Singapore uses AI more than most places — the results are not following

Singapore's workers use AI about as much as anyone in the world — 56 per cent at least several times a week — but the results have not followed the habit. Only 15 per cent strongly believe AI will improve their job, and 71.5 per cent of firms have not adopted it at all. The gains that do show up cluster in work with a repeatable process and an output someone can check: coding, automation, research. The gap is not access to tools. It is method — one recurring task, a written definition of a good result, a prompt written as a procedure, and the time measured before and after so the gain is a fact rather than a feeling.

ADP Research's People at Work 2026 report found that about 23 per cent of workers in Singapore use AI nearly daily, and 56 per cent use it at least several times a week — ahead of the global figures. Microsoft's 2026 Work Trend Index, published in June, found 66 per cent of AI users in Singapore saying they produce work they could not have made a year ago, against 58 per cent globally, rising to 82 per cent among the most advanced users. It also found 88 per cent saying they remain responsible for the thinking.

So the habit is here, and it is not reckless. People are using these tools often, and they are not handing over judgement.

That same ADP research found only 15 per cent of Singapore workers strongly agreeing that AI will positively affect their job responsibilities over the next year. Globally, frequent users report higher engagement, but they are also four times more likely than non-users to feel less productive.

That combination — heavy use, real output, and doubt — is what work without a method looks like. If you cannot tell whether a tool helped, you are not measuring it, and if you are not measuring it, you cannot get better at it.

The Ministry of Manpower's first report on AI adoption among firms, released on 30 April 2026, is blunt: 71.5 per cent of firms had yet to adopt AI. Among the 28.5 per cent that had started, only 3.8 per cent were integrating AI into core processes; most were still planning (7.4 per cent) or piloting (6.0 per cent).

Adoption scales with size — 23.9 per cent among firms with fewer than 25 employees against 76.4 per cent among firms with more than 500 — and concentrates in information and communications (74.1 per cent), professional services (57.5 per cent) and financial and insurance services (56.4 per cent). The pattern is capability, not enthusiasm: the firms with someone whose job is to make a process work are the ones getting it into the work.

Gallup's Q2 2026 figures from the United States make the same point from the other direction. Reported organisational adoption rose to 47 per cent, from 41 per cent the quarter before, with 52 per cent of workers using AI in their role and 15 per cent using it daily. The most common uses were writing and editing (51 per cent), search and research (49 per cent) and general problem-solving (39 per cent) — knowledge support, in other words.

The largest reported productivity gains, though, came from coding and automation. Those are tasks with a defined input, a checkable output and a repeatable process. Writing a first draft is not: it speeds up the start of a task, not the end of it, and the judgement still lands on the person who has to sign it off.

Pick one recurring task rather than rolling out a tool. Write down what a good result looks like before you start. Write the prompt as a procedure — the brief, the constraints, what must not change — so it can be run again by someone else. Check the output for the things that go wrong quietly: facts, numbers, and the details that drift. Then time it, before and after, so the gain is a fact rather than a feeling.

That is the difference between the 66 per cent who say they produce work they could not have made a year ago and the 15 per cent who expect AI to improve their job. The first group has a way of working; the second has an assortment of tools.

Four markers worth watching. You can predict the quality of the output before you generate it. You stop redoing the same task from scratch. You can hand the prompt to a colleague and get a comparable result. And the time saved shows up in a number you recorded, not in an impression.

Teams get there faster than individuals, because the fifth marker is the real one: the prompt library lives somewhere other than one person's memory.

Our own view, after reading all four sources: the doubt is more interesting than the adoption numbers, and we do not think it is distrust. It is the absence of feedback — most people cannot tell whether AI helped them last week, so the belief never forms. That is also why we are sceptical of much of the AI-productivity training on offer in Singapore: it teaches the tools, and the tool is the easy part.

Two cautions on the evidence, because it is worth being straight about. The ADP and Gallup figures are self-reported, so both the enthusiasm and the doubt are softer than they look, and the Ministry of Manpower's report measures adoption, not outcomes. The strongest claim we would make from this data is the modest one: measured, repeated work improves, and everything else in this field is still being guessed at out loud.