Manufacturing has more job vacancies than any other industry in Taiwan, and the working-age population shrinks every year. When positions cannot be filled, the practical move is to remove the work that does not need a person's hands. This article lays out the latest official figures, then covers three gaps AI can take on, and what it cannot do.
01The numbers: the industry with the most vacancies, and a shrinking population
The Ministry of Labor's Job Vacancy Survey for end-June 2026 (ROC year 115, released September 30, 2026) counted 296,000 vacancies across industry and services. Manufacturing accounted for 111,000 of them, or 37.4%, the most of any industry. The manufacturing vacancy rate was 3.7%, above the overall 3.3%, and full-time vacancies took an average of 2.9 months to fill through external recruitment. By occupation, craft workers and plant and machine operators and assemblers made up 68,000 full-time vacancies, 25% of the total. These are the shop-floor operators and technicians.
The Ministry notes in the same release that vacancies measure hiring demand, which is not exactly the same as a labor shortage where demand exceeds supply. On the supply side, though, the direction is clear. Under the medium variant of the National Development Council's Population Projections for the R.O.C. (Taiwan): 2026-2075, the population aged 15 to 64 is 15.76 million in 2026 and falls to 14.16 million by 2035. From 2037, more than half of the working-age population will be aged 45 to 64.
On the factory floor, that means new hires are harder to find while experienced technicians keep retiring. Recruiting still matters. At the same time, it makes sense to take back the hours that technicians and office staff spend on work a system could handle.
02Gap 1: Turning master technicians' experience into something new hires can look up
In many plants, critical knowledge lives in the heads of a few senior people: what to watch for during a mold change, which parameter to adjust for a certain batch of material, what a particular machine noise means. New hires can only watch and ask, and when the technician is busy, learning stops.
The approach is to gather scattered documents into a knowledge base that people can query in plain Chinese. A new hire asks how many purges are needed when changing colors on an injection molding machine, and the system answers from the SOPs and incident records, with citations so the answer can be checked. Technicians can explain procedures verbally or on video; AI turns these into transcripts and draft procedures, which the technician then reviews and approves. The same material can become training content and quizzes.
What to prepare: existing SOPs, work instructions, incident records, quality inspection standards, and a few key processes the technicians are willing to talk through. Paper documents can simply be photographed or scanned.
The limits need to be stated up front. The knowledge base answers what the documents say. Feel, diagnosing problems by sound, and judging material condition still have to be taught on the floor. AI cuts the time spent searching and asking the same questions; it does not replace hands-on operation. If a document is wrong, the answer will be wrong too, so technicians should review the content on a regular schedule.
03Gap 2: Scheduling, daily reports, and inquiry replies with less repetitive work
Office and production-control staff spend a lot of time each day on fixed-format paperwork: turning orders into schedules, compiling line output into daily reports, and answering customer questions about prices and lead times. The rules are clear and the formats are fixed, which makes this work a good fit for a system to produce the first draft.
For example, production figures exported each day from existing Excel files or the ERP can generate a daily report and an exception summary automatically. When a customer emails about lead time or specifications, AI drafts a reply from product data and the current schedule, and sales confirms it before sending. For scheduling, the system produces a first pass based on due dates and machine capacity, and production control adjusts it.
What to prepare: product specification sheets, the rules used to decide prices and lead times, past schedules and report templates, and typical inquiry emails. The clearer the rules, the more usable the drafts.
Quotes, delivery commitments, and final schedules must be confirmed by a person. Rush orders, sudden machine breakdowns, and material delays are shop-floor variables the system only knows about if someone enters them.
04Gap 3: Stop having three departments type in the same data
Sales enters the order, production control enters it again for scheduling, then the warehouse and accounting each enter it once more for shipping and invoicing. This is common in traditional manufacturing. Every extra entry is another chance for error, and mismatched quantities or prices often surface only at month-end reconciliation.
Start by mapping an order from receipt to shipment: which forms and systems it passes through, and which fields get re-entered. Then designate one dataset as the source and pull other reports from it. Handwritten or PDF orders and delivery notes can be read by AI and converted into structured data, with people only checking the result.
What to prepare: the forms and Excel files each department uses, export formats from the ERP or inventory system, and mapping tables for part numbers and customer codes. If part numbers are named inconsistently, standardize them first, or the systems cannot be connected.
What usually holds this up is process and ownership rather than technology. Who maintains master data, and who must be told when a field changes, are decisions the owner or plant manager has to make. Tools only help a process that has already been agreed on run smoothly.
05Where to start
Do not try to do everything at once. Pick one task that repeats most often each week and draws the most complaints about errors, such as compiling daily reports or replying to inquiries. Build a usable version, let the people who actually do the work try it and give feedback, then decide the next step.
Before starting, answer three questions: who does this work now and how many hours a week it takes; where the needed data lives and whether it is complete; and who checks the result when something goes wrong. If you can answer them, you can begin. If you cannot, organizing the data is itself a worthwhile first step.
When Twouring helps manufacturers adopt AI, we start with this kind of review, then decide what goes to AI, what an ordinary system can handle, and what still needs people.
Sources
Figures and regulations in this article come from the official sources below. Please cite the originals.