Executive summary
3
standards governing a machine's safety-related control system, none written with a model in mind
2
separate loops that a well-integrated AI-assisted machine runs, kept apart by design
Key conclusions
- A certified safety function must behave the same way every time, and this must be verifiable. A model's inference time and output are neither fixed nor fully specified in advance.
- IEC 61508, ISO 13849-1 and IEC 62061 all require the safety-related control system to be independently verifiable. This is why the AI has to sit outside it.
- The fix is architectural. Hardwired interlocks and emergency stops (E-stops) stay inside the certified boundary. The AI stays supervisory and reaches an actuator only through a rated safety controller.
When people ask what stops AI from running machinery, the usual answer is a data problem or a model problem. In practice, it is a boundary problem. The machine's safety function and the AI's decision loop are governed by different engineering disciplines and run on different timing guarantees. Most integration proposals never define the boundary between them.
Defining that boundary is the real challenge. If it is drawn correctly, an AI layer can work alongside a machine and do useful work. If it is drawn wrongly, for example by letting a model's output reach an actuator directly, the result cannot be certified, insured or defended after an incident.
Safety systems need a fixed response time that AI inference cannot guarantee
A safety-rated control system is the circuit that stops a press, drops a guard or cuts power to a conveyor. It must respond within a bounded, tested time once a hazard is detected. Certification verifies that response time: fixed logic, on fixed hardware, with a known worst case.
Model inference cannot offer that guarantee. Even a small model returns its answer within a range of times. The range widens under load, on a cold start, or when the input is unlike anything the model has seen before. A control loop built to a certified response time cannot wait for an answer whose timing is known only as a range.
Three safety standards require behaviour that a model cannot guarantee on its own
Three standards cover most machinery in this position, and none of them was written with AI models in mind:
- IEC 61508 sets the Safety Integrity Level for the electrical, electronic and programmable systems performing a safety function, verified against defined failure modes.
- ISO 13849-1 sets the Performance Level for the safety-related parts of a machine's control system. Most mechanical guarding and interlock circuits are certified against it.
- IEC 62061 covers the same ground as ISO 13849-1 for more complex, programmable safety systems, where the logic runs in software instead of a hardwired relay.
- All three require the safety function's behaviour to be fully specified and independently testable. A model's output is drawn from a probability distribution, so a model cannot meet that requirement on its own.
AI belongs in a separate advisory loop
None of this excludes AI from mechanical systems. It excludes one placement: inside the certified loop, with the model's output wired directly to an actuator. Everywhere else, the model does valuable work without having to meet a certification standard that its design cannot meet. Examples include flagging an anomaly, recommending a maintenance window, or adjusting a feed rate within limits a person has already approved.
In practice, this means two loops. The safety-rated loop stays hardwired: E-stops, interlocks and light curtains, running on logic that was verified once and does not change. The AI runs a separate advisory loop alongside it. Any action the AI proposes passes through the rated safety controller, and the controller remains the certified component.
Legacy equipment needs a new interface, with its own certification
Older equipment often has no interface through which the AI can read the machine's state or request an action, safety-rated or otherwise. This is the same gap we cover in our article on retrofitting IoT onto legacy plant equipment. Adding an interface is a small project in its own right, with its own certification and sign-off, and it must be finished before the model sees live data. Teams that skip this step to start a pilot sooner can end up giving a supervisory system more authority over the machine than anyone approved.
Four checks to make before commissioning an AI layer
- Ask the integrator to draw the boundary on paper: which functions stay hardwired and certified, and which run in the advisory layer.
- Confirm the model's output reaches the machine only through a rated safety controller, never through a direct command path.
- Check the response-time budget the certified loop was verified against, and confirm nothing in the new architecture asks it to wait on an inference.
- Where the asset is legacy equipment, treat the interface itself (the sensor, the relay, the wiring) as its own project with its own sign-off.
That boundary usually decides whether an AI layer on a machine passes its first safety audit. The model's accuracy becomes relevant much later.
