A factory run entirely by machines with no people inside, the so-called "lights out" plant, makes a good picture and a poor plan. That is the case Sarah Lee, IDC's Senior Research Director of Manufacturing IT Strategies, makes in an IDC blog write-up, and the evidence she offers supports it: physical AI is being pushed by missing workers and ageing equipment, and what holds it back is mostly data and scale, not robots lacking cleverness. This is analysis of one IDC conversation, so read it as one analyst's view, though a well-grounded one.
A narrower definition than the hype
Physical AI, Lee says, has become a label stretched over nearly anything with a sensor on it. Her definition is tighter. It means AI software (machine learning, computer vision and now multimodal agentic systems) tied to physical hardware in a loop that senses surroundings, reads the data, makes a decision, acts in the real world and then learns from what happened.
That is what separates it from a traditional dashboard. A dashboard flags a problem and waits for a person. Physical AI spots the problem, decides and acts inside the same system. Notice what the definition does not require: the absence of people. It describes a way of working, not an empty building.
The pressure is coming from labor
Asked why this is suddenly everywhere, Lee skips the better-cameras answer and points to workers. Mining and construction firms own expensive equipment that sits idle because nobody can operate it. Factories are losing decades of tribal knowledge to retirement faster than newcomers can absorb it. Tariffs and supply chain volatility add to the strain. Consultancy Roland Berger makes a similar point in its own analysis, saying ageing populations in major manufacturing economies are shrinking the pool of skilled labor.
"The urgency really isn't coming from the tech side," Lee said. "It's coming from a workforce and operating environment that needs to be more resilient when hiring alone isn't enough to solve the problem."
If that is right, the goal is to keep operations running when people are scarce, which is a different goal from removing people. It also hints at why the technology might be used to capture know-how from experienced staff, though that last step is our reading, not a claim Lee makes in the material we saw.
The old-plant problem
The less glamorous obstacle is the equipment. IDC's research, as Lee describes it, finds that most manufacturing facilities in operation are more than two decades old and run legacy programmable logic controllers (PLCs) with no native connectivity. We could not find the survey behind that claim, including its sample size or method, and we have no independent count of plant ages to check it against, so treat that figure as an unverified claim from one analyst. Other research points the same way on the barriers, though. A July 2026 Tata Consultancy Services survey of 300 manufacturing executives in North America and Europe found 68% of manufacturers still in non-deployment or experimental stages, with legacy system integration, data infrastructure and workforce skills named as the main barriers, according to IT Brief's report on it. If Lee's claim holds, the hardest bottleneck is usually not computing power or model quality. It is pulling clean, usable data out of machines never designed to share it.
The money is starting to follow the same logic. Startup Harmoni raised $10 million in September 2026, led by Bessemer Venture Partners, to retrofit existing factories, a market PYMNTS put at roughly 280,000 US plants still running on traditional machinery, a figure it attributed to Business Insider.
The technology did arrive at a useful time, by her account: better vision-language models, better simulation and digital twins, and edge AI that decides in real time without a round trip to the cloud. Those make the loop workable at the speed a shop floor or mine site needs. They do not rewire a 20-year-old plant.
Different industries, different machines
The loop stays the same while what fills it changes. Discrete manufacturing is furthest along, with robots and collaborative robots doing precision work, inspection and material movement. Warehouses and hospitals lean on autonomous mobile robots and forklifts. Construction, mining and agriculture run autonomous drill rigs and haul trucks in harsh, poorly connected places where tolerance for failure and the regulatory bar look nothing like a factory's.
Lee describes adoption as a maturity curve, not one leap, since almost no operation starts from zero, and most companies sit somewhere in the middle. That description is her characterization, and we found no independent data on the stages of that curve, though the TCS survey's 68% still in early stages fits the picture. Her advice to those earlier on is blunt: check that your data is clean before choosing a pilot. "Where do you first deploy? Where do you first do the POCs? It's where your data's clean, your data's ready, and it's relevant, good data."
The strongest counterargument
A fair critic would say that lights-out is still the logical endpoint, and that today's human-heavy setups are a stage, not a destination. Some tasks in controlled settings may well go fully automatic, and the source does not rule that out. It also gives no adoption rates, cost figures or ROI numbers in the text we reviewed, so we cannot test Lee's picture against hard data. IDC's piece is a conversation highlight, not a published dataset, and the maturity-curve stages were not included.
Still, the counterargument has to explain the pilot problem. Lee says the gap between a successful pilot and real return on investment is where most projects die. Gains in productivity, throughput, safety and faster commissioning show up at fleet scale, not pilot scale, and getting there means orchestrating many machines across multiple sites, not proving one robot in one corner of a plant. Roland Berger reaches a similar conclusion, writing that "an AI inspection model cannot remain a stand-alone pilot" if the goal is system-level improvement, and that the real task is connecting isolated use cases into an IT and operational technology architecture that scales across plants. That is a coordination and data challenge that an empty building would not solve.
What would prove this right or wrong
The argument holds if companies that fix their data first scale beyond pilots faster than those that buy robots first. It weakens if fully unattended plants start appearing at meaningful scale in older facilities without heavy data cleanup. Neither test has public numbers yet in what we have seen. Watch for IDC's underlying survey data and for case studies that report fleet-scale results, not single-site demos.
If you work on a factory floor, a mine or a warehouse: is your biggest obstacle finding people, or getting your machines to talk?
Sources
Reported by the WattsUpNext desk from the sources linked below. Spot an error? Tell us at corrections@wattsupnext.com.
The WattsUpNext Brief
Get stories like this in your inbox.
One email each weekday, only the topics you choose. Real news, sourced, no fluff. Unsubscribe in one click.
