If you read my last post, you’ll remember Gary, my nickname for ChatGPT and, unofficially, my extra brain. Gary helps me tackle everything from project documentation to planning vacations. But lately, I’ve been helping Gary understand more about the plant floor. That’s right, I’m teaching Gary.
You see, more and more people are eager to ask Gary big, important questions like: "Why did Plant A outperform Plant B last month?"
It's the kind of question that sounds deceptively simple until you realize it’s not. That one prompt packs in countless assumptions, missing context, and requires a level of domain awareness that, for decades, has only existed in the minds of seasoned plant professionals. People who could answer that question without batting an eye. People who are now retiring.
The good news? AI can help us preserve and scale that knowledge but only if we teach it how to think like a plant person.
From tribal knowledge to AI knowledge
For years, manufacturers have grappled with the so-called "skills gap” with the fear of losing critical operational knowledge as experienced personnel retire. The concern was always real but often felt intangible. Now, however, with the rapid adoption of AI, we’re at a true inflection point. The tools to preserve, structure and scale tribal knowledge finally exist. But it starts with data contextualization.
Large language models (LLMs) are incredible tools for interpreting natural language prompts. But they’re only as smart as the context they’re given. They don’t inherently know how your process works, what your KPIs are, or that “Area 3” means the packaging line on your west campus. We need to train them just like we would a new engineer shadowing operations for six months.
Reverse-engineering human intuition
Let’s go back to the question: Why did Plant A outperform Plant B?
Before an AI system can even begin to tackle that, we need to take a step back and ask ourselves: How would a human go about answering that?
To do this you need to think about issues such as:
1. What systems would they look at?
- They might check the historian for throughput data, MES for batch quality, ERP for order completion rates, maybe even LIMS for raw material consistency.
2. What context do they already know?
- Plant B just switched to a new raw material supplier.
- Plant A’s maintenance lead has a strong track record on proactive PMs. * There was a major thunderstorm that impacted Plant B’s power.
3. How do they phrase things?
- The CIP never hit temp, so we lost the batch. The extruder was acting up and kept tripping on torque. QA kicked back three pallets for label misalignment.
This is the human process of data curation and contextualization. We mentally stitch together disparate data, account for anomalies and interpret results based on lived experience. If we can articulate that process, we can begin to teach AI how to mimic it. The data curation blueprint To do this effectively, we must:
- Map the systems involved: Identify which systems hold relevant information and how they interact (e.g., historian points, MES event records, ERP production schedules).
- Standardize semantics: A tank is a tank until it’s also a reactor or a vessel or a unit. Defining a common vocabulary, both machine-readable and human-understandable, is critical.
- Capture cause-and-effect relationships: Maintenance events that trigger downtime, raw material lots that impact yield, operator changes that influence performance. These relationships need to be mapped.
- Incorporate metadata and domain logic: Not just what happened, but why it mattered in the context of your process.
When we do this, we aren’t just improving data quality. We’re teaching the AI how to be look at a problem like an operator or a plant manager would.