If anyone asks a minister in any capital of a Southeast Asian country about how the country is planning to succeed in the Artificial Intelligence (AI) race, the answer will be in the same old order: data centers, chips investments, and engineers. But the people who will decide whether AI makes the region richer, or only faster at making mistakes, are not the builders. They are the loan officer, the hospital administrator, the secondary school teacher, and the marketing manager who approves a campaign at six in the evening because the draft looked polished.
The scale is bigger than most national strategies admit. In July this year, the International Labor Organization (ILO) predicted that 22.9 percent of employment in ASEAN countries — almost 80 million jobs — is in occupations subject to generative AI, software that writes, condenses, and analyses incidental language instructions. These are not mostly engineers. Women are more than twice as likely as men to hold the most exposed jobs, because they fill so many clerical and administrative posts. In the Philippines, ILO analysis puts exposure among clerical jobs at 93.7 percent. There are arithmetic travels here. Any economy whose growth runs through offices and back-end services, from Naypyidaw to Dili, should check its own version of it. Here is the detail that ought to bother planners. The same ILO brief finds actual use concentrated in technology-heavy occupations, with limited uptake in offices and administration. The people most exposed are using it the least.
But that will not last. It will end the way these things usually end with a software update. One Monday, the payroll system, the customer database, and the email client will all grow an assistant, and a worker who has never been shown how to question it is expected to use it in Southeast Asia. Workers feel that this is coming. In February, Milieu Insight conducted a survey amongst 3,000 professionals from Singapore, Malaysia, Thailand, Indonesia, the Philippines, and Vietnam.
AI training provided by employers was reported to reach between 14 percent to 29 percent, depending on the specific country. Between 34 and 53 percent of participating employees had received none and wanted some. Only 8 to 22 percent called their government very effective at preparing them. A separate Salesforce and YouGov survey of 4,062 ASEAN knowledge workers found just 32 percent were being trained to use AI agents, which are systems that carry out multi-step tasks with little supervision. Workers are asking. Employers are mostly not answering.
The lazy response of the workers is a prompt-writing workshop. I think that is largely a waste of money. Tools change every quarter, and a worker who has memorized one interface will be back at the beginning by Christmas. But what lasts is a habit: doubt, applied quickly, and without embarrassment.
What happened to Deloitte in Australia is pertinent to mention here. The firm delivered a 237-page report to the federal government for A$440,000. It contained a fabricated quote from a federal court judgment and references to academic papers that do not exist, the kind of confident invention that specialists call hallucination. A Sydney University law researcher, Chris Rudge, flagged it. Deloitte then disclosed that it had used generative AI and agreed to refund part of its fee. This is a firm with money, lawyers, and a reputation built on checking other people’s work. What it lacked, when it counted, was someone with both the habit and the standing to ask where a quotation came from. If that can happen there, it cannot be difficult to imagine what will happen at a district health office with two analysts and a vendor demonstration.
AI literacy means something specific and unglamorous: knowing what a tool is good at and where it reliably fails, what the user is pasting into it and who else might see it, and that someone’s name on the output means he/she answers for every line. A nurse reading an AI-drafted handover note, an analyst trusting a summary of forty pages, a teacher who cannot tell a fluent essay from a correct one and so marks the wrong thing. None of them needs to write code. All of them need to be hard to fool.
Universities go first, and the uncomfortable part is that many are behind their own students. Locating AI in the computer science department is a mistake. Every program, from nursing to marketing, should put students in front of these tools on real assignments and then grade the error hunt: what did the machine get wrong, how does he/she know, and what did he/she change? Institutions that ban the tools to protect their assessments are rather protecting the assessment and abandoning the graduate. A student who has never caught a machine bluffing in a safe setting will meet the first one at work, where it costs money.
Employers should stop counting certificates. A completion badge proves someone sat through a video. Organizations should look at whether staff can show a flawed AI output they caught and explain how they caught it. Employers should run the training in paid working hours, because a course taken at nine at night on a worker’s own time is a hobby. Organizations should make the disclosure of AI use in client work, and naming one accountable human on any decision an AI helped shape, mandatory.
Governments have the hardest and most useful job. Governments need to put public money first where exposure is highest, which means clerical and administrative workers, and design programs for the women who hold most of those posts, as the ILO urges, along with support for small and medium businesses that have no training department to speak of. Governments should make AI disclosure a condition of public contracts so that buyers know where to look hardest and publish outcome data on training rather than enrolment counts so that the region learns what actually works.
Some will say that literacy is a soft target next to semiconductor plants, and that economies compete on hardware. Hardware can be bought. The habit of asking “says who?” has to be taught to millions of people who will never write a line of code, and the region that teaches it first will make better decisions with the same tools everyone else is renting. That is slower and duller than cutting a ribbon at a data center, which is exactly why it keeps getting neglected.
DISCLAIMER: The views expressed are those of the author(s) and do not necessarily reflect the official policy or position of the Magazine, or its editorial team. The views published are the sole responsibility of the author(s).
