Let the agent draft the work order. Let the planner still approve it.

Manufacturing Dive ran a piece last week on agentic AI adoption in manufacturing, and the honest read is that most plants aren't ready for an agent to *act*. Machine data is missing, siloed, or inconsistent site-to-site. Deloitte's 2025 survey has 40% of manufacturers planning to invest in data analytics in the next two years, which is another way of saying the foundation isn't there yet.

That doesn't mean agents have no place on the floor right now. It means the near-term use case is narrower than the demos suggest.

The pattern that actually works

IIoT World described the shape of it well: an agent reads a predictive-maintenance signal, queries the CMMS for asset history, checks ERP for parts availability, finds a qualified technician on the schedule, and drafts a work order with all of that context attached. What normally takes a planner two weeks of chasing people takes about 30 seconds.

The agent doesn't dispatch anyone. It doesn't reserve parts. It doesn't change a schedule. It puts a fully-contexted draft in front of the planner, who approves, edits, or kills it.

That's the whole trick. The value isn't autonomy. It's the coordination loop collapsing from days to seconds while a human keeps the final say.

Why draft-and-approve is the defensible line

The moment you let an agent write to a production system — publish a setpoint, release a work order, reserve a spare — the security question changes. HiveMQ's piece on guardrails for agent access to a UNS frames it correctly: you stop asking "who can see this data" and start asking "who or what is allowed to act, under what constraints, and how do we prove it later." That's broker-level policy, a governance layer, and an audit trail per action. It's not impossible, but it's a real project, and you'd better have a reason to take it on.

Draft-and-approve sidesteps most of that. The agent has read access across the historian, CMMS, and ERP. The write goes through a human. The audit trail is the approval itself. You get most of the time savings without expanding your blast radius.

What to have in place before you try it

A few things we've learned building these:

- The tag dictionary matters more than the model. If your assets are named inconsistently across lines, the agent will confidently fetch history for the wrong pump. - Draft quality is a function of context, not cleverness. The useful agent is the one wired into asset history, BOMs, technician skills, and open POs. A smarter LLM on thinner data loses to a modest LLM on complete data. - Make the approval one click, with everything visible. If the planner has to open four tabs to verify what the agent proposed, you haven't saved them anything. - Log every draft, including the rejected ones. Rejections are the training signal for the next iteration.

The takeaway

Agentic manufacturing apps will get more autonomous. They should. But the version that pays for itself this year is boring: an agent that reads widely, drafts carefully, and asks a human to press the button. Start there. The autonomous version is easier to justify once the drafted-work-order version has a year of clean audit trail behind it.