AI Models Struggle with White-Collar Work: Insights from the APEX-Agents Benchmark

Recent research from Mercor has unveiled significant challenges faced by leading AI models in performing white-collar tasks, despite advancements in artificial intelligence. The study introduced a new benchmark called APEX-Agents, which evaluates AI's ability to handle complex tasks drawn from fields such as consulting, investment banking, and law. The findings indicate that current AI models are underperforming, with none achieving satisfactory results in real-world professional scenarios.

TechCrunch

The APEX-Agents benchmark revealed that AI models struggled to answer more than 25% of the questions posed by professionals. The scenarios were designed to reflect the multifaceted nature of knowledge work, requiring AI to navigate information across various platforms like Slack and Google Drive. Brendan Foody, CEO of Mercor, emphasized that the benchmark was structured to mimic real professional environments, where context is often dispersed across multiple tools.

TechCrunch

For instance, one question from the law section involved assessing compliance with EU privacy laws, which required a nuanced understanding of both company policies and legal frameworks. Such complexity poses a challenge even for well-informed humans, highlighting the gap in AI's current capabilities.

TechCrunch

Among the AI models tested, Gemini 3 Flash achieved the highest accuracy at 24%, followed closely by GPT-5.2 at 23%. Other models, including Opus 4.5, Gemini 3 Pro, and GPT-5, scored around 18%. These results suggest that while some models are closer to being viable for white-collar tasks, none are yet ready to replace professionals in high-value roles.

TechCrunch

The APEX-Agents benchmark serves as a critical indicator of the potential for AI to automate knowledge work. Foody noted that the benchmark reflects the real tasks professionals undertake, making it a vital tool for assessing AI's readiness for workplace integration. He expressed optimism about future improvements, likening current AI capabilities to an intern who is gradually becoming more competent.

TechCrunch

Despite the advancements in AI, critics argue that the slow progress in automating white-collar work raises questions about the technology's practical applications. Some experts believe that the complexity of human decision-making and the need for contextual understanding may limit AI's ability to fully replace professionals in these fields.

TechCrunch

Brendan Foody stated, “I think this is probably the most important topic in the economy,” underscoring the significance of the benchmark in evaluating AI's impact on the workforce. He acknowledged the historical trend of AI rapidly overcoming challenging benchmarks, suggesting that improvements are likely as the field evolves.

TechCrunch

With the APEX-Agents benchmark now public, it presents an open challenge for AI labs to enhance their models. As the industry continues to innovate, the focus will be on developing AI systems that can better handle the complexities of knowledge work, potentially reshaping the future of professional services.

TechCrunch