Anthropic tracks how AI really affects your job

For the past two years, the debate over whether AI will steal everyone's jobs has mostly relied on speculative think-tank whitepapers and vague survey data. Nobody was actually opening the server logs to see what people are typing into chatbots all day.

Anthropic just changed that. The AI lab introduced the Anthropic Economic Index, a long-term research initiative mapping how AI touches real-world labor markets. Alongside the launch, Anthropic open-sourced the underlying dataset and invited economists and policy experts to dig in.

Instead of guessing, the team analyzed roughly one million anonymized Claude.ai conversations from Free and Pro tiers, mapping individual prompts directly against occupational tasks.

Coders lead, but the extremes stay offline

To translate raw user queries into labor economics, Anthropic built an automated analysis tool called Clio. Clio categorized chats against the U.S. Department of Labor's O*NET database, which tracks around 20,000 specific workplace tasks.

The findings show usage is heavily lopsided. Computer and mathematical tasks sucked up 37.2% of Claude's workload, dominated by code debugging, software modification, and network troubleshooting.

Arts, design, entertainment, and media came in second at 10.3%, mostly driven by copywriting, drafting, and editing. At the bottom, physical roles like farming, fishing, and forestry accounted for just 0.1%.

Salary data revealed a striking bell-curve pattern. AI adoption is lowest among the lowest-earning and highest-earning jobs. Roles requiring manual dexterity or physical presence, like shampooers and obstetricians, barely register.

Instead, usage heavily clusters around mid-to-high wage knowledge workers, specifically software engineers and copywriters.

Augmentation still beats replacement

Despite ongoing fears of mass layoffs, Anthropic's data paints a more collaborative picture for now.

Around 36% of occupations use AI for at least a quarter of their daily tasks, but only 4% rely on it for 75% or more of their workload. There was no evidence in the data of any job being completely automated out of existence.

When looking at how tasks are executed, augmentation beat direct automation 57% to 43%.

In most cases, people are treating Claude as a second pair of eyes, using it to validate complex work, learn new technical concepts, or bounce ideas back and forth, rather than outsourcing entire deliverables.

Anthropic noted several obvious caveats: Claude's reputation as a top-tier coding model likely skews the sample, and the dataset excludes enterprise and API traffic.

Why it matters

Most AI labor impact discourse tends to swing wildly between apocalyptic replacement predictions and breathless marketing hype. Anthropic's index brings much-needed empirical ground truth to the table by linking actual conversational telemetry to granular labor benchmarks.

For workers, the takeaway is clear: AI is dismantling tasks rather than wiping out whole careers overnight. If you sit in that mid-to-high knowledge salary band, learning to master the 57% augmentation workflow is where immediate leverage lives today.

Strategically, open-sourcing this data is a brilliant flex by Anthropic. It positions the company as a thoughtful, research-first steward in the eyes of regulators and labor economists, raising the bar for OpenAI and Google to be equally transparent.

Still, we think this initial snapshot has blind spots. Without API and enterprise data, it misses the heavy programmatic automations happening quietly inside corporations, and Claude's lack of native image generation understates creative sector disruption. The real transition might move faster beneath the surface than these consumer logs suggest.