Artificial intelligence (AI) agents can complete only about 1 in 4 real-world work tasks, according to a new benchmark from UC Berkeley’s Center for Responsible, Decentralized Intelligence that pitted leading AI systems against actual assignments from working professionals. The findings test a bold prediction: that AI agents will replace human workers across most knowledge jobs within one to two years. “There are predictions everywhere that AI agents will surpass humans in almost all jobs between 2026 and 2027,” said Yiyou Sun, a core author of the benchmark. “So, we created this exam to verify this claim,” Sun told FrontierNews.ai.
Researchers gathered 1,490 real work assignments from more than 250 professionals across 55 industries: the kind of projects that would normally take someone hours to weeks to finish, such as preparing a legal filing, building a financial model, or designing a manufacturing part. Each AI system was graded on whether it delivered a correct, complete, finished product, with no partial credit for a good try. A missing deliverable or something done wrong counted as a fail, according to the paper.
Even the Best AI Agents Fall Short on Real Work
The best-performing system, OpenAI’s Codex tool running on its GPT-5.5 model, completed 26.2% of all the assignments correctly, or roughly one in four, the paper found. On the hardest assignments, the kind that require an AI to keep track of many steps over a long stretch of work and get every part right, pass rates across all the systems tested averaged just 2.6% Even Codex managed only 8.6% on that toughest tier.
Anthropic’s Claude Code tool, running on its Opus 4.7 model, failed every one of those hard assignments. One version of the test threw enormous computing resources at the problem: $630 worth of computing power and 763 million “tokens,” the units AI systems use to process text and information, and still succeeded just 2.9% of the time. The takeaway, according to the researchers, is that spending more money and computing power doesn’t reliably make an AI agent better at this kind of work.
Agents Handle Simpler Tasks Well but Struggle With Long, Multi-Step Work
That gap, from a 26.2% overall pass rate down to a 2.6% average on the hardest tier, shows exactly where today’s AI agents are useful and where they aren’t. On the test’s easier tier, 59 assignments that current AI systems can partially handle, the top performers succeeded on roughly 30% of tasks, the paper reported, with Codex on GPT-5.5 reaching 42.4% on that easier set specifically.
An AI agent might correctly answer a single finance question, then fail completely when asked to carry that answer through an entire multi-step process: gathering the right data, checking its own work, and producing a final report someone could actually use. Staying on track through a long sequence of steps, catching and fixing its own mistakes along the way, and delivering something ready to use remains something today’s AI systems mostly can’t do reliably.
What This Means for Companies Deciding Where to Use AI Agents
The 26.2% overall figure gives businesses a useful signal for where AI agents can be put to work today: on narrower, well-defined tasks, not open-ended, multi-step projects that require judgment calls along the way. Companies can restructure jobs around the tasks AI handles reliably, freeing up employees to focus on the judgment-heavy, accountability-heavy work AI still can’t be trusted with.
This mirrors what’s already happening in corporate finance. More than 8 in 10 CFOs at large companies are already using AI for accounts payable or are seriously considering it, according to PYMNTS Intelligence research. Adoption of AI agents specifically is far more limited: close to 7% of U.S. finance chiefs have already put AI agents to work in live, day-to-day finance operations, with another 5% currently testing the technology before a full rollout, according to another PYMNTS Intelligence study. Those early deployments are concentrated in structured, repeatable tasks with clear rules.
