Let the agent draft the work order. Keep the planner on approve.

Trade press on agentic AI in manufacturing has moved past the chatbot-over-your-historian pitch. IIoT World's recent piece on AI agents in MES frames the shift bluntly: the interesting work is not asking questions, it is agents that reason over MES data and take action. That is also where the risk shows up.

In our experience, the pattern that actually survives contact with a plant floor is narrower than the demos suggest. The agent drafts. A human releases. That is it. Everything else is a variation on where the draft comes from and how fast the human can say yes.

The draft-then-approve loop

Start with the CMMS. A reliability planner's day is a queue of judgment calls dressed up as data entry: which notification becomes a work order, which trade, which priority, which spare, which window. An agent sitting on top of the historian, the CMMS, and whatever condition-monitoring stack you run (Augury, AVEVA PI, a homegrown vibration pipeline, pick your poison) can draft a credible work order from a triggering event. Asset tag, failure mode code, suggested craft, estimated hours, parts to stage, a two-sentence narrative citing the tags it looked at.

The planner reads it, fixes what is wrong, and clicks approve. If the draft is good eight times out of ten, you have given that planner back half a day. If it is good three times out of ten, you have given them a worse inbox. The ratio is the whole game.

Augury's writeup on agentic AI in manufacturing makes a similar point from the reliability side: the value shows up when the agent closes the loop between a detected anomaly and the maintenance system, not when it answers a question about one.

Guardrails before write access

The moment an agent can write to a production system, the design problem changes. A few rules we have not seen anyone regret:

- No autonomous release to any system of record. Work orders, quality holds, recipe changes, schedule moves — all drafted, none released, without a named human. - Scoped credentials. The agent's CMMS user can create and modify drafts in a specific status. It cannot approve, close, or touch cost centers. - An append-only audit trail of every tool call, prompt, and payload. If a bad work order goes out, you need to reconstruct why in minutes, not days. - A kill switch the shift supervisor controls, not a ticket to IT.

None of this is exotic. It is the same discipline you would apply to any integration with write access to SAP PM or Maximo. The agent just makes it easier to forget.

Operator screens are not analyst screens

The other mistake we see is one interface for everyone. The operator at the press needs three buttons and a reason code. The reliability engineer wants to interrogate the tags behind the suggestion. The planner wants a queue. Same agent underneath, three different front doors. Build it that way or watch adoption stall on the floor while the office loves it.

If you are scoping an agentic project on your MES or CMMS this quarter, pick one workflow where a human already approves something today. Put the agent in front of that approval, not in place of it. That is the boring version. It is also the one that ships.