A network of personal AI agents helped 15 teams solve a simulated evidence-sharing task 19% faster than group chat with similar accuracy, researchers from Carnegie Mellon University and Unanimous AI reported in a preprint.
The study involved 105 participants organized into 15 teams. Each group worked through a case about identifying a terrorist, with relevant clues spread among members and some details in conflict or designed to mislead. Teams tried the task through both group chat and Hypercom AI.
In the AI round, each participant used an individual software intermediary that could relay information across the group. Carnegie Mellon University said the agents handled exchanges but did not add facts or draw conclusions, leaving participants to interpret the clues. The preprint gives an average time to reach a correct solution of 11.6 minutes with Hypercom AI, compared with 14.3 minutes in chat.
The authors reported a 19% decrease in decision convergence time, a difference they described as statistically significant (p = 0.023). Accuracy was 93.3% in the agent condition and 92.3% with group discussion.
Anita Williams Woolley, a professor of organizational behavior at Carnegie Mellon’s Tepper School of Business, co-authored the paper with Carnegie Mellon researcher Ganesh Mani and Unanimous AI employees Louis Rosenberg, Hans Schumann and Gregg Willcox. The paper’s funding statement names the U.S. Air Force through AFWERX; Carnegie Mellon lists contract FA864926P0228.
Woolley said the approach may help organizations combine expertise when decisions need to be made quickly. The researchers cited emergency response, disaster relief, cybersecurity, medical triage and military command as possible applications, but the reported experiment tested a simulation rather than those operational settings.
Carnegie Mellon says Hypercom AI is available in beta through Thinkscape, Unanimous AI’s platform.
