Surveys this year put AI use among Thai companies above forty percent, up sharply on the year before. The same surveys report that most of those companies struggle to get past a pilot. In manufacturing the reason is rarely the model. It is that stock lives in one system, costs in a spreadsheet, supplier history in email, and production schedules on a whiteboard, and no forecast can be better than the worst of those sources.
The order of work
Establish the record. Before any forecasting, a plant needs one answer to four questions: what is in stock, what was last paid for it, who supplied it and how they performed. In our engagement with K.C. Metalsheet in Khon Kaen, the first phase is exactly this: a requirements and data audit, then the purchase history, vendor pricing table, product metadata and stock normalisation that everything after depends on.
Automate the paper first. Supplier quotations, invoices and delivery notes arrive as email and PDF. Reading them into the ERP automatically, and reconciling them against the purchase order, removes the manual step that keeps the record stale. This is unglamorous work and it is where most of the value sits.
Then add intelligence. With a record that agrees with itself, the useful models follow quickly: purchase requests raised from a photograph of stock, suppliers scored on price, delivery accuracy, quality and lead time, stock forecasting from warehouse, production and seasonal data, and production scheduling against machine availability and delivery deadlines.
Train the people who run it. Standard operating procedures generated from the system, an onboarding assistant that answers "how do I issue a transfer order", and translation between Thai and English for supplier communication turn the system into something the plant runs, rather than something run for it.
What this means for the boardroom
The question to ask a vendor is not what the AI can predict. It is what the AI will be predicting from. A plant that consolidates its record first will get more from a modest model than a plant that buys the best model and points it at six disagreeing systems. Opsian's engagements start with the audit that answers where the data originates, which is why some of them end there, with a document the manufacturer owns.
