New study shows AI still stumbles on everyday office challenges

When Mercor released its APEX-Agents benchmark, the tech world expected a headline-grabbing triumph. Instead, the results painted a sober picture: even the most celebrated language models faltered on tasks that mimic real-world consulting, legal analysis, and financial decision-making. Researchers fed the systems scenarios that a seasoned professional would navigate-drafting a client brief, interpreting a contract clause, or weighing an investment recommendation-and watched as the AI's responses fell short of basic competency. The experiment, designed to strip away the polished veneer of curated prompts, revealed a gap between glossy demos and the gritty reality of office work. Participants noted that the models often missed crucial context, offered generic advice, or simply guessed when pressed for specifics. Industry insiders, who have long championed the promise of AI as a productivity booster, now face a reckoning. The findings suggest that the technology is still a tool for assistance, not a replacement for human judgment in complex, nuanced environments. As companies weigh the hype against the hard data, the consensus is clear: the road to truly reliable office-level AI is longer than many anticipated.