A new benchmark says AI agents still can't do most real jobs.
Agents' Last Exam, or ALE, is a new benchmark built with more than 250 industry experts to test AI agents on long, real world professional tasks instead of trivia-style questions. It organizes work using the U.S. government's O*NET/SOC 2018 occupational taxonomy, spanning 13 industry clusters, 55 sub-fields, and more than 1,000 tasks with verifiable outcomes. Unlike a typical benchmark, ALE is designed to keep growing, adding new workflows and industries over time rather than freezing at launch. On its hardest tier, the average full pass rate across mainstream agent setups is below 1%.
That number is the real story. For years, strong benchmark scores have been treated as a proxy for whether AI can do your job, even though actual economic deployment across professions has lagged far behind. ALE was built specifically to expose that gap, measuring whether agents can finish real, valuable work rather than just answer a question correctly.
A benchmark assembled by 250+ industry experts clearing less than 1% of its hardest tasks is a useful gut-check for anyone pricing in imminent job disruption.