Manufacturing AI in 2026: seven results, one published price

Seven manufacturers have put an AI result in writing. Only VitrA said what it paid. What the signed results have in common, and the 20 January 2027 deadline.

Seven manufacturers have put a result for a specific AI system in writing in documents we can link to: VitrA Tiles, Unilever, Bosch, Danfoss, Tata Steel, Rockwell Automation and Toyota. Their results range from 56% less scrap on one tile line to new technicians troubleshooting in three months instead of nine. Only one of the seven, VitrA, has said what its system cost: nearly EUR 1 million.

That gap is the most useful thing a 200-person manufacturer can take from this year's evidence. The results are real and specific. The price of getting them is almost never published, which means any business case you are shown that borrows one of these figures is borrowing half a sum.

What is in production, with a figure

Sorted by the job the system does rather than by company, the signed results fall into four groups.

Process settings on the line. VitrA Tiles put machine learning on the milling, spray-drying and sludge stages of one production line at its Bozüyük plant in Türkiye. The company reported 56% less scrap, 19% higher OEE and 14% less energy, and its chief executive put the spend at nearly EUR 1 million for hardware and analytics. Unilever's Indaiatuba laundry powder plant in Brazil reported 20% more capacity and almost EUR 3 million saved in 2024 from a digital twin and machine learning on process parameters. Bosch's Changsha plant cut annual electricity use by 18% with a system that predicts each line's energy draw and schedules production around it.

The order desk and the complaints queue. Danfoss gave the emailed orders that its system-to-system links never caught to AI agents, which now make above 80% of the decisions on them, and its team estimates about five minutes saved per order. Tata Steel says agents that read complaint documents, spot the defect in customer photos and route the case have halved average complaint turnaround.

Maintenance knowledge. Rockwell Automation's Singapore plant built a copilot from fault notes written by engineers with 20 to 30 years' experience, plus manuals and production data. Its chief supply chain officer says new technicians now learn to troubleshoot in three months rather than nine.

Inspection built by plant staff. Toyota let quality and production staff build their own image-inspection models after an outside quote came in above 100 million yen. Staff built 10,000 models across 10 factories in 2024, up from 8,000 the year before.

What the seven have in common

Three things recur, and none of them is the model.

Each result attaches to one number that existed before the project started: scrap rate, energy bill, order turnaround, complaint turnaround, time to a technician's first solo shift. Nobody reported a result against a target invented for the project.

Most of the knowledge was already written down. Unilever's plant had a decade of World Class Manufacturing discipline behind it. Rockwell's copilot is only as good as the fault notes its veterans wrote. Danfoss's agents check orders against an ERP that already held the customer and price data.

The people who knew the process built or shaped the system. VitrA did the modelling entirely with its own staff. Toyota's models are built on the floor. Rockwell's notes came from its own engineers.

Does this apply to a 200-person plant?

Yes, as a filter for where to start. The signed results all sit on one line item with a measured before-number: scrap on one line, energy, order-desk hours, complaint turnaround or technician ramp-up. Pick the one where you already have twelve months of data and written procedures. If you have neither, the first project is writing them down, not buying software.

What is described without a figure

BMW's GenAI4Q system at its Regensburg plant builds a custom inspection checklist for each of the roughly 1,400 vehicles made there each day. BMW calls it a pilot and, in its April 2025 release, gives no result. Tata Steel's halved turnaround comes with no before or after in days and no complaint volume. Danfoss has not published order volumes, error rates or what it paid. Rockwell's plant-wide figures, a third less downtime and about a quarter less servicing and spares cost, cover more than 50 digital and AI tools together, not the copilot alone.

None of that makes the results false. It does mean the gap between a press release and a business case is wide, and the cost line is where it is widest. On our count, six of the seven companies here have published a result without the price of getting it.

Six of the seven manufacturers with a signed AI result have not said what it cost them.

What is theatre

Model counts read as output. Toyota's 10,000 models is a measure of how many staff are building, not of how many models run on a production line or what they catch. Toyota has not published defect or escape rates for them.

Adoption headlines built on a survey change. The Federal Reserve's April 2026 note shows manufacturing with the largest jump of any sector between the old and new versions of the Census business survey. The two versions ask different questions of the same firms, so the jump measures the wording change, not growth over time. On the Fed's figures, roughly one US manufacturer in eight was using AI at all at the end of 2025.

Borrowed plant-wide numbers. Several of these sites are World Economic Forum Lighthouse factories, and their award figures cover dozens of tools at once. Do not put one in your own business case as the return on a single system.

Which vendors keep appearing

The cloud providers sit under most of these results, and they publish the stories. Google Cloud tells the Danfoss, Tata Steel and Toyota stories; Microsoft tells Rockwell's. Danfoss's agents come from Go Autonomous, a Copenhagen software firm. VitrA, Toyota, Bosch and Rockwell built the working parts themselves. For a mid-sized manufacturer that means the deciding factor is less which vendor you choose than whether someone inside the plant understands the process well enough to judge the output.

What regulation is coming, and when

The date that matters is 20 January 2027, ninety-seven days from today. The Machinery Regulation (EU) 2023/1230 applies from then and repeals the Machinery Directive with no transitional period. The July 2026 Digital Omnibus moved the AI Act's high-risk dates back to December 2027 and August 2028, and gave the Commission until 2 August 2028 to write AI-specific safety requirements into machinery law. It did not move the Machinery Regulation's own start date.

The standards picture is thin. According to IBF Solutions, a compliance consultancy that tracks the process, the first list of harmonised standards under the new Regulation is expected before the end of 2026 and should carry over most of the standards already cited under the old Directive. As of August 2026, it reports, no harmonised standards had been submitted for the Regulation's new AI and cybersecurity requirements. A machine builder with a self-learning component near a safety function will meet January 2027 without a ready-made standard for that part.

What this does not mean

It does not mean AI in manufacturing pays only for large companies, or that the unpriced results are inflated. It means the public record tells you where results have come from, mostly single lines and single desks with a known before-number, and almost nothing about what they cost. Ask any vendor who cites one of these companies for the cost line. If they cannot point to a document, plan as if you are paying for the first published price yourself.