90% of US working hours are spent on tasks not covered well by AI benchmarks

A viral story on Hacker News highlights a striking gap in tech's current narrative: 90% of US working hours are spent on tasks not covered well by AI benchmarks. While the industry fixates on coding assistants and automated agents, most human labor consists of nuanced, context-heavy tasks that lack the tight feedback loops required to train modern AI models.

For early-stage founders, this is a reality check against the prevailing AI hype. Software engineers often build tools for themselves because coding offers clear success metrics and fast feedback. But assuming the rest of the global economy operates like a software repo is a mistake. If nine-tenths of human labor sits outside standard benchmarks, broad AI wrappers are going to run into a wall very quickly.

The real venture opportunity isn't trying to outpace big tech labs on standard benchmarks. Instead, it lies in solving the unglamorous, unbenchmarked 90%—building specialized tools for industries where workflows are messy and feedback loops are long.

For backers investing from £500, look for founders who understand this divide. The biggest long-term wins won't come from another generic agent promising a work revolution, but from startups tackling the stubborn, real-world work that AI labs are currently ignoring.