Operational technology is the layer of systems that touches the physical world directly: the controllers running a factory line, the sensors and actuators on a water treatment plant, the safety-instrumented systems that shut down equipment before it fails. IT manages data, applications, and information flow. OT manages machines, and a mistake in OT can mean a physical consequence, not just a corrupted record.
That difference is why OT has historically been kept isolated from IT networks, and why the gap matters enormously once AI systems start operating across both. An AI agent with write access into IT is, worst case, a data or availability problem. The same agent with write access into OT is a safety problem, because a wrong or manipulated action can move a valve, halt a process, or override a physical safeguard. Standards like IEC 62443 formalise this as zone and conduit modelling: OT and IT are segmented into separate trust zones, with tightly controlled conduits between them.
The practical rule for agentic systems is that this boundary should stay absolute for autonomous write access. An agent can read OT telemetry, analyse it, and prepare a proposed change, but the actual write into a safety-instrumented system should go through existing change control and a human, not the agent’s own authority. Convenience is never a good enough reason to let a model close that gap.