As announcements go it is unspectacular, and it is worth more attention than it will get.

Why a training programme is an AI story

Every deployment covered on this site runs into the same constraint eventually. A vision system that flags a defect needs an operator who knows when to believe it. A predictive-maintenance alert needs a technician who can tell a developing bearing fault from a sensor drift. Jay Lee's three barriers, presented at IMTS this month, are all decided by people, none of them by a model.

There is also a regulatory dimension that most industrial operators have not yet costed. The EU AI Act's AI literacy duty has applied since 2 February 2025, it binds deployers as well as providers, and it covers the staff who work alongside an AI system. Unlike the high-risk obligations, which the AI Omnibus deferred to 2027 and 2028, this one is in force now.

The inclusion of critical thinking

The most interesting item in the list is the least technical one. A curriculum that teaches an apprentice to operate a digital twin but not to doubt its output produces exactly the failure mode that makes plant managers distrust AI systems: confident automation, uncritical operators, and a fault that nobody questioned until it reached the customer.

Siemens has not disclosed the cost of the programme or whether it extends beyond Germany.

The Digital Omnibus on AI was published in the Official Journal as Regulation (EU) 2026/1744 on 24 July 2026 and entered into force on 27 July, six days before the original deadline. It amends the AI Act and defers the high-risk regime.

The new dates

| What | New deadline | |---|---| | Stand-alone high-risk systems (Annex III): employment, education, credit, law enforcement, critical infrastructure | 2 December 2027 | | High-risk AI embedded in already-regulated products (Annex I): machinery, medical devices, toys | 2 August 2028 |

The second row is the one that concerns this readership. An AI system that is a safety component of a machine, or that is embedded in one, falls under Annex I and now has until August 2028.

The more consequential change

More important than the delay is a change of rulebook. Under the amended framework, embedded AI subject to the Machinery Regulation is removed from the AI Act's direct application, with AI-related safety measures to be introduced through delegated acts under that regulation instead.

For a machine builder this converts a two-framework compliance problem into a one-framework problem, inside the regulation its engineers already work with. For a plant operator buying that machinery, the practical consequence is that the assurances to ask a supplier for will be phrased in machinery-safety terms, not AI-Act terms.

What is already in force

Deferral is not exemption, and several obligations already apply:

• Prohibited AI practices and AI literacy duties, since 2 February 2025.
• General-purpose AI provider obligations, since 2 August 2025.
• Transparency duties under Article 50, since 2 August 2026.

The AI literacy duty is the one most often overlooked in industrial settings. It applies now, it applies to deployers as well as providers, and it covers the operators and maintenance staff who work alongside an AI system, not only the people who procure it.

This article summarises publicly available legal texts and analysis; it is not legal advice. The authoritative text is the regulation itself on EUR-Lex.

The deployment began with a single finish mill in 2024. It has since been commissioned on six more: four cement ball mills and two vertical mills, one for pre-grinding and one for slag.

What was measured

• Automatic operation rate: above 90% on the finish mills.
• Grinding throughput: up 3%.
• Specific power consumption: down 3%.
• Kiln: specific heat consumption down 3%, and manual operator actions down 70%, following an earlier optimisation reported in 2022.

The plant reports that the gains were achieved while maintaining stable operations, which in a grinding circuit is the harder half of the problem: an optimiser that raises throughput by destabilising the mill does not survive contact with a production schedule.

"We decided to extend the use of the solution to the entire cement mill process." — Junya Doi, Engineering Section, Cement Manufacturing Department, Tokuyama

The system combines model predictive control with AI, and is the same product line ABB has been placing in cement and minerals plants for several years. Hiromichi Yoda, who runs ABB's Process Industries division in Japan, framed the result as a step toward autonomous operation of the plant rather than a one-off efficiency project.

Why it matters

Cement is among the most energy-intensive industries in the world, and grinding is where a large share of a plant's electricity goes. A 3% cut in specific power consumption sounds modest until it is multiplied across seven mills running continuously. The 70% reduction in kiln operator actions is the more telling number: it describes a change in how the control room actually works, not just a line on an energy bill.

Figures are those reported by Tokuyama and ABB. They have not been independently verified.

What is deployed

• AI-powered cameras and sensors along the assembly lines, used for automated quality control and error detection.
• The Brilliant Factory data platform, which monitors plants in real time and produces AI-generated reports for plant managers on where problems are and how to solve them.
• Predictive maintenance, including detection of overheating motors.
• AI-driven staffing optimisation and market demand forecasting.

The people and the numbers

Tony Gabbert, Director of Manufacturing Operations at the Lafayette plant, and Bill Good, Vice President of Manufacturing at GE Appliances, are the executives quoted. Good can see which machines are down and why, how many appliances go to the repair bay, and which parts are scrapped and at what cost.

Good estimates that the company saves between $1.5 million and $2 million a year for every percentage point of improvement, and puts assembly-line downtime at $300–500 per minute. The report also notes a recent expansion adding 600 jobs with a $180 million investment.

Good's assessment of the technology gives the piece its headline, "It can outthink me", though he does not anticipate widespread replacement of the workforce in the foreseeable future. Results are as reported by the company to NPR.

Named industrial customers

• Airbus: commercial aircraft, helicopter, defence and space divisions.
• BMW Group: crash simulation and multimodal reasoning models.
• ASML: lithography diagnostics; ASML is also Mistral's largest shareholder.
• CMA CGM: the shipping group uses Mistral's Search Toolkit.

The physics-AI component uses data-driven models trained on solver outputs to predict physical behaviour in seconds on a single GPU rather than hours, according to the report.

"We have two convictions at Mistral. The first is that in order to deploy AI in the enterprise, you actually need, as an AI provider, to own the full stack." — Arthur Mensch, CEO, Mistral AI

Mistral also stated a €1 billion revenue target for 2026, a €4 billion infrastructure investment, about 1,000 employees and a €11.7 billion valuation, with 200 MW of data-centre capacity targeted by 2027 and 1 GW by 2030. Customer results were not disclosed.

The release also describes Lenovo's own use of the technology: an Automatic Quality Inspection Robotic Cell deployed at Lenovo facilities in Brazil, Hungary and Mexico, and a North American site where the company reports an 85% reduction in lead time, a 42% reduction in logistics costs and a 58% productivity increase.

"Manufacturers don't need more AI pilots. They need AI that runs at scale in production." — Jonathan Wu, Chief Technology Officer of Smart Manufacturing, Lenovo

Products named include the Automatic Quality Inspection Robotic Cell, Multi Purpose Robots, Lenovo iChain, ThinkStation PGX and ThinkEdge, the latter positioned for visual inspection and predictive maintenance on the plant floor. Lenovo exhibited in Hall 15, Stand G76.

Figures are those reported by the vendor; Hisense has not published its own account of the deployment.