AI Assistants Join the Factory Floor

WIRED

24-02-2025

Feb 24, 2025 6:00 AM Manufacturers already have the data. LLM-powered tools could help them make use of it. A factory worker operates a CNC machine in an industrial manufacturing plant. Photograph: Getty Images
The basic machine for grinding a steel ball bearing has been the same since around 1900, but manufacturers have been steadily automating everything around it. Today, the process is driven by a conveyor belt, and, for the most part, it's automatic. The most urgent task for humans is to figure out when things are going wrong—and even that could soon be handed over to AI.
The Schaeffler factory in Hamburg starts with steel wire that is cut and pressed into rough balls. Those balls are hardened in a series of furnaces, and then put through three increasingly precise grinders until they are spherical to within a tenth of a micron. The result is one of the most versatile components in modern industry, enabling low-friction joints in everything from lathes to car engines.
That level of precision requires constant testing—but when defects do turn up, tracking them down can present a puzzle. Testing might show a defect occurring at some point on the assembly line, but the cause may not be obvious. Perhaps the torque on a screwing tool is off, or a newly replaced grinding wheel is impacting quality. Tracking down the problem means comparing data across multiple pieces of industrial equipment, none of which were designed with this in mind.
This too may soon be a job for machines. Last year, Schaeffler became one of the first users of Microsoft's Factory Operations Agent, a new product powered by large language models and designed specifically for manufacturers. The chatbot-style tool can help track down the causes of defects, downtime, or excess energy consumption. The result is something like ChatGPT for factories, with OpenAI's models being used on the backend thanks to the company's partnership with Microsoft's Azure.
Kathleen Mitford, Microsoft's corporate vice president for global industry marketing, describes the project as 'a reasoning agent that operates on top of manufacturing data.' As a result, Mitford says, 'the agent is capable of understanding questions and translating them with precision and accuracy against standardized data models.' So a factory worker might ask a question like 'What is causing a higher than usual level of defects?' and the model would be able to answer with data from across the manufacturing process.
The agent is deeply integrated into Microsoft's existing enterprise products, particularly Microsoft Fabric, its data analytics system. This means that Schaeffler, which runs hundreds of plants on Microsoft's system, is able to train its agent on data from all over the world.
Stefan Soutschek, Schaeffler's vice president in charge of IT, says the scope of data analysis is the real power of the system. 'The major benefit is not the chatbot itself, although it helps,' he says. 'It's the combination of this OT [operational technology] data platform in the backend, and the chatbot relying on that data.'
Despite the name, this isn't agentic AI: It doesn't have goals, and its powers are limited to answering whatever questions the user asks. You can set up the agent to execute basic commands through Microsoft's Copilot studio, but the goal isn't to have the agent making its own decisions. This is primarily AI as a data access tool.
That's particularly valuable in manufacturing, where tracking down a set of errors might mean comparing data across quality assurance systems, HR software, and industrial control systems like kilns and precision drills. Within the industry, this is known as the IT/OT gap: the disconnect between information tech like spreadsheets and the operational tech that's used in a factory. AI companies believe large language models like the Factory Operations Agent will be able to work across that gap, allowing it to answer basic troubleshooting questions in a conversational way.
The Factory Operations Agent is due to leave public preview later this year, making it broadly available to Azure AI users. But there will be plenty of competing systems hoping to play a role on the factory floor. As tech companies look for ways to make money from recent breakthroughs in LLMs, manufacturing has proven to be a tempting target. Last September, Google rolled out an update to its Manufacturing Data Engine specifically aimed at unlocking data held on industrial devices, and both Microsoft and Google maintain platforms where independent developers can test out systems with different fine-tuning strategies and different tolerances for risk.
That competition is good for the field, but the increasing use of industrial AI also raises the stakes for safety—particularly on the factory floor, where malfunctions can be a matter of life or death. Crucially, the Factory Operations Agent only manipulates data rather than directly controlling machinery, but there are still concerns. Speaking in his personal capacity, Duncan Eddy, executive director of the Stanford Center for AI Safety, says the biggest concern for AI models like the Factory Operations Agent is simply that users won't recognize when the system is starting to fail, or won't know how to intervene once they do.
'These systems can fail in new and surprising and unpredictable ways,' he says.