AI agents get better at IT ops, but only with humans in the loop

AI agents are increasingly taking on a larger share of routine IT operations tasks, but they still require human oversight and approval for the most critical actions, according to a new study of human-AI interactions in enterprise IT workflows. Fixify, an automation platform provider, analyzed tens of thousands of interactions involving AI agents and human analysts across 40 companies over a three-month period.

The study found that AI agents are currently responsible for about one-third of IT actions, with human analysts primarily handling high-stakes areas like identity verification, onboarding and offboarding, and hardware environments. However, human approval rates for AI-proposed actions have improved, rising from 23% to 41%, while rejection rates fell from 27% to 16%, indicating that AI is becoming more accurate and trustworthy over time.

Fixify identified four stages in the agentic work process: planning, proposing, approving or declining, and acting on approved steps. Most tasks that AI agents handle are well-understood and low-risk, such as software, application, security, and collaboration work. In these areas, AI excels at automating repeatable and reversible tasks.

Human analysts remain closely involved in more complex and high-stakes areas, such as identity verification and hardware setups. AI agents are also building "scaffolding" that guides IT operations by mapping out multiple possible scenarios and selecting the most appropriate course of action. However, agents often fail when they cannot find the necessary data or resources, such as when users or groups are moved to different teams or accounts are renamed.

This often points to issues with data cleanliness and currentness rather than AI inefficiencies. As AI agents continue to learn and adapt through iterative rejection and approval cycles, their plans become leaner, and they start to re-plan when conditions change. This adaptability, combined with human supervision, allows AI agents to gradually take on more responsibility in IT operations, with humans still controlling high-impact decisions and exceptions.

To effectively integrate AI agents into IT workflows, organizations should focus on investing in clean identity data, building strong review workflows, and establishing reliable integrations. Teams should evaluate agentic tools based on their supervision loops and view rejections as opportunities for AI training rather than failures.

By leveraging AI agents as a support system for human analysts, organizations can gradually transition towards more efficient and automated IT operations while maintaining human oversight for critical tasks.