Anthropic, the company behind Claude, has published some striking figures. Code commits per engineer rose to roughly five times their December 2025 level within six months. By August 2026, employee-directed AI agents were accumulating six hours of runtime for every assumed human working hour. Meanwhile, the proportion of junior hires in technical departments fell from 27% to 9%. [1]
These figures certainly deserve attention. However, I would be cautious about treating them as a benchmark for the average UK business.
Anthropic is not a typical business. It develops frontier AI, employs highly technical people and has good reasons to redesign work around its own systems. More code commits do not necessarily mean more useful or reliable software. Agent runtime is not equivalent to human labour. Nor does a lower proportion of junior hires mean recruitment stopped: the technical workforce was still expanding in absolute terms. [1]
Even with those qualifications, I think the findings matter. AI may change not only how long tasks take, but the way organisations divide up, manage and review the work itself.
The UK picture: rapid take-up, limited transformation
ONS analysis published in July 2026 found that about 35% of UK businesses with at least ten employees reported using an AI technology, up from roughly 12% in late 2023. Yet the average number of technologies used by adopting businesses had risen only modestly, from approximately 1.4 to 1.6. [2]
A separate government survey reported AI use among 51% of small businesses with 10–49 employees that handled digitised data. Yet only 21% of AI-using businesses in that survey said their tools were integrated into existing business systems. Just 5% had a formal written AI policy. The surveys cover different populations and definitions, so their headline percentages should not be compared directly. [3]
This is a useful distinction. An employee drafting correspondence with an AI assistant may complete that particular task more quickly. A business that rethinks how customer enquiries are qualified, priced, approved and delivered is going further. It is changing the process, not simply accelerating an individual activity.
The productivity paradox
Government-commissioned AI Adoption Research, based on interviews conducted in 2025, found that 75% of businesses already using AI reported improved workforce productivity, 57% better processes or operations and 34% lower operating or production costs. Yet 77% reported no change in revenue since adopting AI, whilst 12% reported an increase. [4]
These are self-reported findings, not evidence that AI caused any particular financial result. Revenue is not the only measure of success, of course. Even so, the difference between reported productivity and business outcomes raises an important question: what actually happens to the time being saved?
Consider a hypothetical business producing 60 quotations each week. AI-assisted drafting cuts preparation time from 25 minutes to 15 minutes, apparently releasing ten hours a week. That sounds worthwhile, but the result depends on what happens next. Those hours may be absorbed by checking output, correcting mistakes or waiting for more work. Alternatively, the business may be able to prepare more accurate quotations, respond to customers sooner and win additional orders. The commercial value of the same ten hours could therefore be very different.
Time saved is an activity measure. Cost per completed transaction, service quality, productive use of released capacity and gross-margin improvement help show whether the business has actually benefited.
The hardest problems are often not technical
UK businesses have faced longstanding problems with productivity, investment, management practice and fragmented technology. In its 2023 research, the ONS identified difficulty finding worthwhile AI use cases, cost and a lack of expertise as leading barriers to adoption. Firms with stronger management practices were more likely to follow through on adoption plans. This is an association rather than proof that better management, on its own, causes successful AI adoption. [5]
An AI assistant is unlikely to achieve very much if the underlying business processes are inconsistent, customer information is unreliable or responsibilities are unclear. It may even make existing mistakes happen faster, or allow sensitive information to reach systems that have not been approved for that purpose.
Before selecting an AI product, I would start with three fairly basic questions. What business problem needs solving? How does the work happen today? And what would meaningful improvement actually look like?
What leaders can do over the next 90 days
- Start with one process that matters: quotation preparation, customer onboarding, monthly reporting, service-desk triage or document review, for example.
- Establish a baseline: the volume of work, time taken, staff effort, errors, rework, customer outcomes and current cost.
- Understand the bottleneck before choosing technology. Some problems call for better process design or existing software rather than generative AI.
- Run a limited pilot with real users, approved information and clear human oversight. Include failure cases, not just demonstrations that work as intended.
- Assess the complete picture: tool and implementation costs, checking time, risk, capacity released and benefits actually realised.
There is also a longer-term question about skills. If AI takes on more of the routine execution, how will junior employees develop the judgement expected of tomorrow’s senior specialists? Anthropic’s recruitment figures make this a question about workforce development, not simply technology. [1]
The competitive advantage may be organisational
I suspect the advantage will not necessarily lie with businesses buying the most AI licences, running the most ambitious agent demonstrations or spending the most on training. It may lie with those that identify work worth improving, redesign it sensibly, protect customers and employees, and can demonstrate what has changed.
For founders and boards, there is a worthwhile opportunity here. But the disciplines needed to realise it are familiar: sound strategy, workable processes, reliable information, clear accountability and measured outcomes.
AI does not remove the need for these fundamentals. If anything, it makes them more important.
Sources and methodological notes
[1] Anthropic Institute, How employment is changing at Anthropic (16 October 2026). Source — Internal observational data; not directly representative of UK SMEs.
[2] ONS, Artificial intelligence in UK businesses: 2023 to 2026 (20 July 2026). Source — BICS analysis; businesses with 10+ employees.
[3] DSIT, UK Business Data Survey 2026 (18 June 2026). Source — Survey of businesses handling digitised data; not the same denominator as ONS BICS.
[4] DSIT, AI Adoption Research (28 January 2026, updated 13 February). Source — Self-reported benefits among AI adopters; 2025 fieldwork.
[5] ONS, Management practices and adoption of technology and AI in UK firms (24 March 2025). Source — Associational evidence; barriers refer to 2023.
