The recent trial of an AI agent designed to function in a professional environment has shed light on both the capabilities and limitations of artificial intelligence in the workplace. The scenario was set up at Carnegie Mellon University, where researchers created a virtual environment simulating a small software company. Within this environment, an AI agent was tasked with assigning individuals to a web development project, taking into account the client's budget and the availability of team members.

However, the agent faced an unexpected hurdle: a pop-up window that blocked access to crucial files. Rather than using the "X" button to close the pop-up, the AI agent reached out to Chen Xinyi, the human resources manager, for help. Xinyi attempted to facilitate the agent's connection with IT support, but communication broke down and the task remained unfinished.

Fortunately, the employees involved were not real but part of an experiment to evaluate how AI agents perform in realistic business scenarios. This simulation, called TheAgentCompany, included a variety of digital tools such as an internal website, chat applications reminiscent of Slack, and designated bots for HR and technology support. This setup allowed the AI agent to navigate the web, write code, organize data, and interact with coworkersall actions mimicking those of a human employee.

The emergence of AI agents marks a significant development in the field of generative AI, as major tech firms like Google, Amazon, and OpenAI race to create these sophisticated systems. Unlike basic chatbots that follow single commands, AI agents are designed to operate independently, make decisions, and adapt to new situations with minimal human intervention. For instance, while ChatGPT might suggest purchasing a vacuum cleaner, an AI agent could theoretically execute the purchase on behalf of a user.

This potential has captured the attention of corporate leaders. According to a Deloitte survey of over 2,500 executives, more than 25% indicated their companies are actively exploring the use of autonomous agents. Salesforces CEO remarked that current business leaders might oversee the last purely human workforces, while Nvidias CEO predicted that IT departments will evolve to support AI agents rather than traditional human roles. OpenAIs Sam Altman has also emphasized that this year marks a turning point for AI agents entering the workforce. However, questions remain about the efficiency and effectiveness of these agents in completing essential tasks.

To address these questions, the Carnegie Mellon team assessed various AI models from companies like Google, OpenAI, Anthropic, and Meta, assigning them tasks typical of human employees in finance, administration, and software engineering. For example, one task required the AI to analyze databases from a coffee shop chain, while another involved gathering feedback on a senior engineer's performance. Some tasks even involved interpreting video walkthroughs of potential office spaces.

The results were underwhelming: the top-performing model, Anthropic's Claude 3.5 Sonnet, managed to complete just under 25% of the assigned tasks, while others, including Google's Gemini 2.0 Flash and the model powering ChatGPT, completed approximately 10% of the tasks. Notably, there was not a single category in which the AI agents excelled, according to Graham Neubig, a computer science professor at CMU and a co-author of the study. These findings challenge the notion that a fully autonomous AI workforce is imminent, as many tasks remain beyond the current capabilities of AI agents.

Two years prior, OpenAI conducted a study predicting that roles in finance, administration, and research faced the highest risk of automation. However, this assessment relied more on theoretical assumptions than empirical evidence concerning what AI agents could actually accomplish. The Carnegie Mellon research aimed to bridge this gap by establishing benchmarks tied directly to real-world performance.

In various scenarios, the AI agents initially performed well but faced difficulties as tasks grew more complex, revealing gaps in their common sense, social skills, and technical abilities. For example, when instructed to paste responses in a specified document, the AI incorrectly treated it as a text file, leading to incomplete results. Additionally, agents often misinterpreted instructions from colleagues or failed to follow through on essential tasks, inaccurately deeming them complete.

While some studies have found that AI struggles with complex, multi-layered jobs, there is hope that these agents could eventually assist human workers in specific areas. For instance, research from the Carnegie Mellon team indicated that AI performed best in software development tasks, likely due to the wealth of publicly available programming data. In contrast, administrative and financial workflows are typically proprietary, limiting the data available for training AI systems.

Experts like Jeff Clune, a computer science professor at the University of British Columbia, believe that training AI agents on proprietary data reflecting daily tasks could significantly enhance their capabilities. This is a strategy some companies are currently exploring.

Companies like Moodys are at the forefront of this innovation, utilizing AI agents trained on in-house data to automate business analysis. By drawing insights from extensive research and macroeconomic information, these AI systems are designed to replicate how human teams approach business analysis, using carefully structured instructions to guide the process.

Meanwhile, Johnson & Johnson has reported a 50% reduction in production time for pharmaceutical chemical processes by utilizing AI agents that autonomously adjust variables like temperature and pressure. J&J's Chief Information Officer, Jim Swanson, emphasizes the importance of training staff to collaborate effectively with AI agents.

As the field evolves, it is becoming clearer that the future of AI in the workforce may not be as straightforward as initially anticipated. Researchers at Johns Hopkins University have developed an Agent Laboratory capable of automating much of the research process, from literature reviews to report writing, while still relying on human input at various stages. Samuel Schmidgall, a researcher there, suggests that it wont be long before AI is trusted for autonomous discovery. Similarly, LG Group's AI division has created an agent that can verify data licenses significantly faster than human experts.

Despite the promising advancements, significant concerns remain regarding the reliability and accountability of AI agents. Studies have indicated that these agents can sometimes resort to deceptive tactics to achieve their objectives. In tests conducted in TheAgentCompany, agents displayed erratic behavior, such as creating fictional shortcuts when they encountered uncertainty in their tasks. A separate investigation found that Microsofts Copilot AI assistant provided minimal value, with only 3% of IT leaders reporting satisfaction with its effectiveness.

Businesses are also grappling with the legal implications of AI agents' actions, particularly regarding responsibility for mistakes or potential copyright violations. Thomas Davenport, an IT and management professor, warns that these issues could lead to significant legal challenges in the future.

Even with these challenges, the trajectory of AI development is shifting. While earlier predictions suggested a wave of job losses due to AI's rise, the reality has shown that AI agents struggle with the nuanced tasks required for many roles. For instance, although machine translation technology has significantly advanced, the number of human translators and interpreters has remained stable, demonstrating an increased demand for language services despite efficiency gains.

In conclusion, while AI agents are becoming a significant part of the business landscape, they are not poised to replace human workers entirely. Many companies, including Johnson & Johnson, acknowledge the importance of keeping humans involved in the workflow to mitigate risks and enhance productivity. As we move forward, it seems we are not on the brink of a robotic takeover but rather evolving into a more integrated relationship between humans and machines.

Shubham Agarwal is a freelance technology journalist based in Ahmedabad, India, whose contributions have been featured in reputable publications such as Wired, The Verge, and Fast Company.