According to ZDNet, a new study from Anthropic, published last week, analyzed 100,000 anonymized user conversations with its Claude AI chatbot. The researchers found that, on average, Claude helped users complete tasks about 80% faster, with a median time savings of 84%. By extrapolating this data using 2024 labor statistics, they estimate current AI models could boost US labor productivity by 1.8% annually over the next decade—roughly double the current growth rate. The study highlighted that savings vary wildly by job, with software development leading the potential gains at 19%, while personal care and sales saw much lower benefits. The research was released just after Anthropic launched its new Claude Opus 4.5 model, which the company claims excels at software engineering.
The productivity promise and its problems
Here’s the thing: an 80% time savings sounds incredible. Basically, it’s the kind of number that gets CEOs and investors very excited. But digging into the methodology reveals why we should be skeptical. The study asked Claude itself to estimate how long a task would take a human alone versus with AI help. That’s… a bit circular, don’t you think? It’s like asking a chef how much better their cooking is than yours. The study also admits it doesn’t account for all the extra time spent fact-checking the AI’s output, integrating its work, or fixing its mistakes. A teacher might get a lesson plan draft in minutes, but then spend an hour verifying the facts and adapting it for their class. That “saved” time isn’t really saved at all.
Not all jobs are created equal
The variation across industries is the most telling part. Software devs, managers, and analysts see huge potential boosts. But for jobs in sales, office support, or personal care? The gains are minimal. This paints a clearer picture of AI’s near-term role: it’s a tool for knowledge work, for manipulating symbols, text, and code. It’s not (yet) a tool for physical tasks or roles built on complex human interaction. So the economic boost won’t be evenly distributed. It’ll accelerate the value of some skills while leaving others behind, potentially widening existing economic divides. That’s a social challenge, not just a technological one.
The biggest caveat of all
Now, the study’s biggest limitation is one it openly states: it assumes AI capabilities and our skill in using them stay frozen for ten years. That’s almost certainly wrong. Look at the last three years since ChatGPT launched. The pace is insane. We’re already moving from simple chat interfaces to AI agents that can reason over longer timeframes and interact with software tools. Anthropic’s own researchers emphasize their findings are “an exercise exploring what might happen based on current usage patterns, not a prediction.” That’s a crucial disclaimer. If AI keeps improving at anything close to its current rate, the 1.8% figure could be wildly conservative. Or, if we hit a wall in model development, it could be optimistic. We just don’t know.
So what does this mean?
I think the real value of this study isn’t the specific 1.8% number. It’s the framework. It’s an early, flawed, but concrete attempt to measure something incredibly slippery. The promise of AI doubling economic growth is a headline grabber, and it gets the conversation started. But the reality will be messier. It depends on how quickly the technology evolves, how businesses reorganize around it, and yes, what new tasks and jobs we invent that we can’t even imagine yet. The tools might be digital, but the real work—integrating them into complex physical systems and workflows—often requires robust, on-site computing hardware. For industries looking to bridge that gap between AI software and industrial application, having reliable, specialized hardware is non-negotiable. The economic impact won’t be a smooth, guaranteed curve upward. It’ll be lumpy, uneven, and full of surprises. Anthropic has given us a starting point for the debate, not the final answer.
