As so often mentioned by Jensen Huang, physical AI is seen as the natural follow-on from agentic systems. But, with physical AI suddenly getting all the attention, will agentic AI still get the focus it shoud?

According to Capgemini’s new research on physical AI, two thirds of executives now rank physical AI as a high priority in their automation agenda for the next three to five years. We all knew the age of robotics was coming, powered by major leaps in AI, but other factors are driving the rapid attention on it. Labour shortages, US reshoring pressures, and the natural convergence of foundation models with robotics hardware. Physical AI has arrived as a serious strategic priority for nations and enterprises, not just a research frontier.

But the report, which surveyed more than 1,600 executives across 15 industries, is largely silent on one question of what agentic milestones have to be reached before that ambition becomes operational? The answer, if you follow the technical architecture rather than the press releases, is agentic AI. In some respects they are the same conversation, staged across two timelines.

Physical and agentic AIs are working together

Physical AI, as Capgemini defines it, marks a shift from robots that follow fixed programmed paths to robots that can generalise across tasks, perceive complex environments, and adapt to real-world variation without being reprogrammed. This is meaningfully different from the industrial robotics of the past fifty years. The robot on a traditional assembly line is a sophisticated mechanical arm with a very narrow job description. The physically intelligent robot navigates a dynamic warehouse, handles unpredictable item shapes, and adjusts to conditions it has never seen before.

What makes this possible is not just better hardware, it’s the extension of the agentic paradigm — the same reasoning, planning, and autonomous execution that is reshaping digital workflows — into the material world. NVIDIA’s head of robotics and edge AI put it plainly in the Capgemini report: industrial deployments will increasingly involve fleets of robots with different embodiments working as part of a coordinated system, where each robot acts autonomously while digital AI agents observe across robots, sensors, cameras, and safety systems to orchestrate work.

Physical AI is like the body, agentic the brain, and we are actually seeing them work together for real. Accenture’s Physical AI Orchestrator (built on NVIDIA’s Omniverse) is being deployed at scale where AI agents monitor live digital twins of factory floors, run continuous what-if simulations, and convert insights into precise operational instructions that physical robots then execute. A consumer goods manufacturer using this approach reported a 20% throughput improvement and 15% capex reduction by eliminating iterative, trial-and-error redesign. The physical and the agentic are already intertwined. The question is whether enterprises truly understand the dependency yet.

The sequencing problem

Capgemini’s report contains a warning that the headline number obscures. Only 4% of executives say they are already operating physical AI at scale. Nearly 80% identify scaling as a significant challenge, primarily due to a lack of technology and operating readiness. And at least 80% of physical AI projects are not yet in production.

Like issues with agentic deployments, it’s not because the models fail but because the operational and governance infrastructure around them hasn’t been built out. For physical AI, the stakes of that failure mode are even higher. A digital AI agent that makes a wrong decision can be rolled back. A robot that makes a wrong decision in a live warehouse, on a construction site, or in a healthcare setting cannot. Deploy safety mechanisms independent of the AI layer, not built on top of it, is highlighted by Capgemini as the route to take. But it requires an orchestration and governance architecture to fully implement, and the architecture of course is agentic AI.

The dependency runs deep

The relationship between the two domains goes beyond the obvious case of agents controlling robot fleets. It runs through data, simulation, and learning infrastructure that the physical AI build-out depends on entirely.

Robotic foundation models — the technology that gives physical AI its ability to generalise — are trained using massive simulation environments. Those simulations are managed, iterated, and optimised by agentic systems. The AI-robot-data flywheel that Capgemini identifies as one of the core drivers of the physical AI inflection point where deployed robots generate real-world data that feeds back into improved models, requires agentic infrastructure to close the loop in production. The fleet-level orchestration of thousands of robots across a facility requires multi-agent systems making real-time decisions that no human team can make at that speed or scale.

What this means for enterprise leaders

The 66% physical AI priority figure from Capgemini is a legitimate signal of where executive attention is moving. Physical AI is tangible, competitive, and addresses real operational pressures. But organisations that treat physical AI as a separate strategic initiative from their agentic AI ambitions are likely to hit the same wall twice. Governance frameworks for autonomous agents, the observability and audit infrastructure, the rollback and safety protocols, are not just relevant to software agents running inside enterprise systems anymore. They are the foundational layer that physically intelligent robots will depend on to function reliably.

The organisations that will move fastest on physical AI are not necessarily the ones investing most heavily in robotics hardware either. They are the ones that have done the agentic groundwork by building the orchestration infrastructure, governance, and operational frameworks that allow autonomous systems to be safely trusted.