Your tag dictionary is the ceiling on any plant-floor agent

Manufacturing Dive reported recently that agentic AI is moving from pilots into production on plant floors, but the same infrastructure gaps keep showing up: fragmented historian, MES and ERP data, inconsistent tag naming, and unclear data ownership. That matches what we see. The model isn't the hard part anymore. The dictionary is.

Why historian chat is not document RAG

Document RAG works because a PDF has structure a language model can lean on. Headings, sentences, paragraphs. Chunk it, embed it, retrieve it, answer.

A process historian has none of that. A tag called `TT_4471_PV` is a float and a timestamp. The model doesn't know it's the melt temperature on extruder 3, sampled every second, in degrees C, with a deadband of 0.5 and a quality code you have to filter before you average anything. It doesn't know that downtime on this line is state code 3, not 0, or that the shift changeover at 06:00 always produces a two-minute gap that isn't a fault.

Ask it "why did we lose yield on line 3 last night" and it will happily hallucinate a plausible-sounding answer against tag names that don't exist. The demo looks great because somebody hand-mapped twelve tags for the demo.

The wrapper problem

ARC Advisory Group's writeup on the industrial copilot landscape is worth reading in full. Their point, after dozens of vendor briefings: most of what's being sold as an industrial copilot is a thin wrapper over a general-purpose LLM, with no real integration to OT data or workflow systems. The ones doing serious work spent most of their engineering budget on the plumbing — tag semantics, unit normalization, joining a work order in the CMMS to the batch in the MES to the tag range in the historian to the operator on shift.

That plumbing is unglamorous and it's where the value lives.

Guardrails when the agent can touch things

Read-only is a good place to start and a fine place to stay for a while. A shift supervisor asking "what were the top three downtime causes on press 2 this week" and getting a correct answer in ten seconds beats another dashboard nobody opens.

When you do give an agent write access, keep it boxed. Draft a work order in the CMMS, don't dispatch it. Propose a setpoint change, don't push it. Every action reversible, every action logged, a named human on the approval. The agent's job is to save the planner twenty minutes, not to replace the planner's judgment.

Where to start

Before you scope an agent, scope a tag dictionary. Pick one line. Write down what each tag means, its units, its quality rules, and how it joins to the work order and the batch. If that document doesn't exist, no model is going to invent it for you.

That's the boring work. It's also the ceiling on everything you build next.

If you want a second set of eyes on whether your data is ready for an agentic layer, our discovery call is free. No pitch, just diagnosis.