An LLM Will Answer Your Question
That Is the Problem

Ask a chatbot whether a second palletizer gets Line 3 to next year’s volume and it’ll tell you, fluently and with a number, in seconds. It has no model of your line. It has read a great deal about palletizers and is producing the most plausible continuation of your sentence. You can’t tell that answer from a correct one by looking at it.

What it’s actually doing

A language model predicts text. Given your question, it produces what a well-informed answer to that question tends to look like, assembled from everything it has read. It’s a remarkable capability, and for a lot of tasks it’s the right tool.

But your line isn’t in its training data. It doesn’t know about the buffer between your filler and your capper, the changeover you lose forty minutes to on Tuesdays, or the fact that your palletizer is starved rather than slow. It has never seen your historian. When it returns “roughly a 12–15% throughput improvement,” that figure isn’t a calculation. It’s the kind of number that shows up in sentences like the one you asked for.

Why that’s worse than being wrong

A tool that fails visibly is manageable, because you see the error, stop, and find another way. Here the wrong answer looks just like a right one. It has the right shape, the right units and a confident tone, and often a plausible-sounding justification underneath.

So the usual defense, “we’d check it,” doesn’t work, because there’s nothing to check it against. Checking means comparing to something that knows your line, and if you had that, you wouldn’t have asked the chatbot.

In most contexts a confidently wrong answer costs you an afternoon. In a capital decision it costs you the decision, and you find out two years later when the line still can’t make the volume.

The question isn’t whether the model is clever. It’s whether anything it said was derived from your plant.

What we do instead, and where we do use AI

We offer an AI / MCP add-on that connects an assistant to your validated model. You can ask it questions in plain language and run scenarios conversationally, against a local model if you want one. Use ReliaSim® from Claude or ChatGPT shows how to connect one to the public reference models.

The difference is what produces the number:

We don’t use a language model to build the model. The model is built from your line event data (the stops and starts your historian is already recording), with distributions fitted per failure mode. An assistant helps you ask questions of it and read the answers back in plain language. It’s the interface, never the engine.

The test that separates them

One question is worth asking of any tool that offers to answer a what-if about your plant. What did it reproduce before it predicted?

A peer-reviewed WSC 2020 plant model matched a full year of measured OEE within one percentage point, and so did its ReliaSim rebuild. The claim isn’t about cleverness. It could have been falsified, and it wasn’t. A chatbot can’t make that kind of claim, because it didn’t reproduce your year. It only produced a sentence about it.

Where a language model helps

The add-on is for jobs like these:

• Asking the model questions without learning its interface. “Which failure mode costs the most on Line 2?” is a better way in than a menu.
• Reading results back in plain language, including the ones you didn’t think to ask for.
• Getting oriented in an unfamiliar field, such as what a buffer does, why micro-stops matter, or what OEE leaves out.
• The writing around the work: the summary, the business case, the email to your plant manager.

Every one of those sits on top of something that knows your line. So the assistant is an add-on to a validated model, not a product of its own.

What to ask a vendor

1. “What is this answer derived from?” If the honest answer is “a language model,” it’s a guess in a confident voice.
2. “What did it reproduce before it predicted?” A tool that can’t match a period you already measured hasn’t earned a question about one you haven’t.
3. “Would it give me the same answer twice?” A simulation run again on the same inputs behaves the same way, and text generation doesn’t.
4. “Can I see the assumptions?” A model has inputs you can argue with, and a sentence doesn’t.

Related: why predictive maintenance can’t answer what-if makes the same argument about a different tool. That tool knows your plant intimately and still can’t tell you what a change returns.