Why AI pays off for some and not others

Two datasets published within a day of each other at the end of September describe the same thing from opposite ends of a company. Gallup found that the more different purposes a person gives AI, the more likely they are to report a productivity gain — 45% at one or two purposes, 90% at seven or more — and that only 36% of employees in AI-integrating organisations strongly agree their manager supports its use. Boston Consulting Group's 2026 Applied AI Index, presented the same day, found about half of companies now deliver some or significant value from AI, but only a small share are future-built, and those track value with clear measures. The gain follows a decision about where the tool is applied and who owns the result.

Gallup's page on AI and workplace productivity, updated on 30 September 2026, reports that a clear majority of employees in organisations that had implemented AI said it had a positive effect on their productivity and efficiency. The more useful figure is underneath. Among United States employees who use AI at work, the share reporting a positive effect rises with how many different work purposes they give it — from 45% at one or two to 90% at seven or more. Gallup reads the benefits as sitting at the level of individual tasks, not whole systems. Leaders are likelier than everyone else to feel the gain, and they use it far more often than managers and individual contributors.

Gallup also breaks the gain down by what people use AI for, and the ordering is worth reading before anyone buys a licence on the strength of an average. Coding assistance and automation score highest, then presentations and data work. Writing and editing sit lower, and search or research lower still — the two uses most people start with, and the two where a wrong answer is hardest to notice. The gain tracks the task, so a team rolling out one assistant should expect an average. An average across tasks tells you almost nothing about the task you actually have in front of you.

Boston Consulting Group's 2026 Applied AI Index, presented on 30 September 2026 from a survey of C-level and senior respondents, is the more encouraging of the two: about half of respondents now deliver some or significant value from AI, with a small share classified as future-built and a larger group merely scaling. What separates the leaders is unglamorous: they use clear measures, track the value, and actively manage what AI costs. The warning in the same data is that very few companies have a full set of AI controls in place, while many expect autonomous agents within a few years. We would rather copy the boring part than the headline.

Gallup measures one organisational condition more sharply than the rest. Manager support is one of the two factors most strongly associated with frequent AI use, alongside whether AI is integrated into the workflow, and in 2026 only 36% of employees in AI-integrating organisations strongly agreed their manager actively supports their team's use of it. Where they do agree, they use AI far more often and are much likelier to say it has transformed how work gets done. A tool roll-out cannot produce that. It is the cheapest intervention on this list and the one most often skipped. That is a management task, not a software one.

Singapore's local evidence is about adoption, not breadth of use, and we will not stretch it. Answering parliamentary questions on 9 September 2026, Minister for Trade and Industry Tan See Leng said around three in ten firms had adopted AI, and that among them about seven in ten reported improvements in worker productivity. He added that more firms are redesigning roles and creating AI-related jobs than reducing headcount. The size split is where the local gap shows: smaller firms have adopted AI far less than large ones.

The finding we would act on is the spread from 45% to 90%, because it is a statement about how a person is set up rather than which assistant they were sold. It also explains the argument at every team meeting: one colleague says AI saves them a day a week, another that it is a faster way to write something they have to rewrite. Both are reporting honestly; they are using it for a different number of jobs. Breadth is itself a vanity metric — using AI for seven purposes badly is not a gain — so we would count the net after checking, not the number of uses.

As for the half of companies now seeing value, we hold it lightly: these are self-reports from senior leaders, in an industry whose earlier self-reports produced the pilot-to-pilot cycle everyone is trying to leave behind. The part we trust is the boring part — named measures, a manager who uses the thing, and one programme instead of eleven experiments.