Anthropic ran a privacy-preserving analysis of 100,000 real Claude.ai conversations to estimate how long people’s tasks would take with vs. without AI help. Claude’s own scoring says the typical task would take about 90 minutes unassisted, and that using Claude cuts completion time by roughly 80% on average. They also mapped tasks to O*NET occupations and paired them with wage data, estimating a median implied labor value of about $54 per conversation (with big swings by job type).
The interesting part isn’t just “AI saves time,” it’s where and how unevenly. Claude appears to be used for longer, higher-wage knowledge-work tasks—management tasks averaging ~2 hours unassisted and legal ~1.8—versus food prep tasks around ~0.3–0.5 hours. Savings also vary sharply: healthcare support tasks showed ~90% faster completion while hardware-related issues were closer to 56%; some tasks barely moved (e.g., checking diagnostic images at ~20%), while others fly (compiling info from reports at ~95%). For a concrete “this is why people use these tools” example: curriculum development estimated at 4.5 hours got compressed to 11 minutes, with an implied $115 labor value based on teacher wages.
Then comes the macro extrapolation: assuming universal adoption over 10 years (a huge assumption) and holding model capability constant, they estimate a 1.8% annual boost to US labor productivity—about double the post-2019 run rate—with an implied ~1.1% annual lift in total factor productivity (using a 0.64 labor share). The big caveat: this method can’t see the extra human time spent verifying, editing, or implementing work outside the chat, and Claude’s time estimates still compress short/long tasks. For advertisers, the takeaway is less “everyone will be 80% faster tomorrow” and more “AI will speed up the some parts of marketing first, and slow down others. And whatever it doesn’t speed up becomes the new bottleneck.”
Read more at Anthropic’s Blog.