Adding AI to Operational Technology? Start Here.

AI is showing up on the plant floor: predictive maintenance, quality inspection with computer vision, energy optimization, and assistants that help technicians troubleshoot. The upside is real. So is the new risk, because now you're connecting a fast-moving technology to systems that were built to change slowly and fail safely.

Here's what I'd want in place before AI touches an OT environment.

1. Decide what the AI is allowed to do: advise, or act?

There's a big difference between a model that recommends a maintenance window and one that can change a setpoint. Start with "advise only," with a human approving any change to the physical process. Earn your way toward automation with evidence, not enthusiasm.

2. Keep the data path one-way where you can.

Most AI use cases need data out of OT, not commands into it. Design for that. Push data out through a segmented zone (often a DMZ between IT and OT), and be very deliberate about anything that flows back in. ISA/IEC 62443 zones and conduits are a good way to frame the conversation with your engineers.

3. Treat the model and its data as part of the control system.

If a model influences how equipment runs, then its training data, its updates, and its vendor are part of your safety story. Data poisoning or a bad model update can cause the same harm as a tampered controller; it's just harder to spot. Ask who can change the data and who can push a new model version.

4. Watch your vendors and remote access.

Many OT AI tools are delivered by vendors who want cloud connectivity and remote support. That's understandable; it's also a door. Know which vendor connections exist, who approved them, and how they're monitored and shut off.

5. Plan for the model being wrong.

Sensors drift, processes change, and models degrade. Define what "wrong" looks like, who gets alerted, and how operations falls back to manual control. If your incident response plan doesn't mention the AI system yet, it should.

The bottom line

AI in OT can pay for itself, but only if safety and availability stay first. Start with advise-only use cases, one-way data flows, and clear ownership, and you'll be in a strong position to expand. The OWASP AI Exchange is a good free reference for the AI-specific threats.

If you're planning AI in a plant, utility, or other operational environment, I offer a complimentary 30-minute discovery call for qualified organizations. Book a time here.

Dan Sorensen is a cybersecurity leader, Forbes Technology Council contributor, OWASP AI Exchange core team member, and fractional CISO.