AI Agents Make Proposals, Humans Make Final Decisions

While AI is taking on a significant portion of model development work, humans still hold the final judgment. A research team that conducted a detailed analysis of its own AI model development records revealed this reality. The team examined logs from 769 tasks to investigate how autonomously AI agents operate.

Analysis results showed that AI agents handled up to 55% of method proposals. In other words, AI generated more than half of the ideas about "what to do next." Meanwhile, humans made over 85% of the final decisions to "adopt this approach." The data clearly demonstrates a significant gap between proposals and decisions.

Notably, the research team concluded that approximately one-third of all tasks would not have been attempted at all without AI agent support. This indicates that AI does more than just streamline work—it enables efforts that humans alone could not reach, demonstrating substantial quantitative contribution.

In recent years, AI research and development environments have seen growing adoption of "AI agents" that autonomously handle multiple tasks. As large language models (LLMs) improve in accuracy, more organizations are delegating tasks such as data collection and experimental planning to agents, and this has raised expectations for improved development efficiency. This research can be positioned as a systematic documentation of this field's current state.

However, the research team cautions against conflating activity volume with autonomy. An increase in the number of tasks involving AI agents does not necessarily mean AI is doing more of the "thinking and acting independently," the team points out. The roles of making proposals and making final decisions remain clearly separated.

These findings offer important implications for designing AI development management systems. In environments where automation is advancing, the misconception "AI will handle it automatically" easily arises, but concrete data now supports the reality that human judgment remains a persistent bottleneck. For organizations considering expanded adoption of AI agents, this underscores the importance of role design—determining what to automate and what humans should handle.

Going forward, discussions are likely to intensify about how and whether the scope of decisions made by AI—beyond just proposals—should expand. The accumulation of empirical studies like this one can serve as evidence for determining the appropriate scope of autonomous AI deployment.