Imagine an artificial intelligence acting as your Chief Financial Officer, making split-second decisions to keep your business afloat for over a decade. Now imagine it fails spectacularly. A new study reveals a startling reality about the capabilities of current large language models when put to the ultimate test of enterprise management.

Researchers have launched EnterpriseArena, the first benchmark designed to evaluate how well AI agents handle long-term resource allocation in dynamic business environments. Unlike simple chatbots that answer questions instantly, this system simulates a complex company with 132 months of financial data, market signals, and strict operating rules. The goal is purely survival: allocate money wisely over time without running out of resources or making bad strategic choices.

Here lies the conflict that breaks most sophisticated AI systems. The environment forces agents to make tough trades between gathering more information and saving scarce capital. While smaller models struggled even more, larger, state-of-the-art AI did not automatically win. In a stunning display of inconsistency, only 16 percent of test runs managed to survive the full simulation horizon. This suggests that despite their size, bigger models are not necessarily smarter at navigating uncertainty than their smaller counterparts.

The takeaway is clear: there is a significant gap between today's advanced AI and the true ability to manage long-term enterprise risk under pressure.

Source: Can LLM Agents Be CFOs? A Benchmark for Resource Allocation in Dynamic Enterprise Environments by Yi Han et al., https://arxiv.org/abs/2603.23638