Why AI use at work isn't saving time

Barclays said on 1 October 2026 that Claude Code will be in half its developers' hands by the end of 2026 and most of its engineers in 2027, adding to deployments that already answer colleagues' questions and sort client email. Every figure counts use, and none of them is a saving: the bank has published no time or money returned, and in a half when it reported about £350m of cost efficiency savings, group operating costs still rose 6%. Singapore's official reading is patient in the same way, and only 3.8% of firms surveyed by the Ministry of Manpower have integrated AI into their core operations. Our view is that a count of seats is not a result, and the industry is still publishing the first while calling it the second.

Barclays and Anthropic said on 1 October 2026 that the bank is extending Claude across its global operations — software development, legacy-system modernisation, operational efficiency — with one headline target: Claude Code reaching half of its developer population by the end of 2026, then a majority of software engineers in 2027. The rest of the release lists what is already running: a knowledge assistant on Claude since 2025 that answers colleagues' questions, and a platform in Global Markets that classifies and routes incoming client email. Anthropic's chief commercial officer, Paul Smith, described the tool helping colleagues find answers and sorting email, and the bank's group co-chief operating officer, Anne Marie Darling, said the true measure of any technology is the impact it has on customers, clients and colleagues. No such impact is in the announcement — no hours returned, no fall in engineering effort, no cost avoided. The figures measure how many hands the tool is in.

The gap sharpens beside the half-year numbers. Barclays is working to a multi-year efficiency target, and reported about £350m of cost efficiency savings in the first half of 2026 — a half in which group operating costs rose 6% on business growth, inflation and investment spend. The savings were real, and so was the growth they partly offset, and nothing in the AI announcement separates the two. Others reading the release noted the same thing: the bank has not quantified time saved in support or development, has not defined adoption as regular use rather than licences issued, and has not said whether it will publish the result. That is the shape of enterprise AI reporting in 2026 rather than a Barclays flaw. Counting seats is cheap and rises reliably. Counting rework avoided needs a baseline, months of measurement, and someone willing to be judged by it.

AMRO's 2026 annual consultation report on Singapore, reported on 30 September 2026, estimates a modest long-run gain to the country's potential output — a partial offset to an ageing workforce rather than a boom — and concludes that redesigning workflows, not buying tools, decides how much of it arrives. Its adoption figures show the familiar pattern: small and medium-sized firms' AI use climbing fast from a low base while larger firms sit further ahead, and only 3.8% of firms surveyed by the Ministry of Manpower having integrated AI into their core processes, most of the rest still at planning or pilot stage. AMRO also flags that the transition could squeeze some workers, fresh graduates among them. The cost side is being competed away meanwhile: on 1 October 2026 Singtel's RE:AI launched a Token-as-a-Service offering, hiring out AI models on sovereign infrastructure here on a token-based subscription. Cheaper tokens remove one excuse for not adopting. They do not touch the 3.8%.

Singapore also has the first local example of what happens when AI-assisted work ships without a reviewer. The Personal Data Protection Commission's voluntary undertaking, published on 21 September 2026, sets out the country's first notified AI-related data breach. A Bee Cheng Hiang employee asked a generative AI tool for a script to send a mass marketing email in batches, without saying that recipients should not see one another's addresses. A missing bracket grouped each batch into a single To: field, so the mailing went out with every address in a batch visible to everyone else in it. The employee tested by reading activity logs rather than the content of a test email. The commission found the tool had not malfunctioned; the prompt and the missing review had. Bee Cheng Hiang now requires two staff to verify bulk mailings before sending.

We would not plan a project around an adoption target like Barclays' 50%. A number stated in heads is a statement about procurement, and it is the number that can go into a press release on the day; the number that matters — hours, rework, cost per task — is the one nobody has yet. Barclays is a competent institution running a wide rollout with governance attached, and its own results show why the honest version is hard: costs still rose in a half when it found savings, and nobody can say how much of either came from Claude.

The Singapore figures are the ones we find genuinely useful, because they separate breadth from depth: a small share of firms touching AI, against a much smaller share that has rebuilt a core process around it. That describes most organisations we meet, including the ones that call themselves AI-first. AMRO's modest long-run gain is not a headline, and it is probably the right scale. We are also unconvinced that cheap tokens close the gap; falling prices mostly make it easier to spend more on the same undirected work.

The Bee Cheng Hiang case is the one we would put in front of a team, because it is the smallest version of the failure. Nobody had to be malicious or unusually careless: someone asked a tool for a script, trusted the output, and tested it by reading a log instead of the thing itself. Our rule is unglamorous and cheap. Open the actual output and look at it before it goes out, and give someone other than its author the job of saying yes.