Malaysia’s New Industrial Master Plan 2030 sets a minimum target of 3,000 smart factories by the end of the decade. The Ministry of Investment, Trade and Industry has repeated it often, and MIDF has put a rough price on it: about RM5 million per company, or RM15 billion in total.
The distance to that number is large. Reporting on the Smart Tech-Up programme described 21 fully operational smart factories identified so far, with a first-phase aim of certifying at least 40. One industry voice quoted by MIDA said plainly that he was not optimistic about reaching 3,000.
None of that means the direction is wrong. It means most factories are not going to get there by buying a smart factory. They will get there one workflow at a time.
The gap is not machinery
Walk through a mid-sized Malaysian plant and the machines are usually fine. What is missing is the record of what the machines and the people did. Test results are written on a form. The form is typed into a spreadsheet at the end of the shift, or the end of the week. A certificate is assembled by someone in the office who was not there when the test happened.
AI cannot help with data that does not exist. A vision model, an anomaly detector or a forecasting tool all need the same thing first: a reliable digital record, per unit, created at the point of work.
That is why the highest-return first project in most factories is unglamorous. Capture the record properly, once.
What “capture it once” looks like
We built a quality control platform for a water reticulation products manufacturer. The whole workflow hangs off a QR code on the product:
- A worker scans the code with a phone camera and the product record opens.
- They enter the pressure applied, pass or fail, a remark, and photos if needed.
- A product can be retested. The latest result is the one that counts.
- Only a product whose latest test passed can be added to a delivery order. The server enforces this, so no one can ship around it.
- The test certificate for that delivery order is generated in under a second, with a QR code that opens a public verification page.
More than 20,000 products and delivery orders have been traced through it. The certificate that used to take manual effort to prepare is now a by-product of doing the test.
Where AI fits after that
Once every unit has a clean record, the AI projects become cheap and low-risk, because they are reading structured data instead of guessing from paper.
Reading incoming documents. Supplier mill certificates, purchase orders and authority letters arrive as PDFs and scans. A language model can extract the fields into your system for a clerk to confirm. This is the same technique we use to turn unstructured job postings into structured records for a healthcare client, where it cut manual entry and review time by 76%.
Catching drift before failure. If a product type normally passes at a certain pressure margin and a batch starts passing by less, that is visible in the data long before anything fails. This does not need deep learning. It needs the history.
Answering the purchaser. “Which delivery order did unit X ship on, and where is its certificate?” is a lookup when the chain from product to test to delivery to certificate is linked.
Visual inspection. Camera-based inspection is real and works, but it is the fourth project, not the first. It needs labelled examples, and the labels come from the QC records.
What this means if you run a factory
Start with the document that leaves the building. Whatever your customer or the authority asks for, trace it backwards to where the information is first created, and capture it there.
Insist that rules live on the server. A rule that only exists as a greyed-out button will be bypassed the first time someone is in a hurry.
Make the proof verifiable by an outsider. A certificate with a QR code that anyone can check is worth more to a purchaser than one with three signatures.
Then ask what AI could do with the data you now have. The answer will be more specific, and much cheaper, than it was before.