As AI transforms organizations, the future of enterprise AI is going to drastically evolve. Gone will be the days when a company simply deploys a single chatbot; the future of enterprise AI will be more like a pool of specialized AI coworkers. But some of the challenges associated with a human workforce will still remain.
In fact, according to a new study from NTT Research and Harvard University’s Center for Brain Science, a successful AI organization has many of the same management challenges that can be found in a human workforce—from communication to coordination to team design.
Enterprises are accelerating the adoption of agentic AI, which has led to the deployment of AI agent teams that collaborate on tasks like software development, cybersecurity, customer service, scientific research, and business operations. Though organizations are asking themselves how they can build more capable AI models, the new study posits a different question: How should enterprises organize AI agents to maximize performance?
The research, which was presented at the AI4Good Workshop at the International Conference on Machine Learning (ICML), concludes that the effectiveness of multi-agent AI systems is dependent on organization, communication, and guidance from humans. However, the number of AI agents being deployed does not necessarily contribute to how successful these systems will be.
Businesses have been increasingly incorporating agentic AI into their operations. Oftentimes, they aren’t just settling for a single AI assistant, but instead experimenting with specialized AI agents that are designed to divide work, share information, and collectively solve problems. While this approach promises greater autonomy and scalability, it also introduces the complexity of how to manage the way AI systems interact with each other.
To better understand these dynamics, researchers examined how AI agent populations make group decisions. This goes beyond evaluating how single language models work in isolation. Instead, concepts from physics, mathematics, and machine learning were used to study how collective intelligence emerges when AI agents collaborate.
One way to illustrate this challenge was the study’s “Flag Game” experiment. Each AI agent is given a small, randomly-assigned section of a country’s flag with no information that would allow it to identify the correct answer. This forced the group to exchange observations and work toward coming to a shared conclusion.
Initially, including additional agents improved the group’s performance by increasing the amount of available information. However, after the optimal size had been reached for the task, additional agents actually decreased the group’s efficiency. Communication became more complex, competing interpretations emerged, and consensus became more difficult to reach. As a result, after the Flag Game had more than about 16 AI agents, accuracy began to decline.
The challenges the researchers discovered mirrored what is experienced when managing staff in a growing organization. Hiring more people doesn’t translate into greater productivity if there are communication breakdowns and decision-making bottlenecks. This same principle applies to AI organizations.
“As organizations begin deploying hundreds or even thousands of AI agents, one of the most important questions becomes how collective intelligence emerges from their interactions,” explained Dr. Hidenori Tanaka, Group Leader of the Physics of Artificial Intelligence Lab at NTT Research and the Physics of Intelligence Program at Harvard University’s Center for Brain Science.
Tanaka says that simply adding AI agents doesn’t automatically improve performance. Communication becomes exponentially more difficult among growing AI organizations because competing interpretations can emerge and collaboration requires greater coordination. This means enterprises should not solely focus on the number of AI agents they deploy, but also how those agents communicate, how they’re organized, and how human oversight shapes the decision-making that leads to the best outcomes.
The study identifies several factors that influence the success of multi-agent AI systems. Organizational design is as significant as scale, since communication patterns and information-sharing impact collective performance. Organizations do benefit from using multiple AI agents, with researchers finding that teams made up of diverse AI models consistently outperformed groups built around a single model. Enterprises benefit from combining complementary capabilities, rather than standardizing one AI system for every task.
These study findings highlight how, in the next phase of enterprise AI, orchestration will be as relevant as model selection. As organizations move away from isolated AI agents to coordinated AI workforces, CIOs and Chief AI Officers will need to define the ways agents share information, the division of responsibilities, the applications of human judgment and the governance frameworks that prevent conflicting information that slows decisions and creates unnecessary risk.
The research reflects the broader mission of NTT Research’s Physics of Artificial Intelligence Lab, which seeks to understand the scientific principles underlying intelligence itself. Rather than focusing solely on building larger or more powerful AI models, the lab studies how intelligent systems learn, reason, communicate, and collaborate to develop the scientific foundation for more trustworthy and scalable AI systems.
This study is a good reminder that AI scaling isn’t only about deploying more agents. Effective AI organizations, much like human ones, must have a thoughtful structure, diverse expertise, clear communication, and strong leadership. The design of AI organizations may be just as important as the models that power them.
