Anthropic’s Economic Index offers a look at how organisations and individuals are actually using large language models. The report contains the company’s analysis of a million consumer interactions on Claude.ai, plus a million enterprise API calls, all dated from November 2025. The report notes that its figures are based on observations, rather than, for example, a sample of business decision-makers or generic survey.
Limited use cases dominate
Use of Anthropic’s AI tends to cluster around a relatively small number of tasks, with the ten most frequently-performed tasks accounting for almost a quarter of consumer interactions, and nearly a third of enterprise API traffic. There’s a focus on the use of Claude for code creation and modification, as readers might expect.
This concentration of use of AI as a software development tool has remained fairly constant over time, suggesting that the model’s value is largely based around these types of tasks, with no emerging use of Claude for other purposes of any empirical significance. This suggests that broad, general rollouts of AI are less likely to be successful than those focused on tasks where large language models are proven to be effective.
Augmentation outperforms automation
On consumer platforms, collaborative use – where users iterate on queries to the AI over the course of a virtual conversation – is more common than using the AI to produce automated workflows. Enterprise API usage shows the opposite, as businesses attempt to gain savings through automating tasks. However, while Claude succeeds on shorter tasks, the observed quality of outcomes declines the more complex the task (or series of tasks) is, and the longer the required ‘thinking time’ required.
This implies automation is most effective for routine, well-defined tasks that are simpler, require fewer logical steps, and where responses to queries can be quick. Tasks estimated to take humans several hours show significantly lower completion rates than shorter tasks. For longer tasks to succeed, users have to iterate and correct outputs.
Users breaking down large tasks into manageable steps and posing each separately (either interactively or via API) have improved success rates.
The company’s observations show most queries put to the LLMs are associated with white-collar roles (although poorer countries tend to use Claude in academic settings more commonly than, for instance, the US). For example, travel agents can lose complex planning tasks to the LLM and retain elements of their more transactional work, while some roles, such as property managers, show the opposite: routine administrative tasks can be handled by the AI, and tasks needing higher-judgement remain with the human professional.
Productivity gains lessened by reliability
The report notes that claims of AI boosting annual labour productivity by 1.8% (over a decade) are likely best to be reduced to 1-1.2%, due to the need to factor in extra labour and costs. While a 1% efficiency gain over a decade is still economically meaningful, the need for activities such as validation, error handling, and reworking will lower success rates and therefore there should be a similar adjustment in the minds of a business’s decision-makers.
Potential gains to an organisation deploying AI also depend on whether tasks given to the LLM complement or substitute work. In the latter case, the success of substituting an AI for tasks normally done by a human depends on how complex the work is.
It’s noteworthy that the report finds a near-perfect correlation between the sophistication of users’ prompts to the LLM and successful outcomes. Thus, how people use AI shapes what it delivers.
Key takeaways for leaders
- AI implementation delivers value fastest in specific, well-defined areas.
- Complementary systems (AI+human) outperform full automation for complex work.
- Reliability and necessary extra work ‘around’ the AI reduce predicted productivity gains.
- Changes to workforces’ makeup depend on the mix of tasks and their complexity, not specific job roles.
(Image source: “the virtual construction worker” by antjeverena is licensed under CC BY-NC-SA 2.0.)
