Anthropic measures how AI is already changing work

Anthropic has just published a detailed look at how Claude is being used and what that could mean for work. The report analyzes two million interactions collected between November 13 and 20, 2025: one million conversations from Claude.ai and another million records from its enterprise API.

The main conclusion is less spectacular and more useful than many predictions about the future: AI is already speeding up real tasks, but its impact depends on which task it performs, how successfully it completes it and how much human judgment is still needed.

AI is still concentrated in programming

Programming remains Claude's dominant use. On Claude.ai, computer and mathematical tasks account for 34% of conversations, while on the enterprise API they reach 46%.

The most common task is modifying software to fix errors. On its own, it represents 6% of use on Claude.ai and roughly one in ten requests on the enterprise API.

Use is also highly concentrated. The ten most common tasks account for 24% of Claude.ai conversations and 32% of API traffic. That means that although Claude can respond to thousands of different activities, companies and users still mainly turn to it for a small group of tasks with clear value.

Education is the second-largest area of use on Claude.ai. Conversations related to courses, tutoring and teaching materials rose to 15%, up from 9% in January 2025.

People are working with AI again, not just handing tasks over to it

In November, 52% of Claude.ai conversations involved collaboration: the user reviews, corrects, learns or iterates alongside the system. Another 45% were classified as automation, where Claude receives an instruction and completes the task with little interaction.

The result contrasts with August 2025, when automation had temporarily overtaken collaboration. Anthropic links the change to features such as file creation, persistent memory and tools for customizing workflows.

Businesses show the opposite pattern. The API is mainly used to automate processes such as:

• Managing and classifying emails.
• Processing documents and invoices.
• Scheduling meetings.
• Generating sales responses.
• Creating and maintaining customer service systems.

For you, the difference matters. A conversation on Claude.ai is more like working with an assistant that helps you think. An enterprise integration is more like building AI into a process that runs automatically.

Claude performs better on simpler tasks

The report introduces a measure that has been missing from this type of analysis until now: the success rate. Anthropic asks Claude to estimate whether a task was completed correctly and combines that data with its duration, complexity and level of autonomy.

The pattern is clear: the longer a person would need to complete a task, the less likely Claude is to solve it successfully.

For API requests, the success rate falls from around 60% for tasks taking less than one hour to approximately 45% for tasks that would require five hours or more from a person. The point at which success drops to 50% appears near 3.5 hours of human work.

On Claude.ai, where there is more back-and-forth between the user and the model, the result is different. The 50% success rate would be reached near 19 hours, according to an extrapolation in the report. Anthropic warns that this is not a controlled test: users choose which tasks to bring to Claude and tend to avoid those they think will not work.

The practical lesson is simple: asking AI to speed up a complex task can save a lot of time, but it also increases the need to review the result.

Estimated productivity falls when reliability is taken into account

In an earlier analysis, Anthropic calculated that widespread use of Claude could increase annual labor productivity growth in the United States by 1.8 percentage points over the next decade.

After discounting tasks that Claude does not complete successfully, the estimate falls to:

• 1.2 percentage points for use on Claude.ai.
• 1 percentage point for enterprise API traffic.

These are not results observed in the economy. They are estimates based on usage patterns and assumptions about how improvements in specific tasks translate into overall productivity.

The effect could also be smaller if some tasks act as bottlenecks. A teacher might prepare materials faster with AI, for example, but that does not remove the time needed to teach and support students.

The impact will not be the same for all workers

Claude is used for tasks that require, on average, more education than tasks across the economy as a whole. The tasks observed in the report require around 14.4 years of education, compared with 13.2 years for all the tasks analyzed.

If AI removes the most specialized parts of a job first, the remaining work may become less skilled. Anthropic calls this a possible deskilling effect.

A travel agent might retain routine tasks such as printing tickets or collecting payment while Claude handles complex itinerary planning. A real estate manager, by contrast, might delegate accounting and focus on negotiating contracts and dealing with other parties, making the remaining work more specialized.

That is why it is not enough to ask which jobs are exposed to AI. You also need to look at which part of the job is automated and which tasks remain for people.

The geographic gap remains

Claude use remains concentrated in countries with higher GDP per capita. Anthropic estimates that a 1% increase in GDP per capita is associated with a 0.7% increase in Claude use per capita, although the relationship does not show that one variable causes the other.

In lower-income countries, educational use carries more weight. In wealthier countries, personal and professional uses are growing more. This suggests that AI is not being adopted in the same way everywhere: some users employ it to address specific needs or study, while others incorporate it into a wider range of work and everyday activities.

Adoption also remains uneven within the United States, although differences between states narrowed between August and November 2025. If the observed pace continued, use per capita could approach national parity in two to five years. Anthropic stresses that the estimate is based on only three months of data and carries considerable uncertainty.

The figure worth watching is not just how many people use AI, but whether they can use it well. The education needed to understand users' instructions and Claude's responses is almost perfectly related, with a correlation above 0.92. In other words, the quality of what you get depends to a large extent on how you frame the problem.

The report points to a specific scenario: AI can increase productivity, but its benefits will be greater for people with the skills, context and time needed to supervise it. The next important signal will be whether models improve their reliability on long tasks and whether companies move those gains from chat into complete work processes.