Agentic AI came on like a storm over the past year or so, but blazed a trail littered with failed projects and cutting-edge high-tech junk that companies are still trying to sort out. So it’s perhaps no surprise that tech industry execs are urging enterprises to move cautiously with physical AI, where mistakes can have far-reaching business and societal consequences.

“We need clear boundaries…, definitions, and rules,” said Tianlan Shao, CEO of Mech-Mind Robotics, which makes industrial robots.

Companies can’t afford to experiment with something like a chainsaw-wielding robot, unless it’s in a controlled environment and monitored by humans to avoid mistakes, Shao said at a panel session at the recent World Economic Forum (WEF).

More than half of companies globally already use physical AI in some form, and that figure will grow to 80% in the coming years, Deloitte said in its State of AI in the Enterprise study, published last month. Uses include robots, drones, inspection devices, intelligent security cameras, forklifts, and other industrial applications.

“Physical AI use cases…in controlled domains such as factories and warehouses tend to progress much faster than use cases in open, real-world environments, where the challenges and risks are far greater,” Deloitte said in its study.

Physical AI discussions have been more focused on actual pilots and outcomes rather than “Jetsons-like” futuristic robots, said Francisco Martin-Rayo, CEO of Helios AI, who tracked conversations about the technology at WEF in Davos. “The emphasis has been on deployment in constrained environments: logistics, agriculture, energy, and manufacturing, where labor shortages and efficiency gains are very real problems today,” Martin-Rayo said.

Physical AI is expected to advance more slowly than software AI, but with deeper, longer-lasting consequences for how societies function once it does, Martin-Rayo said.

Physical AI has challenges outside of trials that could also slow progress, said Nacho De Marco, CEO of BairesDev, who also tracked the hot topics of conversation at WEF. There’s no “ChatGPT for robots” yet, and hardware challenges such as power consumption, mobility, and cost hinder development of the technology.

“Many believe we’re still in the floppy disk stage of [physical] AI,” De Marco said.

Though potential use cases are emerging in eldercare, logistics, and industrial automation, the chatter at WEF was “more foundational than flashy, which, frankly, is a good sign,” De Marco said. “We’re missing a standardized development layer for physical AI. Everyone’s building their own ecosystem, which slows adoption.”

Questions also remain about how well software in a virtual world marries with the  physical world, said Jinsook Han, chief agentic AI officer at Genpact. “I think it’s that layer that will take time for us to define,” Han said. “The question is more about what we are willing to let physical AI do and how much.

“I’m not a futurist to say seven years out or five years out. I think we are very much getting closer,” Han said.

The building blocks for physical AI were laid more than a decade ago, said Beena Ammanath, global head of Deloitte AI Institute. It started with IoT and sensors, followed by robotic process automation and data science, and most recently autonomous execution through agents.

“The underlying foundation was laid…12, 13 years ago, which is now enabling us to push more of this intelligence to that foundation,” Ammanath said.

Physical AI is becoming especially prominent in monitoring and security systems, where intelligence cameras can handle alarms by injecting intelligence into the devices. Beyond that, many collaborative robots can now communicate with each other and make decisions. And there are increasing instances of physical AI use in retail stores.

“In consumer shops…, there is more push towards having your returns or point of sales being automated and being able to have conversations. There’s always some form of LLM in the back powering it,” Ammanath said.

AI has been driving industrial processes for years, and productivity gains are showing, said Deepak Seth, director analyst at Gartner. Newer automobile plants, for instance,  are totally dark inside because robots work all day and don’t need to see each other, saving on electricity bills.

“The next step would be making it more humanoid and more part of our daily lives, like an AI-based robot that figures out what you want to eat and cooks it for you,” Seth said.

Genpact, which is well established in agents, is also looking to advance in physical AI, Jinsook Han said. “Our heritage is GE. From supply chain to manufacturing, we are absolutely vested in physical AI and it has been in our thoughts for a long time,” Han said.