By: Sushant Shetty, General Manager – Data Scientist at Findability Sciences
Somewhere on a Hyundai assembly line, a robot is going to break down next week. The plant already knows which one.
Hyundai’s predictive maintenance system now forecasts robot failures up to five days in advance, with accuracy above 90 percent. It started at the Asan plant. It is now moving to Ulsan, Jeonju, India and the company’s EV lines, and should cover roughly 1,400 robots by the end of the year.
The prediction is clever. The roll-out is the real story. For most of the past decade, manufacturing AI lived in a pilot graveyard: a promising proof of concept at one plant, an enthusiastic board presentation, and then silence. 2026 is the year that pattern finally started to crack.
The question has changed
The debate about whether AI belongs on the shop floor is over. Rootstock’s latest State of Manufacturing Technology survey found that 94% of manufacturers now use some form of AI, with predictive AI adoption up 12 points to 48%. Deloitte says 84% are already seeing measurable value. In a separate Deloitte study of 600 executives, eight in ten plan to commit at least a fifth of their improvement budgets to smart manufacturing this year. Augury’s survey of 501 leaders points the same way: more pilots are crossing into full deployment.
So the question on the plant floor is no longer “Does it work?” It is “Why does it work in one plant and not in twelve?”
What scale actually looks like
Procter & Gamble shows what getting it right looks like. Working with Siemens, it has taken an AI quality inspection system beyond a single site and is now rolling it out across its manufacturing plants worldwide. The system checks products in real time at full line speed and can flag or reject a defective unit on its own. Scrap is down 10 to 20 percent, depending on the product. The more telling number is speed: new lines come online five to ten times faster than with conventional machine vision. Plant engineers retrain the models themselves, so a new site does not need a data science team flown in.
That is the difference between a pilot and a capability. A pilot proves something can be done once. A capability can be done again, by someone else, somewhere else, next quarter.
The same logic shows up wherever leaders are pulling ahead. At BMW’s Spartanburg plant, Figure’s humanoid robots supported production tied to more than 30,000 X3s, and Leipzig is next. Schaeffler has signed a binding agreement to deploy thousands of humanoids from UK start-up Humanoid across its network. TRMC reports 20 to 50 percent improvements in computational lithography costs using NVIDIA AI.
Some of the biggest gains are happening away from the line altogether. ABB E-mobility has cut quotation turnaround by 80 to 95 percent. For via Hella shrank a component-placement design task from four days to about four minutes. Once AI agents sit inside SAP and Oracle planning systems, AI stops being a project and becomes part of the plumbing.
Why most pilots stall
Across more than 250 enterprise implementations, our experience has been consistent: pilots rarely fail because the model is wrong. They fail because the data is.
A pilot runs on a clean dataset that one dedicated team spent months preparing. Then it meets the real plant: three generations of machines, sensors nobody has calibrated in years, a quality database that has never spoken to maintenance. The algorithm was never the hard part.
The manufacturers who scale do four things from the start. They tie each use case to a number the CFO already tracks. They build on live production data rather than a tidy extract. They put a business leader, not IT, on the hook for the result. And they design for the tenth plant while building the first. That last one is skipped most often, and it matters most.
The 2027 test
As 2027 budgets are signed off, boards should change the milestone they ask for. Proof of concept has been met. Proof of repeatability has not.
The winners of the next five years will not be the companies with the cleverest model. They will be the ones that can take what worked on Line 3 in Pune and have it running on Line 7 in Ohio before a competitor finishes reviewing its pilot.
The pilot era gave manufacturing its proof. The scale era will decide who gets paid for it.