AI Agents Fail to Manage Simulated Businesses in Carnegie Mellon Study
A Carnegie Mellon University experiment tested AI agents managing a simulated business. Claude 3.5 Sonnet completed only 24 percent of tasks, highlighting current limitations in autonomy.
The global business world is intensely debating whether artificial intelligence is poised to take over all jobs, prompting many companies to rush into adopting new technologies while others remain skeptical. To test the reality behind these grand promises, researchers at Carnegie Mellon University designed an intriguing experiment, published as a preliminary paper on arXiv, in which AI agents were handed the reins of a simulated business. The outcome, to put it mildly, was less than stellar.
Simulating Corporate Roles
As part of the experiment, researchers placed AI agents into various organizational roles including financial analyst, project manager, and software engineer. To create a realistic simulation, simulated colleagues in other departments such as human resources were also introduced, whom the agents could consult when necessary. The AI agents were assigned a wide range of tasks, ranging from reviewing files for database analysis to conducting virtual tours in order to select new premises for the business.
The numerical results presented clear data. The Claude 3.5 Sonnet model led the list, yet managed to complete only 24 percent of the tasks assigned to it. Even when factoring in attempts that were only partially completed, its success rate stood at 34.4 percent. The Gemini 2.0 Flash model took second place, registering a completion rate of just 11.4 percent. No other AI agent managed to cross the ten percent threshold. In terms of operational costs, Claude 3.5 Sonnet accumulated an expense of $6.34, compared to Gemini 2.0 Flash which cost a mere $0.79.
Limitations and Human Oversight
"The agents repeatedly failed to grasp the implicit aspects of instructions, stumbling over common social and digital navigation challenges." — Carnegie Mellon Study
The experiment highlighted the classic and well-known limitations of artificial intelligence systems. As the researchers noted, these agents frequently failed to understand the implicit parts of prompts. For instance, when asked to save a result in a file with a ".docx" extension, they failed to deduce that they needed to use the Microsoft Word format.
Social skills are not a strong suit of these systems, and many tasks remained unfinished because the AI could not cope with the human and social element. Furthermore, web navigation tripped them up, particularly in the face of pop-up windows. These findings confirm that while AI can deliver impressive results in specific, well-defined tasks, the systems remain a long way from becoming fully autonomous digital workers.
