Anthropic Economic Index UK: what Claude data shows about work

News · 3 Jun 2026 · MTW Editorial Team

The Anthropic Economic Index is the clearest public read we have on how a frontier AI model is actually used at work, and its numbers carry real signals for British jobs, skills and small firms. Rather than survey opinions, Anthropic studied millions of anonymised Claude conversations and mapped them onto real occupations and tasks. The result is a rolling dataset that tells us less about what AI might do one day and more about what people are paying it to do right now.

Key facts: what the Anthropic Economic Index is and why it matters now

Anthropic published the first Economic Index on 10 February 2025, with an updated cut using Claude 3.7 Sonnet data on 27 March 2025 and a larger geographic and enterprise report on 15 September 2025. Each release samples real Claude usage, classifies the underlying tasks against the US Department of Labor O*NET occupational framework, and then asks a simple question: is the model doing the job for someone (automation) or doing it with them (augmentation). That distinction is the spine of the whole project, and it is the reason the data is more useful to a UK reader than another round of speculation about robots.

For British workers and owners of small firms, the value is practical. If you can see which tasks are already being handed to AI, and which ones still need a human in the loop, you can plan training, pricing and hiring around evidence instead of hype. We have read the reports in full so we can separate the headline figures from the marketing, and the sections below set out what the data says, where it is silent, and what we think it means for the UK. Anthropic is an American company reporting global usage, so the UK-specific reading here is our analysis, clearly labelled as such, not a figure lifted from the reports.

Automation versus augmentation: the headline split

The first report found that roughly 57% of Claude tasks looked like augmentation and 43% looked like automation. Augmentation covered patterns such as task iteration, learning and validation, where a person stayed involved and used the model to think, draft or check. Automation covered directive and feedback-loop patterns, where the model was handed a task and asked to complete it with little further input. The 3.7 Sonnet update kept augmentation at about 57% of usage, with the learning share rising from roughly 23% to around 28%, which suggests people increasingly reach for the model to understand something rather than just to offload it.

That split matters because automation and augmentation imply different things for jobs. A task that is fully automated can, in principle, be removed from a role. A task that is augmented tends to make the person faster or better rather than redundant. The headline reading, then, is reassuring on the surface: most current use leans towards augmentation. We would temper that, though, with the later finding that the picture shifts sharply once you move from consumer chat to business systems, which we cover below. For UK readers weighing how AI fits a knowledge-work role, our guide to the best AI writing assistant in the UK for 2026 walks through where these tools genuinely help with drafting versus where they still need close supervision.

Which occupations and tasks show up most

Usage is heavily concentrated. In the first report, Computer and Mathematical tasks accounted for 37.2% of Claude queries, dominated by software modification, code debugging and network troubleshooting. Arts, design, entertainment, sports and media came next at 10.3%, followed by education and library work at 9.3%, office and administrative support at 7.9%, life and physical sciences at 6.4%, and business and financial operations at 5.9%. By the September 2025 report, coding still dominated the overall sample at about 36%, so the centre of gravity has held steady.

The other important number is depth of use across the labour market. Anthropic found that roughly 36% of occupations showed some Claude use across at least a quarter of their tasks, while only about 4% of occupations used it across at least three-quarters of their tasks. In plain terms, AI is touching a wide range of jobs lightly and very few jobs comprehensively. That pattern fits what we see in British workplaces, where a solicitor or accountant might lean on a model for a handful of well-defined tasks while the bulk of the role stays human. Our practical walkthroughs for Claude for UK solicitors and Claude for UK accountants show exactly which tasks tend to fall into that quarter, and which should stay off the model entirely.

The UK in the global picture

The September 2025 report added country-level detail for the first time, using an Anthropic AI Usage Index that compares a country’s share of Claude use against its share of the working-age population. The United Kingdom came in at 2.67 times its expected usage, which places it among the heavier adopters globally but behind the leaders. Israel topped the list at about 7.0 times, with Singapore at 4.57, Australia at 4.10, New Zealand at 4.05, South Korea at 3.73 and the United States at 3.62. Large emerging economies sat far lower, with India at 0.27 and Indonesia at 0.36.

A UK score of 2.67 is a useful reality check. Britain is clearly engaged with frontier AI, well above the global average, but it is not at the frontier of adoption itself. The report also noted that higher-income, higher-usage countries tend to use AI more as a collaborator and lean less on full automation, which is our read on where the UK sits: more augmentation than directive hand-off, at least in consumer use. We would not over-read a single index, and Anthropic measures Claude alone rather than the whole market, so the figure understates total UK AI activity. Still, it is one of the few hard data points on British adoption, and it argues against both complacency and panic.

The gap between the UK and the leading adopters is partly cultural and partly structural. Israel and Singapore combine dense technology sectors with policy that actively encourages experimentation, while Britain’s adoption is spread more thinly across a larger, more cautious economy. For individuals, the lesson is that being an early, fluent user of these tools is still a differentiator here in a way it may not be in Tel Aviv or Singapore. The same logic applies at firm level, which is where the consumer-versus-enterprise gap becomes the most important number in the whole dataset.

The enterprise gap: where automation jumps

The single finding UK small firms should sit up for is the difference between casual chat use and programmatic business use. On Claude.ai, consumer automation rose from about 27% in late 2024 to roughly 39% by August 2025, still a minority of activity. But across enterprise API traffic, where companies wire the model directly into their workflows, automation accounted for about 77% of use, against roughly 50% for consumers. Computer and mathematical tasks made up around 44% of that API traffic, with office and administrative work the next largest slice at about 10%.

That 77% figure reframes the comforting headline. When you stop talking to a model in a chat box and start embedding it in a system, the centre of gravity swings hard towards automation. For a UK SME, this is the difference between an employee occasionally using Claude to draft an email and the business building a pipeline that handles invoices, support tickets or first-line code without a person in the loop. The productivity gains can be large, but so is the change to how work is organised. We have written about what large rollouts teach smaller teams in our look at Microsoft Copilot lessons from the Accenture rollout for UK enterprises, and the cost side is laid out in our coverage of the Microsoft 365 Copilot UK price rise on 1 July 2026, both of which sharpen the build-versus-buy question.

What this means for UK workers and skills

Read together, the data points to a clear personal strategy rather than a vague call to upskill. The tasks most exposed to automation are well-defined, repeatable and text-heavy: boilerplate code, routine drafting, data lookups, basic troubleshooting. The tasks that stay augmented involve judgement, context, client relationships and accountability. For a British worker, the sensible move is to push your time towards the second category and to become the person who supervises, edits and signs off AI output rather than the person who produces the first draft by hand.

The rising learning share, up from roughly 23% to around 28% in the 3.7 Sonnet data, is encouraging on this front. It suggests many people are using the model to build understanding, not just to dodge effort, and that is exactly the habit that keeps a worker on the augmentation side of the line. Practical fluency matters more than abstract awareness here. If you want to see what disciplined, sign-off-style use looks like in a regulated profession, the solicitor and accountant guides linked above are a better starting point than generic advice. And for anyone tempted to spend money chasing AI features, our argument that you should not buy a new phone for an Android AI upgrade alone is a reminder that the skill, not the hardware, is the asset.

What this means for UK small and medium firms

For SMEs, the enterprise automation figure is both an opportunity and a warning. The opportunity is that the same tasks consuming a disproportionate share of staff time, support triage, routine document work, first-pass code, are precisely the ones the data shows AI handling well. A small firm that automates the right slice can free skilled people for work that actually grows the business. The warning is that 77% automation in API use shows how quickly a model moves from helper to operator once it is embedded, and an operator needs oversight, error handling and clear accountability.

Our advice is to start where the data is strongest and the risk is lowest: well-bounded internal tasks with a human checkpoint before anything reaches a customer. Pilot one workflow, measure it honestly, and only then widen. The lessons British small firms drew from Anthropic’s own developer event are worth borrowing here, and we gathered them in Code with Claude 2026: five lessons for UK SMEs from the London keynote. If your team builds on Claude directly, the steady pricing in our note on Claude Opus 4.8 in the UK makes budgeting a pilot more predictable than it would have been a year ago.

Key takeaways at a glance

Those figures are the load-bearing ones, and they reward a second look before any business decision. Notice how the breadth-of-use line and the enterprise line pull in different directions: most jobs are touched only lightly, yet where firms commit to integration, the model takes over the majority of the work. Both can be true at once, and holding them together is the key to reading the index well rather than cherry-picking the comforting half.

Where to check next in the UK

This is a data story rather than a product launch, so the useful next steps are sources and tools rather than shops. Anthropic publishes each Economic Index report and the underlying dataset on its own research pages, and reading the September 2025 geographic report in full is the best way to see the UK figure in context. For the official model and pricing details a UK buyer would weigh before a pilot, the Anthropic website is the primary source, and our own pricing and feature coverage cross-checks it.

If you are comparing AI assistants before committing budget, John Lewis, Currys and Amazon UK all sell the hardware these tools run on, but the meaningful comparison is the subscription, not the laptop. Check each provider’s UK pricing, data-handling terms and any business-tier commitments directly on the provider site, confirm where your data is processed, and note the contract length before you sign. For Microsoft’s stack, the price and rollout pieces linked above set out what to verify; for Anthropic’s, the Opus 4.8 note covers the current UK position. Treat any quoted figure as a prompt to confirm at source, because AI pricing and terms move quickly.

What we like and what we would watch

Our verdict

Our view is that the Anthropic Economic Index is the most grounded public dataset on AI at work, and we would treat it as a planning input rather than a forecast. The honest headline is mixed: most current use augments people, but the moment firms embed the model in their systems, automation dominates. For individual workers, we would act now to move up the value chain towards judgement and sign-off, because that is where the data says human work holds. For small firms, we would pilot one well-bounded workflow with a human checkpoint, measure it, and scale only on evidence.

We would change that advice if a future report showed augmentation falling sharply in consumer use, or if a UK-specific occupational breakdown revealed concentrated exposure in a particular sector. Until then, the pragmatic position is engaged but disciplined: learn the tools deeply, automate the boring and well-defined, and keep a person accountable for anything that reaches a customer. The firms and workers who treat this index as a map, not a verdict, will be the ones who come out ahead.