A robot patrols a network facility and flags an anomaly before any human operator sees it. An autonomous system reroutes traffic before an alert reaches the operations team. A base station reconfigures itself in real time based on a model running metres from the antenna.
These are no longer experimental scenarios. They are operational realities already emerging across telecom, infrastructure, logistics and industrial environments. And they represent something fundamentally different from the AI most organizations still believe they are deploying.
This is not simply cloud AI running closer to the edge. The intelligence is local, the decision is immediate, and the consequences are physical. That distinction matters more than most leadership teams realize.
Most people hear “Physical AI” and think about AI embedded into devices: robotics, drones, sensors or smarter cameras. But the defining characteristic is not where the model runs. It is whether the system can act autonomously in the real world, without human intervention, and produce operational consequences at machine speed.
A traditional software failure might generate an error log or degrade a user experience. A failure in Physical AI can reroute traffic incorrectly, trigger an automated response, misconfigure infrastructure or escalate an operational incident before a human fully understands what happened. That changes the nature of operational risk entirely.
Most organizations are still governing AI as if it behaves like traditional software. Traditional governance assumes systems are deterministic, predictable and stable unless someone intentionally changes them. Physical AI breaks that assumption.
These systems are probabilistic by design. Their behaviour can drift silently over time, not necessarily because the code changed, but because the environment around them did. A model trained on one operational reality can gradually become misaligned as network conditions, infrastructure patterns or physical environments evolve.
The dangerous part is that organizations may continue trusting decisions that are no longer aligned with operational reality because the system still appears to function normally. At the same time, these systems are increasingly deployed across distributed physical infrastructure: telecom towers, street cabinets, industrial facilities, transport infrastructure and remote operational sites. The attack surface is no longer centralized. And when these systems fail, the outcome is no longer just a cybersecurity issue. It becomes an operational event with real-world consequences: outages, autonomous misconfigurations, service instability or automated actions no human explicitly authorized.
This is where the governance gap becomes critical. Organizations are accelerating AI adoption faster than they are evolving governance. Autonomous capabilities are being integrated into operational workflows while accountability structures, oversight mechanisms and risk models remain largely unchanged.
That gap is where the next major incident will come from. The EU AI Act makes this shift even more significant because accountability does not stop with the technology vendor. Responsibility increasingly extends across the deployment chain, including integrators, operators and the organizations deploying these systems into production environments.
For organizations operating AI within critical infrastructure or high-impact operational contexts, this is no longer simply an engineering discussion. It is a governance issue with executive and board-level implications.
The uncomfortable reality is that many leadership teams still cannot answer a few relatively basic questions with confidence:
- Who is accountable when an autonomous system makes the wrong decision?
- Do we know which models are operating across our infrastructure today, and whether they still behave as intended?
- Can we demonstrate oversight, traceability and governance if regulators, customers or auditors ask for it?
Most organizations can answer those questions for traditional software environments. Far fewer can answer them for autonomous AI systems operating at the edge.
The next competitive phase of AI will not be defined solely by who has access to the most advanced models. It will be defined by which organizations can operationalize AI responsibly at scale while maintaining visibility, governance and control over increasingly autonomous environments. Because the perimeter already moved.
The next major AI incident is unlikely to begin in a chatbot or a cloud application. It will emerge at the edge, inside an autonomous system operating in the physical world, in an environment leadership assumed was still under control.
Physical AI is already making decisions inside your industry.
The question is whether leadership is still in the loop.
Written by Edgar Mateus
