日本語 
NEC announced the development of “cotomi Act,” an agent technology for automatically executing web-based tasks.
Results from the WebArena web browser operation benchmark show that by incorporating tacit knowledge gained from the behavior of veteran employees based on operation history and logs, even non-expert employees can perform tasks with the same level of skill. An AI agent incorporating this technology surpassed the success rate of humans in the international web operation benchmark, WebArena, for the first time in the world.
In developing cotomi Act, tacit knowledge about the specialized tasks of individuals and organizations was extracted from browser operation history and logs. The collected data was analyzed using technology including NEC’s generative AI, “cotomi.” By extracting data from logs, complex and highly specialized tasks that would be difficult to perform manually, can be performed autonomously and with high accuracy. Furthermore, even when users provide ambiguous instructions, the AI agent can identify the necessary knowledge and automatically search and select information and procedures. The “WebArena” used for the benchmark was proposed by a research team led by Carnegie Mellon University.
Also Read: XerpaAI Launches World’s First AI Growth Agent for Web3
In an environment similar to actual internet usage, the AI navigates through pages and fills out forms on e-commerce sites, message boards, collaborative development, map searches, and more, assessing whether it can operate the web according to natural language instructions. WebArena measured 179 tasks that were officially evaluated by humans, and found that the success rate for humans was 78.2%, while the AI agent using cotomi Act achieved 80.4%. Conventional AI agents only achieve success rates of around 40% to 70%, making this the first AI agent in the world to surpass humans. NEC is currently conducting demonstrations of this technology within the NEC Group, with the aim of providing the service in fiscal 2026.
SOURCE: Yahoo
English