The robot arm putting parts together on a factory floor is the easy image. It’s not the whole story, and according to Sarah Lee, IDC’s Senior Research Director of Manufacturing IT Strategies, thinking it is the whole story is exactly what’s slowing companies down.
Physical AI has become one of those terms that gets stretched to cover almost anything with a sensor attached to it. Lee’s definition is narrower and more useful: It’s the convergence of AI algorithms (machine learning, computer vision, now multimodal agentic systems) with physical hardware, built into a closed loop. Sense the environment. Interpret the data. Decide. Act in the physical world. Then feed what happened back into the system so it gets better next time.
That loop is the whole point: physical AI senses the problem, decides what to do about it, and takes the action itself, inside the same system, something a traditional dashboard that just flags an issue for a person to fix was never built to do.
Watch the full conversation with Sarah Lee below, or keep reading for the highlights.
The pressure isn’t coming from the tech side
Knowing what counts as physical AI doesn’t explain why it’s suddenly everywhere at once. Ask most people, and they’ll point to better vision models or cheaper sensors. Lee points somewhere else first: labor.
Mining and construction companies are sitting on expensive equipment that’s idle because there’s nobody to operate it. Manufacturing floors are losing decades of tribal knowledge to retirement faster than new workers can absorb it. Add tariffs and supply chain volatility on top, and automation stops being a nice-to-have. “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.”
Most of the equipment this has to run on wasn’t built for it. IDC’s research puts a finer point on the brownfield problem: the majority of manufacturing facilities in operation today are more than two decades old, running legacy PLCs with no native connectivity. That means the hardest bottleneck usually isn’t compute power or model quality. It’s getting clean, usable data out of machines that were never designed to share it.
The technology caught up at a useful moment. Better vision-language models, better simulation and digital twins, and edge AI that can make real-time decisions without a round trip to the cloud. All of it converged to make the closed loop actually workable at the pace a shop floor or a mine site needs.
One architecture, very different floors
The underlying loop stays consistent, but what fills it changes by industry. Discrete manufacturing is furthest along, with robots and cobots doing precision work, inspection, and material movement. Warehouses and hospitals are leaning on autonomous mobile robots and forklifts for material handling. Construction, mining, and agriculture are running autonomous drill rigs and haul trucks in harsh, disconnected environments where the failure tolerance and regulatory bar look nothing like a factory floor.
Lee frames adoption as a maturity curve rather than a single leap, because almost no operation is starting from zero:
- Instrumented operations: sensors feed data back, but a person still makes the call
- Closed-loop automation: AI starts interpreting and triggering action within bounded limits
- Embodied autonomy: machines perceive, plan, and act on their own
- Coordinated systems (still emerging): multiple machines, people, and systems sharing one context and safety state
Most companies are somewhere in the middle of that curve, and Lee’s advice for anyone earlier on it is blunt: check whether your data is actually clean and ready before you pick 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.”
Where ROI shows up (and where it stalls)
The gap between a successful pilot and real ROI is where most projects die. The gains (productivity, throughput, safety, faster commissioning) show up at fleet scale, not pilot scale. Getting there means solving for orchestration across multiple machines and sites, not just proving one robot works in one corner of the plant.
What IDC’s research is picking up right now
That orchestration problem is exactly what IDC’s own research has been tracking, and the numbers show how fast the ground is shifting underneath it. In a May 2026 survey, physical AI ranked as the second-highest AI investment priority for the next 24 months, behind only GenAI assistants and copilots. IDC’s 2026 Worldwide CEO Survey found something similar from the top of the house: 45% of CEOs named physical AI among their top three emerging technologies for investment over the next 12 to 24 months, trailing only agentic and generative AI at scale, and sustainability technology.
Digital twins are moving past visualization, and IDC’s research describes exactly how: the workflow is inverting from sim-to-real toward real-to-sim, where a plant’s own production data trains and validates the AI before it ever touches the physical environment. That turns site-specific plant data into a durable competitive asset. And the share of manufacturers running ten or more agentic AI solutions is expected to double by the end of the year.
Lee pushed back hardest on one framing: “lights out” factories. What she’s actually seeing is closer to lights transformed, with workers moving out of manual execution and into supervision, exception handling, and governance.
“
The most important automation decisions manufacturers will make over the next few years are not about which arm to buy, but about which control plane to trust, whose simulation to build on, and who owns the data their robots generate.
That’s the shift underneath everything else here: the arm was never the hard part. The foundation around it is.
