AI Agents: Control Over Autonomy as the Key to Success

As the deployment of enterprise AI agents accelerates, the assumption that "higher autonomy leads to better results" is beginning to collapse in actual operational environments. Research firm Gartner predicts that over 40% of AI agent projects operational in 2026 will be discontinued before reaching 2028. The primary cause of failure is not insufficient model performance, but rather increased costs, unclear business value, and inadequate risk management.

This situation becomes even clearer when considered against the trends of the past two years. From 2024 to 2025, competition among enterprises focused on "how quickly can we deploy more autonomous agents." However, from 2026 onward, this axis is shifting. What enterprises are now competing on is not the sophistication of agent functionality, but their ability to obtain approval from risk management, legal, and compliance departments for production deployment and maintain that status continuously. This represents a type of engineering challenge that many enterprises had not previously anticipated.

McKinsey's "AI Trust Maturity Survey" conducted in 2026 points in the same direction. While AI agent adoption is accelerating across all industries, the maturity of "responsible AI" remains at an average of 2.3 on a 4-stage scale. Only approximately 30% of organizations have achieved maturity level 3 or above in governance and AI agent control. In other words, the expansion of capabilities significantly outpaces the establishment of management systems.

Structural problems are also emerging from a technical perspective. According to Gartner's analysis, while thousands of products are marketed as agent AI, only approximately 130 products actually possess autonomous capabilities, with the vast majority being existing automation tools or chatbots merely rebranded. Furthermore, even in genuinely autonomous agent systems, autonomy and accountability are inherently contradictory. The more autonomous an agent is in executing multi-step processes, the more difficult it becomes to retroactively trace which decision was problematic when something goes wrong.

This challenge is particularly acute in domains such as financial reconciliation, compliance operations, manufacturing quality checks, and medical document creation. In these fields, the ability to explain decision-making processes after the fact—that is, "traceability"—is a prerequisite for operations. The dilemma that higher autonomy makes explanation more difficult is the fundamental reason why approaches designed to grant agents broad and ambiguous authority prove unworkable in actual operations.

What is common among enterprises actually achieving results is that they restrict the authority granted to agents and clearly define scope and operating rules. Rather than designing universal, high-flexibility agents, they design agents with specific responsibilities that operate within defined boundaries—this direction has emerged as a practical solution at present.

The nature of competition surrounding AI agents is changing. For enterprises that have pursued only functional excellence and deployment speed, the new challenge of establishing governance, risk management, and accountability has entered a phase where it carries equal or greater weight than technical considerations. Going forward, how to design agents for "approvability" and "sustained trustworthiness" will become a critical perspective determining success or failure of deployment.