday
The edge retrofit. Category 4. Edge Impulse on an Arduino Opta, or the vendor-equivalent of your choice, targeted at a single known failure mode on a single existing line.
The reasoning has nothing to do with AI and everything to do with the shape of the bet. The edge retrofit is low ceiling, high floor. You cannot burn the plant with it. You cannot embarrass yourself in front of the VP with it. It will either catch the failure mode or it won't, and either answer is useful. The vendor copilots are high ceiling, low floor — they can do more and hurt more, both.
Ship the low-floor bet first. Earn the right to try the high-ceiling one.
Next issue
Issue 02 runs the $240 edge-ML retrofit I staged on a two-decade-old spindle motor. Specific hardware, specific data pipeline, specific cost line. Ships next Monday.
Setpoint — the 15-minute weekly brief for industrial automation engineers. setpoint.news · Independent · Weekly.