The Evolution of IT Operations: How Agentic AI is Redefining the Human-Machine Partnership

The promise of artificial intelligence in the enterprise has long been framed as a binary choice: either the machine replaces the human, or the human remains the sole operator. However, a comprehensive new study of nearly 150,000 real-world agentic actions suggests a more nuanced reality. AI is not "replacing" IT departments; it is scaffolding them, handling the repetitive, low-risk groundwork while human analysts evolve into supervisors, strategists, and final arbiters of consequence.

Data from Fixify, an automation platform provider, provides a granular look at this transition. By analyzing over 18,000 distinct operational plans and 147,000 individual actions across 40 enterprises, the study reveals that while AI agents are increasingly embedded in enterprise IT workflows, their role is defined by iterative human guidance and a clear division of labor based on risk.

The Architecture of Agentic Work: Main Findings

The study identifies a distinct operational rhythm in modern IT departments. AI agents are currently executing approximately one-third of all IT tasks. These tasks are overwhelmingly concentrated in areas that are "repeatable and easy to reverse," such as software license management, routine application maintenance, security group updates, and collaboration tool administration.

The methodology of these agents follows a four-step framework:

1. Planning: The agent maps out potential paths to resolve a ticket or request.
2. Proposing: The agent presents a recommended course of action to a human analyst.
3. Approving/Declining: The human analyst reviews the proposal, acting as a gatekeeper.
4. Acting: The agent executes the approved steps or pivots based on human feedback.

Crucially, the study found that agents are highly efficient planners but conservative actors. On average, an agent might map out 15 potential scenarios for a given request but will only execute two. This "scaffolding" approach allows the system to prepare for multiple contingencies while only committing to actions that align with the current reality of the IT environment.

Chronology of Adoption and Refinement

Over the three-month period of the study, the relationship between human analysts and AI agents underwent a measurable maturation. At the beginning of the period, human analysts rejected roughly 27% of AI-proposed actions. By the end of the study, that rejection rate plummeted to 16%, while the rate of approval for AI-proposed actions surged from 23% to 41%.

This improvement is not necessarily due to a sudden leap in the underlying large language models (LLMs) or neural network architectures. Instead, it is the result of what researchers call "iterative feedback loops." As humans continue to approve or reject actions, they are effectively training the system, refining its ability to navigate the specific context of their organization.

The study tracks this progression through the lens of sophistication. Early in the adoption phase, agents relied on "pre-scripting"—attempting to anticipate every possible contingency before taking a single step. As they matured, the agents shifted toward "in-the-moment adaptation." They began to re-plan dynamically when conditions changed, a behavior that indicates a move toward genuine agentic intelligence rather than simple rule-based automation.

Supporting Data: Where AI Succeeds and Where It Stumbles

The data provides a clear map of the current limitations of AI in the enterprise. When agents are tasked with "running a skill"—physically performing a task—they are highly effective, accounting for nearly 40% of their activity. A significant portion of the remaining work is purely coordinative: sending messages, leaving comments, or routing tasks to the appropriate human expert.

However, the study highlights a stark divide in reliability. While tasks related to hardware and connectivity are rarely problematic, "identity-lifecycle" tasks—such as onboarding, offboarding, and complex identity-access management (IAM)—are prone to failure at a rate three to nine times higher than other IT functions.

The Anatomy of Failure

The research team identified the primary drivers behind these failures, which occurred in roughly 23% of cases where the AI’s recommendation diverged from human judgment:

• Target Not Found (50%): The most common failure, often stemming from poor data hygiene. When user groups are restructured, accounts renamed, or tickets processed without being reflected in the source system, the agent cannot find the target. This is a "data truth" problem, not a failure of intelligence.
• Invalid Inputs (29%): Errors stemming from malformed requests or incomplete data passed to the agent.
• Systemic Breakage (21%): This includes denied permissions, unhandled errors, or misconfigured integrations that indicate a fundamental break in the IT environment’s connectivity.

These findings suggest that the bottleneck for AI maturity is not the agent’s "brain," but the quality of the organizational data it is forced to interact with.

Official Responses and Strategic Outlook

Matt Peters, co-founder and CEO of Fixify, emphasizes that the goal of this technology is not to create a "world-ending hive mind" but to provide actionable intelligence that empowers human teams.

"That may sound less dramatic than replacing the help desk," Peters noted in a recent blog post accompanying the study, "but it is a much more credible path to changing how IT work gets done."

Peters argues that the most effective way to utilize these tools is to "point a little conversational intelligence in the right direction." By allowing the AI to handle initial triage, ask clarifying questions, and route tickets to the correct departments, IT teams can save hundreds of hours of manual labor without surrendering the high-stakes decisions—such as security overrides or permanent access changes—to an automated system.

Implications for the Future of IT

As agentic AI moves from experimental to foundational, the implications for enterprise workflows are profound. The study outlines several key areas where organizations must adapt to maximize the return on these tools:

1. The Shift to "Supervisory" Roles

The role of the IT analyst is moving toward that of a supervisor. Future job descriptions will increasingly focus on the ability to interpret AI proposals, manage feedback loops, and handle complex exceptions. The "review surface"—the interface where humans approve or decline actions—has become the most important piece of software in the IT stack. If this surface is confusing or poorly designed, the entire efficiency gain of the AI is negated.

2. Investing in Data Hygiene

Enterprises that wish to scale their use of agentic AI must invest in "identity hygiene." Because the primary point of failure is "target not found," the quality of the company’s directory services, account management databases, and ticketing systems is now a direct constraint on automation capacity. Clean, real-time data is the prerequisite for autonomous success.

3. Redefining Success Metrics

Organizations should stop measuring success solely by "tickets closed" and begin measuring it by "supervision quality." Rejections should be viewed as training data rather than operational failures. A high rejection rate early in the process is a sign that the system is learning the nuances of the organization’s unique environment.

4. Governing Autonomy

The study highlights that the hardest requests—those requiring deep contextual judgment or repeated replanning—will likely remain human-led for the foreseeable future. "Governed autonomy" is the path forward: allow the AI to build the scaffolding and handle the routine, but maintain rigid human control over the "high-impact" actions.

Conclusion: A Collaborative Future

The era of the "all-or-nothing" AI deployment is ending. The data makes it clear that the most successful IT departments are those that treat AI as a collaborative partner. By allowing the AI to map out paths and suggest solutions while keeping the human analyst in the driver’s seat for final approval, enterprises are achieving a higher level of operational stability than they could through either manual labor or full-scale, unmonitored automation.

As these systems continue to iterate, the "scaffolding" will become more sophisticated, the agents will become more adept at dynamic replanning, and the human role will become more focused on high-level architecture rather than day-to-day administrative firefighting. The future of IT is not a machine-run department; it is a human-led, AI-accelerated ecosystem.