It’s 2:40 a.m. at a packaging plant, and a filler on Line 3 has faulted for the third time this week. The technician on call, let’s call him Marcus, has been at the site for eight months. The veteran who used to handle this machine retired in the spring.
Marcus pulls up the plant’s new artificial intelligence (AI) assistant on a tablet and types in the question. Within seconds he has a clean, confident answer: the lockout sequence, the steps to clear the jam, the torque value for reassembly. It reads like it was written by someone who knows the machine. In a sense, it was.
What the answer doesn’t tell him is that the torque specification changed six weeks ago, after the filler heads were upgraded. Engineering revised the procedure, and the updated document even made it into the assistant’s library. So did the old one, along with the original equipment manual and a folder of shift notes from 2022. The assistant pulled from the wrong source, and nothing on the screen gave Marcus a reason to doubt it.
The scenario is illustrative, but anyone who has worked in operations will recognize the pieces. It raises the question that automation teams need to answer before they scale AI guidance on the plant floor. The issue isn’t whether the model is smart enough. It’s whether the knowledge behind it is governed well enough to trust.
The answer that looks right
Most conversations about AI risk focus on hallucination, where a model invents something with no basis. Marcus’s problem is different, and in industrial settings it may be more common. The assistant didn’t make anything up. It gave him an answer grounded in a real document that used to be correct.
That’s what makes this kind of drift so hard to catch. The answer uses the right terminology, names the right equipment and matches how the machine was configured six months ago. Operational knowledge changes constantly and in small pieces: a retrofit, a supplier part change, a lockout step revised after a near miss. If AI guidance isn’t tied to the same change process as the procedures behind it, the two quietly drift apart.
Controls engineers already know how to handle this kind of problem. No one would push a logic change to a running programmable logic controller (PLC) without review, testing, documentation and a way to roll back. Management of change exists because undocumented edits to live systems come back at the worst possible moment. The knowledge behind AI guidance deserves the same treatment, which is what a governed operational knowledge layer provides.
Replaying the night shift
Go back to Marcus at 2:40 a.m. and picture the same moment with the governed operational knowledge layer in place.
The answer on his tablet now points to its source: the revised procedure, version 4, approved six weeks ago by the maintenance engineering lead. If he wants to check, he can open it. Traceability works in the other direction, too. When engineering revised that procedure, the organization could see every piece of guidance, training and job aid that depended on the old version.
The old torque specification is no longer there to compete with the new one. When version 4 was approved, version 3 was retired from active guidance, the same way a superseded program comes off a controller instead of running beside its replacement. The 2022 shift notes and the original manual may still hold useful insight, but they don’t become operational guidance until someone has reviewed them.
That review is the next safeguard. Loading a document into an AI system isn’t the same as approving it. The subject-matter expert who knows the filler, the safety owner and the site that runs it should sign off first. AI can speed this up by drafting a response—flagging where two documents conflict and highlighting what changed—but the approval belongs to a person with the authority to give it.
And if something still looks off to Marcus, he knows where to take it. His flag goes to the named owner of that procedure, someone who can fix the source, rather than disappearing into a chat log. AI can generate, summarize, translate and check. It can’t be accountable. People can, and someone who is accountable should be assigned to every category of operational knowledge.
Same night, same fault, same technician. The difference is that the answer on the screen is one that the organization stands behind.
Governance is what makes speed possible
It’s easy to see review gates and version control as friction that will slow AI adoption. In practice, they tend to do the opposite. Without governance, every procedure change sets off a scavenger hunt. Someone has to find each training module, job aid and assessment that references the old version, revise them one by one and hope nothing was missed. With one governed source feeding all of them, a single approved change reaches everywhere it’s needed.
That idea is at the center of Cicero Orchestrator, the governed AI knowledge platform we recently introduced at CGS Immersive. In one deployment, an organization compressed a workflow-update process that had taken roughly 150 human days of effort into a 3-day cycle, with expert and compliance review still in place. The speed didn’t come from skipping oversight. It came from reviewing knowledge once, at the source, instead of rebuilding and revalidating it in a dozen places.
Remember the veteran who retired
The detail in Marcus’s story that should concern plant leaders most isn’t the torque specification. It’s the retirement. Much of what keeps a plant running lives in the judgment of experienced technicians: how to handle an exception and what an alarm really means on a particular line. Many organizations are working hard to capture that knowledge before it walks out the door, and that’s the right instinct. But capture is only half the job. Once expertise is captured, it needs an owner, a review process and a version history. Otherwise, it becomes one more source of confident, unverified answers.
Where to start
Teams don’t need to govern everything at once. Pick one high-consequence area, such as lockout/tagout, changeovers or troubleshooting on a critical asset, and walk through your own version of Marcus’s night:
- If a technician asked the AI about this tonight, could they see which approved source the answer came from?
- If the procedure changed tomorrow, how would every downstream use of it find out?
- Who owns this knowledge, and who can approve a change to it?
- When someone says the guidance is wrong, where does that report go?
If an answer is unclear, that’s a gap to close before AI guidance scales.
Automation professionals have spent decades learning how to change live systems safely. AI-generated guidance is fast becoming part of how the work gets done, and it should be held to the same standard. The next Marcus is going to trust what’s on the screen. Our job is to make sure it deserves that trust.
