Anthropic data shows AI boosts complex work fastest, with uneven impact across jobs and countries

New findings from Anthropic’s Economic Index suggest AI is accelerating higher-skilled tasks more than routine work, raising fresh questions about productivity, job design, and who benefits most from workplace automation.

Anthropic, the AI research company behind the Claude large language model family, has released its fourth Economic Index report, offering one of the most detailed looks to date at how artificial intelligence is being used in real work, education, and personal contexts.

The January 2026 report is based on a privacy-preserving analysis of two million AI conversations, split evenly between consumer use on Claude.ai and enterprise use through Anthropic’s first-party API. Most of the data reflects usage from November 2025, largely powered by Claude Sonnet 4.5.

Unlike many workforce studies that rely on surveys or self-reported adoption, Anthropic’s approach analyzes what users actually ask AI systems to do. The company says this allows it to measure productivity, task complexity, and automation potential at a more granular level than traditional labor market research.

New “economic primitives” add detail to AI impact analysis

The latest report introduces what Anthropic calls economic primitives. These are five foundational measures used to evaluate AI’s economic effects over time: task complexity, skill level, purpose of use, AI autonomy, and task success.

Anthropic applies these primitives to every conversation in its sample by asking Claude to evaluate the nature and outcome of the task. This allows the company to examine not just whether AI is used, but how effectively it supports different kinds of work.

According to Anthropic, these measures act as an early signal for how AI could reshape jobs before changes appear in employment or wage data.

One of the clearest findings in the report is that AI currently delivers the biggest productivity gains on more complex tasks, rather than routine or low-skill work.

Anthropic measures task complexity using estimated years of education required to understand the task. On Claude.ai, tasks aligned with a high school education level were completed about nine times faster with AI assistance. Tasks aligned with a college-level education were completed about twelve times faster. Speedups were even larger in enterprise API use.

While success rates decline slightly as task complexity increases, the drop is modest. Claude successfully completes college-level tasks sixty-six percent of the time, compared with seventy percent for tasks requiring less than a high school education. The report finds that productivity gains still scale more strongly with complexity than success rates decline.

This pattern suggests that AI’s near-term productivity benefits are accruing disproportionately to work that requires higher human capital, reinforcing existing trends in white-collar AI adoption.

AI time horizons are extending as users break work into steps

The report also examines how long AI systems can sustain productive work on a task. Longer tasks remain harder for AI to complete, but Anthropic’s data suggests these time horizons are expanding as models improve and users adapt how they work with them.

Using its own data, Anthropic finds that Claude achieves a fifty percent success rate on tasks that would take a human nearly three and a half hours when accessed through the API. On Claude.ai, the same success rate extends to tasks estimated at around nineteen hours of human effort.

Anthropic attributes this gap partly to user behavior. On Claude.ai, users often decompose complex problems into smaller steps, allowing the model to correct mistakes and continue. The company also notes that users are more likely to bring tasks to Claude when they believe AI will succeed, creating selection effects that differ from benchmark evaluations.

AI use reflects national income and development levels

The report finds clear differences in how Claude is used across countries, closely tied to economic development.

In higher-income countries, AI use is concentrated in work and personal tasks. In lower-income countries, a larger share of usage relates to education and coursework. Anthropic frames this as an adoption curve, where early AI use focuses on skill-building before expanding into broader applications.

The findings align with previous research from Microsoft linking educational AI use with lower per-capita income and leisure use with higher income. Anthropic also references its partnership with the Rwandan government and ALX, where participants begin with AI literacy training before transitioning into broader professional use.

Nearly half of occupations now show meaningful AI exposure

Looking at occupations, Anthropic finds that AI exposure is increasing across the workforce, but not evenly.

In early 2025, thirty-six percent of occupations in the sample used Claude for at least a quarter of their tasks. When pooling data across reports, that figure has risen to forty-nine percent. However, when task success rates and time weighting are factored in, the picture changes.

Some roles, such as data entry keyers and radiologists, show much higher effective AI exposure than raw task counts suggest. Others, including teachers and software developers, appear less affected once reliability and task duration are considered.

Anthropic notes that this analysis only captures tasks performed through Claude and does not fully map onto real-world job redesign, an area the company says requires further study.

AI is more often used on higher-skilled parts of jobs

The report also challenges assumptions that AI primarily automates low-skill work.

Tasks covered by Claude require an average of 14.4 years of education, compared with an economy-wide average of 13.2 years. This indicates that AI is currently more involved in higher-skilled components of jobs.

Anthropic models what would happen if these AI-supported tasks were fully removed from roles. In many cases, this would leave behind lower-skill task mixes, producing a short-term deskilling effect. Occupations such as technical writing, teaching, and travel services would be among those most affected.

The report stresses that this is not a prediction of inevitable deskilling but a signal of near-term pressures if AI adoption outpaces job redesign.

Productivity gains remain significant but smaller after reliability adjustments

Anthropic revisits its earlier estimate that widespread AI adoption could add 1.8 percentage points to annual US labor productivity growth over the next decade.

When adjusting only for task speedups, that estimate holds. However, once task reliability is included, the projected gain falls to around 1.2 percentage points for Claude.ai tasks and roughly 1.0 percentage point for enterprise API use.

Even at that lower level, Anthropic notes that the impact would be substantial, returning productivity growth to rates last seen in the late 1990s and early 2000s. The report also emphasizes that these figures do not account for future model improvements or deeper integration into business workflows.

AI use remains concentrated but is slowly spreading

The report also updates trends tracked throughout 2025. AI use remains concentrated in a small number of tasks, with the top ten accounting for nearly a quarter of all Claude.ai conversations.

Computer and mathematical tasks continue to dominate, representing about one-third of consumer usage and nearly half of enterprise API traffic. The balance between augmentation and automation has shifted slightly toward augmentation, though automation has grown steadily over the year.

Geographically, the United States, India, Japan, the United Kingdom, and South Korea continue to lead in usage. Within the US, however, Claude adoption has become more evenly distributed across states. Anthropic’s modeling suggests national usage could equalize within two to five years if current trends continue.

Anthropic concludes that AI’s impact on work remains highly uneven across countries, occupations, and task types. While productivity gains are real, they are concentrated in specific contexts and depend heavily on task complexity, reliability, and user behavior.