Walk into any MES or ERP renewal conversation right now and within ten minutes someone on the vendor side will mention an “AI agent” or a “copilot.” It’s in the MES upgrade quote, it’s in the historian contract, it’s baked into the ERP renewal terms. The pitch is seductive: point the model at your data, ask it questions in plain English, get root-cause analysis or scheduling recommendations back. The demos work great. Then the pilot starts on your actual plant floor and stalls out somewhere between the historian and the data science sandbox, and it’s not because the model is bad.
It’s because nobody packaged the data for it.
This is the uncomfortable truth of the 2026 renewal cycle: the bottleneck was never going to be model quality. Large language models and the agent frameworks wrapped around them are commodity-grade good at reasoning over structured, well-labeled data. What they cannot do is figure out, on their own, that “Line3_Temp_PV” means the barrel temperature on extruder 3, that it’s only valid when a batch is running, that a value of -1 means the sensor is offline rather than actually negative, or that the tag got renamed eighteen months ago when someone re-templated the UDTs. Your Unified Namespace solved the plumbing problem — getting data from PLCs and SCADA up through MQTT Sparkplug B or OPC UA into a broker where anything can subscribe. It did not solve the meaning problem. And meaning is what an AI initiative actually runs on.
The dashboard mindset doesn’t transfer
Most UNS implementations were built, reasonably, to feed dashboards and alerting. A human looks at the number, has context from twenty years on the floor, and interprets it correctly even when the tag name is garbage and the units are ambiguous. That human-in-the-loop tolerance for sloppy metadata is exactly what breaks when you hand the same feed to an AI agent that has no plant history and no tribal knowledge to fall back on.
A dashboard forgives an undocumented unit conversion. An AI agent making a maintenance recommendation off that same field will confidently produce a wrong answer with no visible seam where the error crept in. That’s a worse failure mode than a stalled pilot — it’s a quietly wrong one.
So the practical shift plant IT teams need to make isn’t “get more data into the UNS.” Most plants already have plenty of data in the UNS. The shift is treating a defined, bounded slice of that namespace as a data product: something built, owned, versioned, and supported the way a software team supports an API, not the way a controls engineer maintains a tag list.
What actually makes something a data product
A data product isn’t a database view and it isn’t a wider export. It’s a contract. At minimum, that contract needs four things, and if any one of them is missing, don’t be surprised when the AI layer produces garbage or the pilot just quietly dies.
Schema
A stable, documented structure — field names, data types, allowed ranges — that doesn’t shift when someone re-tags a PLC or a controls engineer “cleans up” the UNS namespace. ISA-95 equipment hierarchy and ISA-88 batch models are the natural backbone here; they give you a shared vocabulary for site, area, line, cell, and unit that both your MES and any AI layer can reason against without guessing.
Semantics
What the schema means, not just what it’s shaped like. Units, valid states, what null actually represents, how a downtime code maps to a real-world condition. This is where Sparkplug B’s built-in metadata helps but doesn’t fully solve the problem — you still need a human-authored data dictionary or ontology layer sitting on top, because Sparkplug tells you the payload structure, not the business meaning.
Freshness SLA
An explicit, published statement of how current the data is and how it’s allowed to degrade. A yield calculation refreshed every few seconds and one refreshed nightly are not interchangeable for an agent trying to reason about “what’s happening right now” versus “what happened this week.” If your AI initiative is going to make or influence decisions, it needs to know how stale is too stale, and that number has to be enforced, not aspirational.
Ownership
A named accountable owner — not “IT” as an abstraction, but a role or person who is on the hook when the schema changes upstream and breaks every downstream consumer. Without an owner, data products rot the same way undocumented tag databases rot: silently, until something built on top of them breaks in production.
Versioning is the part everyone skips
Treat every data product like an API and version it like one. When you change a schema — add a field, rename a state, alter a unit — that’s a new version, published with a changelog, with the old version supported for a defined deprecation window. This sounds like unnecessary process overhead until the first time an AI agent silently starts misinterpreting a renamed field and nobody notices for a production shift or two. Versioning discipline is what turns “the data changed and broke the model” from a recurring fire drill into a routine, expected event that consuming systems can handle gracefully.
This is also where plant IT and controls engineering need a real governance conversation, not a courtesy heads-up. The controls team that owns the PLC program and the IT team that owns the AI integration are now, functionally, co-maintainers of a shared interface. That’s a new working relationship for a lot of plants, and it needs an actual owner and an actual change-approval process, not an informal Slack message when someone reconfigures a UDT.
What to actually do this renewal cycle
Don’t try to productize your whole UNS. Pick two or three high-value slices — OEE inputs for a bottleneck line, genealogy data for a quality-sensitive product family, energy consumption for a specific area — and build them out fully: schema, semantics, SLA, owner, version history. Get one real AI use case working end to end against a properly contracted data product before you let any vendor’s agent loose on the broader namespace.
That’s a smaller, slower start than the renewal pitch decks suggest. It’s also the difference between an AI pilot that produces a defensible result and one that quietly generates plausible-sounding nonsense off data nobody actually vetted. Vendors are not going to slow down and tell you this during a renewal negotiation. It’s on plant IT to build the interface layer that makes the AI feature you’re about to sign for actually usable — because the copilot in the contract is only as good as the data product standing behind it, and right now, for most plants, that product doesn’t exist yet.
This article was written with the assistance of artificial intelligence. While we aim for accuracy, the information may be incomplete, out of date, or incorrect, and should be independently verified before you rely on it for any decision. It is provided for general information only and does not constitute professional advice.
