Siemens' Erlangen plant: AI cut ramp-up engineering by up to 85%

Siemens says AI on its own electronics line cuts ramp-up engineering by up to 85% and planning costs by up to 20%. What is measured, and what is a ceiling.

Siemens said on 19 October 2026 that AI tools built with NVIDIA and tested in live production at its electronics factory in Erlangen, Germany, cut the engineering effort of ramping up a new circuit board by up to 85 percent, cut time-to-market by up to 40 percent, and cut production-planning costs by up to 20 percent. The plant has about 1,000 staff. Siemens did not say what the AI cost, and every result is stated as a maximum, not a typical figure.

That last point is the whole story for an operator. Erlangen is a real factory with a real result, but it is also the showroom for software Siemens sells. Read the figures as the best a supplier has achieved on its own line, with its own engineers, and plan accordingly.

What Siemens actually changed on the line

Three jobs moved into software, according to the release.

The first is the ramp-up. When Erlangen introduces a new circuit board, the line is now prepared in a digital copy of the factory before any physical trial runs. The physical ramp-up then starts from a line that has already been tested in software. This is where the up to 85 percent less engineering effort and up to 40 percent faster time-to-market come from.

The second is inspection. An optical inspection system needs pictures of faulty boards to learn what a defect looks like, and a brand-new board has no defect history. Siemens says preparing inspection used to take several weeks. AI now generates artificial images of faulty boards so the inspection model can be trained before the first real defect appears.

The third is planning. An AI agent now supports production planning, working end to end from orders to the schedule on the line. Siemens puts the saving at up to 20 percent.

Around these three sit more than 100 other AI applications already in use across production, logistics and inspection at the site, which Siemens has been counting since at least 2024.

The useful question for a buyer is not whether 85 percent happened. It is how often it happens.

What do the 85% and 20% figures actually measure?

They are ceilings from Siemens' own plant. Every product figure in the release says "up to", the release does not say how many new boards have gone through the virtual ramp-up, and it does not give a baseline in hours or euros. The 85 percent is a best case on at least one product, not an average across the line.

Two things in the release are stated without a ceiling. Cedrik Neike, who runs Siemens' Digital Industries business, says the plant raises productivity by 7 percent a year. And the joint team built its first new tools in under ten months.

The plant-level number has history. In October 2024 Siemens said Erlangen's productivity had risen 69 percent over four years and its energy use had fallen 42 percent. Our arithmetic: 69 percent over four years is about 14 percent a year compounded, so the 7 percent a year now quoted is lower. The two figures may cover different periods and different definitions, and Siemens does not reconcile them. Neither is attributed to AI alone.

What did it cost?

Siemens has not said. The only money figure is 500 million euros that Siemens is investing in the Erlangen transformation through 2030. In 2024 Siemens described that sum as research and infrastructure, and it includes a semiconductor cleanroom built on site. It is not an AI budget.

Our arithmetic, for scale only: 500 million euros across about 1,000 staff is roughly 500,000 euros per employee. No mid-sized plant should read that as the price of the software. It is the price of turning one plant into a permanent demonstration site.

The humanoid is still a test

The other Erlangen headline this year was a wheeled humanoid robot. In April Siemens said the robot, built by the start-up Humanoid, met its targets in a logistics test at the plant: 60 tote moves an hour, more than eight hours of uptime and pick-and-place success above 90 percent. The release calls this a step that paves the way for production targets. It is a pilot, not a deployment.

For comparison, a success rate above 90 percent still leaves up to one pick in ten that a person may have to finish. That is the figure to weigh against a conveyor or a cart, which do not pick at all but rarely need rescuing.

Does this apply to a 200-person plant?

Partly. The virtual ramp-up needs an accurate digital model of your line, which most 200-person plants do not have and would pay a lot to build. The inspection piece is closer to reach: if you introduce new boards or parts often and lose weeks training inspection, generated defect images attack a real cost you can measure.

Agent-based planning sits in between. It depends on clean, current data on machines, orders and materials. If your planner works from a spreadsheet updated by hand, fix that first; an agent will not.

The Erlangen results also came with resources no customer will have: in January Siemens said it was committing hundreds of its own industrial AI experts to the partnership. The same January release said Foxconn, HD Hyundai, KION Group and PepsiCo were evaluating some of these tools. None has published a result yet.

What to do on Monday

Write down how many engineering hours your last new-product ramp-up took, from first drawing to stable output. Without that number, no supplier's percentage means anything.

Count the weeks it takes to get inspection working on a new part. If it is under a week, the synthetic-image tools are not your problem to solve.

When a supplier quotes an up-to figure, ask three questions in writing: what was the median across all projects, how many projects is the figure drawn from, and was the customer also the vendor. For Erlangen, the answer to the third is yes.

If you are offered a humanoid pilot, measure it against the cart or conveyor you would otherwise buy, at 60 moves an hour, with the cost of the person who catches the one tote in ten.