Key Takeaways
- Anthropic’s analysis of 1 million real conversations shows 57% of AI use augments human work; 43% automates tasks directly
- Only 4% of occupations use AI across 75% or more of their tasks, while 36% use it for at least 25% of tasks
- Computer and mathematical tasks dominate Claude usage at 37.2%, far outpacing every other job category
- By November 2025, augmented use rose back to 52% of conversations, reversing an automation spike seen in August 2025
Anthropic’s Economic Index is the first large-scale dataset built from how AI is actually used, not how users say they use it. Drawing on anonymized analysis of approximately 1 million real Claude conversations, it reveals a labor market in early-stage structural change, where the impact is concentrated, uneven, and measurably accelerating.
What the Anthropic Economic Index Measures and Why It Matters
Launched in February 2025, the Anthropic Economic Index uses a privacy-preserving analysis system called Clio to map real Claude.ai conversations to over 20,000 specific work tasks in the US Department of Labor’s O*NET database. This approach gives researchers and policymakers direct behavioral data rather than survey responses or theoretical projections.
The January 2026 report expanded the methodology by introducing five new “economic primitives”: task complexity, human and AI skill levels, use case (work, education, or personal), AI autonomy, and task success rate. Together, these metrics provide a more precise picture of where AI is genuinely capable of substituting for human labor and where it currently falls short.
Where AI Is Concentrated in the Workforce
The data shows sharp concentration in a narrow set of occupational categories. Computer and mathematical tasks accounted for 37.2% of all Claude queries in the original February 2025 dataset, covering activities like software modification, code debugging, and network troubleshooting.
The second-largest category was arts, design, entertainment, sports, and media at 10.3% of queries, which primarily reflected writing and editing tasks. At the opposite end, physical and manual labor categories saw minimal AI usage: farming, fishing, and forestry represented just 0.1% of queries. Both the lowest-paid manual roles and the highest-paid specialist roles such as obstetricians showed low AI usage, reflecting the limits of current AI capabilities and practical barriers to adoption.
The Automation vs. Augmentation Split
One of the most closely watched metrics in Anthropic’s research is the balance between automation, where AI directly performs a task, and augmentation, where AI collaborates with a human to complete it.
In the initial February 2025 data from Claude.ai Free and Pro users, augmentation held a slight lead at 57% versus 43% for automation. By August 2025, that balance shifted: automated use exceeded augmented use on Claude.ai for the first time, driven by a rise in “directive” interactions where users delegated tasks to Claude with minimal back-and-forth, reaching 39% of all conversations. Three months later in November 2025, the ratio reversed again: augmented use climbed back to 52% and automated use fell to 45%, with directive conversations dropping 7 percentage points to 32%.
Product changes during this period, including file creation capabilities, persistent memory, and Skills for workflow customization, may have shifted usage toward more collaborative, human-in-the-loop interactions.
How Deep AI Penetration Actually Is
The depth of AI use within individual occupations is more limited than public discourse often suggests. Only approximately 4% of occupations use AI for at least 75% of their associated tasks. Moderate use is more widespread: roughly 36% of occupations see AI applied to at least 25% of their tasks.
Importantly, the February 2025 data found no evidence of any occupation being entirely automated. Instead, AI is diffused across many tasks in the economy, with stronger effects for specific clusters of work. The January 2026 report added nuance by factoring in task success rates: Claude generally succeeds at the tasks it is given, but its success rate falls as task complexity increases, measured by how long a human would take to complete the same task without AI assistance.
Which Occupations Face the Most Significant Exposure
The January 2026 report introduced a refined measure of occupational exposure that weights task coverage by both success rates and the importance of each task within the overall job. This produces a more realistic picture than simple task-count exposure measures.
For some occupations, including data entry keyers and database architects, Claude shows proficiency across a large portion of the job’s core tasks. The report also examined what happens to the skill composition of jobs when AI-assisted tasks are removed. Travel agents face a deskilling risk: if AI absorbs complex itinerary planning, workers are left with routine ticket purchasing and payment collection. Property managers face the opposite outcome: when AI handles bookkeeping tasks, workers shift toward contract negotiations and stakeholder management, which represents upskilling.
How Usage Patterns Differ Between Consumers and Enterprises
The January 2026 report tracked both Claude.ai consumer data and first-party API enterprise data, revealing meaningfully different usage profiles.
On Claude.ai, the primary use case is work-related, but educational use accounts for a substantial and growing share: conversations involving educational instruction tasks rose from 9% in January 2025 to 15% by November 2025. Enterprise API usage is overwhelmingly automation-dominant, reflecting its programmatic, batch-processing nature. The most notable trend for API enterprise customers between August and November 2025 was a 3-percentage-point rise in office and administrative support tasks to 13% of traffic, suggesting businesses are increasingly deploying Claude to automate routine back-office workflows such as email management, document processing, and scheduling.
Geographic Concentration of AI Use
AI adoption is not uniformly distributed across the workforce or across regions. Within the US, the top five states by Claude usage account for nearly 50% of all usage despite representing only 38% of the working-age population. States with higher concentrations of computer and mathematical workers, such as Washington D.C., Virginia, and Washington state, show systematically higher usage per capita.
Globally, the US, India, Japan, the UK, and South Korea lead in overall Claude.ai use. Worldwide, usage per capita remains strongly correlated with GDP per capita, and there is no evidence yet that lower-income countries are closing the gap with higher-income countries. Within the US, however, there are early signs of regional convergence: states with lower usage are growing adoption faster, and modeling suggests usage per capita could equalize across US states within two to five years if current trends hold, a pace roughly ten times faster than the spread of major 20th-century technologies.
What This Means for Workers Right Now
The Anthropic data points toward specific patterns workers should understand when assessing their own exposure:
- Mid-to-high wage roles face the highest current AI engagement, not low-wage ones. Computer programmers, data scientists, and copywriters appear most prominently in usage data
- Roles concentrated in high-volume, codifiable cognitive tasks face the clearest automation pressure. This is where directive usage is rising fastest in enterprise deployments
- Jobs where AI absorbs the most complex tasks face deskilling risk over time. The travel agent pattern from Anthropic’s January 2026 data illustrates how competency can erode if workers do not actively maintain skills in AI-assisted areas
- Jobs where AI absorbs routine sub-tasks can create upskilling opportunities. Property managers in Anthropic’s data moved toward higher-value tasks when AI handled lower-skill components
- Educational use of AI is growing fastest in lower-GDP countries, which may reflect early-adopter, high-value technical use rather than broad workplace integration
Limitations of Anthropic’s Data
Anthropic’s own research acknowledges important constraints in interpreting these findings. The dataset covers only Claude.ai Free, Pro, and Max users plus first-party API customers. It does not capture usage of ChatGPT, Gemini, Microsoft Copilot, or thousands of industry-specific AI tools, meaning total AI labor market impact is likely larger than any single provider’s data reflects.
The research team also cannot confirm with certainty whether every conversation in the dataset reflects paid work rather than personal or hobby use. The “automation” classification in the data may overstate actual automation because some users who ask Claude to produce complete outputs then edit them afterward, a behavior that functions as augmentation but appears as automation in the transcript. Additionally, because Claude is specifically marketed as a strong coding tool, software tasks may be overrepresented relative to the broader AI market.
iOS 26.3.1 (23D8133): Every Confirmed Fix and Change Apple Just Shipped to Your iPhone and iPad
Frequently Asked Questions (FAQs)
What is the Anthropic Economic Index?
The Anthropic Economic Index is an ongoing research initiative that analyzes anonymized real-world conversations with Claude to map how AI is being used across occupational tasks in the economy. It uses the US Department of Labor’s O*NET task database as a framework, covering approximately 20,000 specific work tasks.
Does Anthropic’s data show AI automating more than it augments?
As of November 2025, augmentation is again the dominant pattern at 52% of Claude.ai conversations, versus 45% automated. This reversed an August 2025 spike when automated use briefly exceeded augmented use for the first time since tracking began in early 2025. The underlying trend still leans gradually toward greater automation over time.
Which jobs show the highest AI usage according to Anthropic’s data?
Computer and mathematical occupations dominate at 37.2% of all Claude queries, followed by arts, design, entertainment, and media roles at 10.3%. Mid-to-high wage occupations, including computer programmers and copywriters, show the highest per-capita AI usage. Both the lowest-paid and highest-paid roles show lower usage.
What does “deskilling” mean in the context of Anthropic’s research?
Deskilling occurs when AI absorbs the most complex components of a job, leaving workers with lower-skill tasks. Anthropic’s January 2026 report uses travel agents as a concrete example: if AI handles complex itinerary planning, the remaining work shifts toward routine ticket purchasing and payment processing, reducing the skill intensity of the role over time.
How does AI usage differ between US states?
Usage per capita within the US correlates strongly with the share of workers in computer and mathematical occupations. States with more tech workers, such as Washington D.C. and Virginia, show higher usage. The top five states account for 50% of all US Claude usage despite holding only 38% of the working-age population. Regional convergence is occurring but uneven.
Does Anthropic’s data show AI replacing entire jobs?
No. The February 2025 data found no evidence of any occupation being fully automated by AI. Only approximately 4% of occupations use AI for 75% or more of their tasks. AI is penetrating specific task clusters within jobs rather than eliminating whole occupational categories.
How fast is AI adoption spreading geographically?
Within the US, early modeling based on November 2025 data suggests usage per capita could equalize across states in two to five years if current convergence trends continue. Globally, the picture is different: adoption remains strongly tied to GDP per capita, and lower-income countries are not closing the gap with higher-income countries at this time.
Is enterprise use of AI more automation-focused than consumer use?
Yes, substantially. First-party API enterprise traffic is automation-dominant by design, reflecting programmatic and batch-processing use cases. The fastest-growing enterprise category between August and November 2025 was office and administrative support tasks, rising 3 percentage points to 13% of API traffic.
