Sunday Essay — Singapore Fills the Seats, Skips the Stopwatch
Singapore's SkillsFuture programme can tell you precisely how many adults enrolled in AI courses, completed them, and rated their trainers. It cannot tell you whether any of them do their jobs differently. This essay argues that measuring training inputs while ignoring task performance is a structural choice — and proposes tying a share of course funding to a single instrumented measure agreed with the employer before the course begins.
Singapore knows more about who sits in its adult training rooms than almost any country on earth, and on the question that matters most for AI it has chosen not to look past the door. It can tell you who enrolled, who finished, who rated the trainer, and who used their credit. It does not try to find out whether the person who learned to use a model on Tuesday does the job any differently on Thursday. I think that choice is a mistake, and an odd one for a state that measures almost everything else it pays for.
This is a country that built its workforce policy on data the way other countries build theirs on slogans. So the gap is worth taking seriously, because if Singapore cannot close the measurement loop on AI training, countries with thinner data and smaller budgets will struggle to do better.
Note from a written answer to Parliament
The cleanest evidence comes from the government's own replies. In February, the Ministry of Education answered a question from Ng Chee Meng, the NTUC secretary-general and MP for Jalan Kayu, on how many people were taking AI courses. About 1,600 AI-related courses are supported by SkillsFuture Singapore. Last year, 137,000 training places were taken up by more than 105,000 individuals. And then the sentence that frames everything else: "we do not set targets specifically on the number of workers who go for AI training."
Two months later the follow-up came from Dr Wan Rizal of Jalan Besar GRC, and it was a better question. He asked whether the ministry tracks two things for people who complete AI-related SkillsFuture programmes: employment in AI-related roles within six months, and median wage change within twelve months. The written reply of 7 April is worth reading in full, because it is honest about where the measurement stops.
For the long-form career-switch courses, the SkillsFuture Career Transition Programmes, SSG tracks placement. Of 8,000 learners who completed an ICT-related programme between June 2022 and March 2025, 44 percent had found new roles within six months, as of 30 September 2025. Wages are not tracked for that group because "there can be multiple factors that affect their wages." For shorter courses, the ones meant to sharpen people in the jobs they already hold, placement is generally not tracked at all. What is collected is the Training Quality and Outcomes Measurement survey, TRAQOM, which asks learners about programme quality and their perceived outcomes: improved job effectiveness, better work performance, taking on more responsibility.
Put those pieces together and the shape of the problem appears. The courses that most AI training money flows into are short and aimed at incumbents. Their outcome measure is the learner's own opinion of whether they got better. Productivity economics learned long ago to distrust exactly that instrument.
What firms told the Ministry of Manpower
The employer side has the same texture. In April, MOM published its inaugural report on AI adoption among firms, covering private establishments with at least ten employees. Some 28.5 percent had started on AI. Among those, 70.7 percent reported improved worker productivity. The report puts the share of firms integrating AI into core processes at 3.8 percent, with 18.9 percent redesigning roles and 6.2 percent reporting reduced headcount. Among smaller adopters, 46.6 percent were focusing on training.
Hold those numbers side by side. Seven in ten adopting firms say their people got more productive. Fewer than one in five say they changed what anybody's job is. That combination is possible, but it is also the classic signature of a self-reported gain: the tool is in people's hands, everyone feels faster, and nothing in the organisation has moved to capture it.
Workers say something more complicated. PwC's Hopes and Fears survey for Singapore, published on 12 October from 1,000 local respondents, found 74 percent had used AI at work in the past year and 72 percent said it improved the quality of their work. In the same sample, 50 percent said AI increased the complexity of their work and 46 percent said it added to their workload. Confidence in job security over the next twelve months fell to 69 percent, from 84 percent in 2024. Only 65 percent said they had access to the learning resources they needed.
So the worker's own ledger reads: better output, more complexity, more work, less security. An efficiency gain rarely feels like that from the inside. A job with a new tool bolted on, and nothing taken off, usually does.
This past week added two more readings, both in the same direction. The American Chamber of Commerce in Singapore's 2026 Manpower Survey, run with Gallup and released on 11 November, found 80 percent of member employers upskilling existing staff and 68 percent expecting AI to raise employee productivity, while the share citing a skills mismatch as a key manpower concern rose to 46 percent from 34 percent a year earlier. Employers are training harder and feeling less matched. The expectation of productivity is a forecast, and it is doing a lot of work in that sentence.
Note from the job-posting data
The sharpest reason to care about measurement comes from the Institute for Adult Learning. Its July 2026 analytics report with SUSS and the Burning Glass Institute compared roughly 15.3 million job postings in Singapore and the United States from 2020 to 2025, across seven roles in finance and manufacturing.
Singapore's skill requirements are turning over much faster. Prevalence-weighted skill churn ran at 5.4 percent against 2.3 percent in the US, about 2.4 times higher. Singapore's Data Analyst postings listed 3.1 fewer skills on average after 2022, and Compliance Officer postings 2.4 fewer, while American postings held steady. The share of entry-level postings fell by about one percentage point in Singapore against 0.4 points in the US. The authors also observe that explicit AI skill mentions fell in some roles after 2022, which they read as AI disappearing into domain tools rather than being listed as a separate skill.
That last finding matters more than it first appears. If AI shows up as a feature inside the risk platform, the audit tool, or the MES on the plant floor, then a generic AI-fluency course teaches the concept while the job changes in the tool. The course goes stale faster than the credential printed at the end of it. A country with 2.4 times the American churn rate needs a shorter feedback loop between what it teaches and what changed at the desk, not a longer list of courses.
The authors are careful to note their limits: postings show employer signalling, not hiring or actual skill use. Fair. But that is my point turned around. Even the best skills-analytics shop in the region is reduced to inferring task change from job advertisements, because the task itself goes unrecorded.
The targets are all inputs
Look at how each agency defines success. NTUC and the Institute of Banking and Finance, announced by Deputy Prime Minister Gan Kim Yong on 25 June, aim to give up to 100,000 finance professionals AI skills over three years, through an asynchronous e-learning course with IBF funding of up to 70 percent. The release names no completion target, no assessment design, no outcome metric. IMDA and AI Singapore's AIxTech programme, launched in Senior Minister of State Tan Kiat How's speech of 8 May, will upskill 40,000 tech professionals on tools that, in his words, "can already generate large parts of a codebase." NTUC's Skills Lab, launched alongside the new e2i centre at Ang Mo Kio Hub, trained close to 100 participants on its first run on 14 May, and pairs courses with a subsidy covering half of selected AI tool subscriptions, up to $500 a year for members aged 40 and above.
In accountancy, ACRA's Accounting Entities Survey, published on 11 November, reports 34,000 professionals and students enrolled in the sector's AI Fluency Programme. The same survey of 238 firms, covering 92 percent of sector revenue, shows the workforce grew by 579 positions to 22,362 in 2025, median first-year pay for audit assistants rose from $3,000 to $3,400, and 67 percent of firms named AI application, ethics and governance as their top skills priority for the next three years. Those are good numbers for the profession. None of them tells you whether the 34,000 do an audit differently.
Every one of these targets counts something the funder can see from its own systems: seats, enrolments, completions, subsidies paid. There is nothing cynical in that. A training bureaucracy measures what it can see without walking onto anyone's floor. But it means the whole national effort reports activity and calls it capability.
The solid arrows are where the data lives. The dotted ones are where the productivity is supposed to come from.
Note from my consulting years
I have seen this pattern before, well before generative AI. Through the 2000s and early 2010s, more than one large client I worked with reported its technology training in learner-days, because learner-days were what the learning management system could export and what the budget committee could compare year on year. The ERP rollout would go live, the learner-days would be impressive, and eighteen months later the finance team would still be keying journals into a spreadsheet and uploading them, because the month-end close had never been redesigned around the new system. The training was real. The skill was real. The work had not changed, so the skill had nowhere to land.
AI training without job redesign is the same arrangement at higher speed, and with a tool that, unlike the ERP, quietly does part of the work whether you were trained or not.
The paper that argues back
The strongest objection to my argument comes from one of the most cited studies in the field. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered rollout of a generative-AI assistant to 5,179 customer support agents. In Generative AI at Work, productivity measured as issues resolved per hour rose 14 percent on average and 34 percent for novice and low-skilled workers, with minimal effect on the most experienced. The authors also report signs of worker learning.
Taken at face value, that undercuts the case for courses. The novices were not sent on a course. They were handed a tool that embedded the tacit knowledge of the best agents, and they improved on the job. If the tool is the teacher, why is Singapore spending its credits on classrooms, and why should anyone care about measuring them?
I take the finding seriously, and I think it supports my position more than it undermines it. Look at what made the result possible. The firm already logged resolutions per hour for every agent, before and after. The workflow was narrow and repetitive, and the AI was built into it rather than offered as a separate chatbot. The gain was visible because the task was instrumented before the tool arrived. That is precisely the condition Singapore's short courses lack. A compliance officer at a mid-tier bank, a process engineer in Tuas, a manager at a logistics firm: none of them has a resolutions-per-hour counter, and none of the funded AI courses asks for one.
And the novice result cuts in a second direction. If the tool raises the floor for juniors and barely touches seniors, then the mid-career workers whom Singapore most wants to protect may gain least from generic training. Liang Kaixin of the Institute of Policy Studies, writing about her own part-time master's in AI, put the mood plainly: "As middle-aged middle managers, we often feel that we'd be the first on the chopping block." She also notes that on short courses, trainers' qualifications and certificate recognition can be unclear. A system that cannot show those managers a measured gain in their own work is asking them to trust a certificate.
The other large study points the same way from the opposite end. Anders Humlum and Emilie Vestergaard linked adoption surveys to Danish administrative records in NBER Working Paper 33777, revised in March 2026. Most employers in exposed occupations had launched chatbot initiatives, and workers reported productivity benefits. Yet earnings and recorded hours showed precise null effects, ruling out changes larger than 2 percent two years after ChatGPT's launch. The mechanism they describe is task reorganisation: new work in content generation, AI oversight and integration. Reported benefits, real adoption, no movement in pay. Denmark is not Singapore, and two years is short. But it is a warning against reading self-reports as returns.
What SSG already has, and what I would change
Here is the stake, stated flatly. Singapore should stop treating funded AI training for incumbent workers as a seat-filling exercise, and should tie a meaningful share of that funding to one instrumented task measure agreed with the employer before the course begins. Where the employer will not name a measure, the course should be funded as general literacy at a lower rate, and reported as such.
This is less radical than it sounds, because the machinery exists. SSG tightened its funding guidelines from 31 December 2025 for courses that build in-demand skills for a person's current job, the exact category where placement is not tracked. At the same conference, SSG estimated that about a quarter of work tasks can be significantly aided by AI. If a quarter of tasks are in scope, then a funded course can name which tasks, and the employer can record how long they took before and after. Turnaround on a KYC file. Time to close a set of bank confirmations. First-pass yield on a quality inspection report. Tickets resolved per shift.
The Company Training Committee structure, which already brings union, management and workers together on transformation plans, is the natural owner of that baseline. NTUC's new arrangement with KPMG, signed on 22 October and reported by People Matters from the two organisations' joint release, makes the consultancy an approved adviser under the Enterprise Workforce Transformation Package, whose job-redesign grant co-funds up to 70 percent for SMEs. That is the right instrument pointed at the right thing. The pity is that it sits beside the training targets as a separate channel, when it should be the condition that gives the training money something to measure against.
I would also ask SSG to publish TRAQOM's perceived-outcome scores next to whatever instrumented measures exist, course by course, so that the gap between what learners believe and what their work shows becomes a quality signal in its own right. A course whose graduates report large gains while their tasks show none is telling you something about the course.
The objection I hear most often in reply is that measurement burdens small firms, and that SMEs will simply stop sending people. It is a fair worry. MOM's own data shows smaller firms lag on adoption and cite lack of strategy. But a single task timing, recorded twice, is a smaller burden than the grant paperwork those firms already file. And the alternative, funding hundreds of thousands of seats on the strength of the learner's own opinion, is a large public bet placed without a scoreboard.
None of this disputes that Singapore's effort is serious, generous and well organised. It is. The ambition to make a whole workforce comfortable with these tools is right. My quarrel is with what the country has chosen to count, in an era when the skill decays faster than the certificate and the tool does part of the job regardless of who took the course.
For the audit senior on Monday
You finished the AI Fluency Programme in September with others from your batch, and the certificate is already in your HR file. On Monday you open the confirmations folder for the December year-end client, the one with forty-odd bank accounts across three entities. Before you start, note the time. Write down how long the matching took you last year, from the old working papers, and how long it takes you now with the tool your firm licensed in July. Note which steps the tool did, which ones you redid because the output was wrong, and which ones are new to you this year. Keep it in the file next to the certificate. When the partner asks at appraisal what the course did for you, you will be the only person in the room with a number that came from the work.
Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.