AI Agents in IT Operations: How Human-in-the-Loop Supervision Drives Success

AI agents now perform approximately one in three actions within enterprise IT workflows, according to a study by automation platform provider Fixify. While human analysts still reject about one-quarter of AI-proposed actions, that rejection rate is declining as feedback loops improve and agents handle more routine, low-risk tasks.

The study analyzed nearly 18,000 plans and more than 147,000 actions executed by agents across 40 companies over a three-month period. Fixify found that agents are most active in software, applications, security, and collaboration work, where tasks are typically repeatable and easy to reverse.

Human-AI Interaction and Approval Trends

Human analysts currently manage the most consequential actions, handle exceptions, and supervise the agentic systems. According to Fixify, AI agents primarily carry out routine executions. This division of labor creates a path toward changing IT operations without immediately replacing the help desk, according to Fixify co-founder and CEO Matt Peters.

The data shows a steady increase in AI adoption and accuracy over the three-month study period. Human approval of AI-proposed actions rose from 23% to 41%, while the rejection rate dropped from 27% to 16%.

The Four Steps of Agentic Work

Fixify identified a specific four-step framework for how these agents operate: planning, proposing, approving or declining, and acting on approved steps. The company noted that agents often build a form of scaffolding around changes, mapping out significantly more scenarios than they actually execute.

Typically, agents map 15 possible actions but only execute two. The study stated that the agent maps the paths a request could take, then walks down the path that makes the most sense as it meets reality.

Fixify categorized AI automation into six types of actions:

• Running a skill (actual execution): 39.4% of actions
• Sending messages to human requesters: 27.7% of actions
• Leaving initial comments: 13.2% of actions
• Giving instructions to human analysts: 9.8% of actions
• Waiting: 8.8% of actions
• Running entire workflows: 1.1% of actions

Failure Points in Identity and Data Hygiene

AI agents struggle most with identity-lifecycle work, such as onboarding, offboarding, and identity-access management (IAM). Fixify found that identity-lifecycle changes fail three to nine times more often than hardware or connectivity changes.

The study found that agent recommendations diverged from human judgment about 23% of the time. Analysis of these failures revealed that nearly 50% fell into a target not found category. In these cases, the agent could not locate a user, group, account, or resource because the system data was outdated or incorrect.

Invalid inputs accounted for approximately 29% of failures. Other errors included denied permissions, unhandled errors, or invalid configurations, which Fixify described as signals of real breakage in integrations.

Evolution of Agent Sophistication

As systems iterate, agents are becoming more sophisticated by leaning their plans and re-planning in real-time when conditions change. Fixify described the ability to adapt in the moment as a more advanced behavior than pre-scripting every contingency.

This evolution allows human analysts to shift their focus toward controlling agent behavior and making high-impact decisions. However, the study noted that the hardest requests remain human-heavy, particularly those requiring contextual judgment or repeated re-planning.

To adapt to these tools, Fixify advised enterprises to invest in clean identity data and reliable integrations. The company suggested that teams view rejections as a training process and design review surfaces that allow analysts to make quick, informed decisions on proposed AI actions.