Study Finds AI Agent Usage Remains Cautious
A new study from Anthropic reveals that despite the advanced capabilities of AI agents, real-world adoption remains conservative. Most user sessions are reportedly short and tightly scoped, with users maintaining close human oversight and frequently reviewing outputs. While use cases are expanding beyond coding into finance and marketing, this pattern suggests that organizational trust and process design are shaping autonomy as much as technical ability.
- The Anthropic study analyzed millions of real-world interactions and found that approximately 73% of AI agent tool calls appear to have a human in the loop, indicating a preference for oversight. - As users gain experience, their trust in AI agents increases; after roughly 750 sessions with Anthropic's Claude Code, over 40% of user sessions are fully auto-approved, up from about 20% for new users. - Experienced users, while auto-approving more, also interrupt agents more frequently than new users—about 9% of the time compared to 5% for novices—suggesting a shift to a more active monitoring strategy. - The market for AI agents is projected to grow significantly, with one forecast estimating an increase from $5.4 billion in 2024 to over $50 billion by 2030. - Major barriers to wider enterprise adoption include difficulties integrating agents with legacy ERP and CRM systems, concerns over data security and privacy, and the unpredictable, non-deterministic nature of some AI models. - Despite challenges, a 2025 survey found that 96% of enterprise IT leaders plan to expand their use of AI agents within the next year, with key interest in performance optimization and security monitoring. - While software development currently accounts for about 50% of agentic tool use, applications are expanding into cybersecurity, finance, and research. - The duration of autonomous work is increasing, with the longest, top-percentile sessions on Claude Code nearly doubling from under 25 minutes to over 45 minutes in a three-month period.
