Siemens has set out plans to turn its Erlangen Electronics Factory into what it describes as the world's first fully AI-run adaptive production site, with its Industrial Copilot working the floor around the clock. The plans were presented in 2026, and 2026 is the date attached to the site. In parallel the company launched Digital Twin Composer, a simulation platform built on Nvidia's AI technology that lets manufacturers build complete virtual copies of factories, lines and plants and test changes before touching anything physical. PepsiCo ran its US facilities through it, catching up to 90% of potential problems before they appeared in the real world and raising productivity 20% in three months. Both products are aimed at one figure: a shortage of skilled specialists scaled at $8.5 trillion.
Industrial Copilot is not new. Siemens introduced it in 2023, embedded directly into TIA Portal, the software environment its engineers already use to program and configure factory equipment. The engineer describes the behavior they want in ordinary language — Siemens offers examples like a conveyor that should stop automatically when it detects a jam, or a warning raised when operating temperature exceeds a safe level — and the assistant produces the automation code.
More than 100 companies now use it, according to Siemens. One is Krause Automation, formerly ThyssenKrupp, which makes car batteries and transmissions. Industrial Copilot generates the control code for its automation equipment, and work that used to take several hours now takes roughly 30 seconds.
That 30 seconds is the number everyone will repeat, so be precise about what it measures. It is a vendor figure, for one customer, covering code generation. It is not the time to review, test and commission that code, and Siemens itself says engineers still have to check and refine what the assistant writes, while technicians still make the physical changes to the equipment. The honest version of the claim is that drafting collapsed, not that the job did.
The digital twin side is the less discussed half and possibly the more durable one. Digital Twin Composer lets Siemens customers and its own plant operators price a change before making it: model the line, run the change, see what breaks. PepsiCo used it to model its US production facilities, surfacing up to 90% of problems before they occurred and lifting efficiency at its Gatorade plant. The value here is not speed but cheaper failure. Experiments stop costing production time, and the question of whether a change will actually produce the intended effect can be answered before anyone reconfigures a machine.
Taken together, the two products describe a specific strategy: lower the skill floor for automation work and shorten the loop between idea and deployment. Experienced engineers cover more tasks, junior staff become useful faster, and the number of machines a given headcount can program, update and maintain goes up without headcount going up with it. This is why Siemens keeps saying augmentation rather than replacement, and in the narrow sense it is right — the verification work does not disappear.
The three things Siemens did in sequence are the transferable part. It started from a problem that hurts the company and its customers identically: not enough qualified workers, and hiring and training too slow to close the gap. It then put the AI inside the tool its engineers already open every day rather than shipping a new system alongside it, which removes most of the friction that kills internal adoption. And it ran the thing on its own site in Erlangen, on production problems it already understood, where success and failure are both immediately visible — unlike a customer deployment, where the vendor waits for feedback from conditions it does not control.
Here is my read. The Erlangen plan is the strongest signal in the announcement and the weakest documented. A target year identical to the announcement year is either high confidence or a sign the build was substantially done before anyone said so, and the announcement does not distinguish between them. "First fully AI-run adaptive production site" is Siemens' own framing of Siemens' own factory, assessed by Siemens. That is not a reason to dismiss it — running your technology on your own line is the most credible thing a vendor can do, and far more credible than a pilot slide — but it does mean the only entity capable of grading the result is the one with the most invested in a good grade.
Notably absent from all of this is any number about being wrong. Nowhere in the 100-plus customer figure, the 30-second figure or the 90% figure is there a defect rate for generated code, a description of what a bad line of automation logic costs on a live line, or an account of who signs off when the assistant is confidently incorrect. The $8.5 trillion shortage is the justification for the whole program, but the shortage is precisely of the people qualified to catch the assistant's mistakes. The tool expands what a scarce engineer can supervise; it also expands what a scarce engineer must supervise.
Which is the tension Erlangen will actually resolve. If the first fully AI-run adaptive plant is Siemens' own, then the first serious failure of one will also be Siemens' own — and that is the real argument for doing it there rather than at Krause Automation or PepsiCo. Customers evaluating this should ask for the Erlangen incident log, not the Erlangen demo.