Why AI needs manufacturing context

A general-purpose AI model knows what a work order is. It does not know your work orders: which machines can run them, which materials they consume, which defect reasons your quality team uses, what “late” means for a customer with a 48-hour window.

Without that context, AI in manufacturing produces plausible answers — the most dangerous kind.

Context is a model, not a prompt

The usual fix is to paste more information into a prompt. That works for a question; it does not work for a plant. What AI needs is a governed model it can rely on:

• Structure — plants, departments, work centres, machines, operators, calendars.
• Process — products, routings, bills of materials, parameters, controls.
• Reality — what was produced, by whom, on which machine, with which lot, with how much scrap and which stops.
• Rules — the actions that are allowed and what they record.

When those exist as entities with names and relationships, AI can work with them: suggest which entities a request refers to, write a query that respects the model, propose a screen bound to the right objects.

Answers you can trace

Context also makes answers verifiable. “Scrap on line 3 doubled this week” is only useful if you can see which work stages, which reasons and which lots are behind it. AI grounded in production history can show its sources — and a person can check them.

Where Metalya applies it today

In Metalya, AI features work inside the governed model: natural-language questions turned into charts over production data, suggestions of which entities a screen should bind to, drafts of widgets and queries for an expert to review. Composing whole experiences from a request is being developed on the same principle — always with human approval.

The model comes first. The intelligence comes from using it well.