AI agents get better at IT ops with humans in control

AI agents now handle one-third of IT actions, but human control remains essential

AI agents already perform roughly one in three actions in enterprise IT workflows. That is meaningful automation, but it is not autonomous IT. Fixify’s study shows a more practical operating model: AI handles routine execution, while people retain control over decisions where mistakes carry greater cost.

Fixify analyzed nearly 18,000 agent plans and more than 147,000 agent actions across 40 companies over three months. The data shows a clear division of work. Agents execute repeatable tasks and coordinate requests. Human analysts approve important changes, resolve exceptions, and correct agent decisions when business context matters.

The balance is also changing. During the three-month study, human approval of AI-proposed actions increased from 23% to 41%. Rejection fell from 27% to 16%. This suggests that the system became more useful as agents operated within feedback and review processes. It does not show that human oversight became unnecessary. It shows that supervised automation can earn a larger role over time.

Matt Peters, co-founder and CEO of Fixify, frames this as a more realistic goal than replacing an IT help desk. “That may sound less dramatic than replacing the help desk,” Peters wrote. “It’s also a much more credible path to changing how IT work gets done.”

For executives, the important metric is therefore not the percentage of IT jobs that AI can replace. It is the percentage of work that can move safely from human execution to human supervision. These are different objectives. The second can deliver faster service and greater capacity without requiring an agent to make every decision independently.

This also changes the role of IT staff. Analysts increasingly need to inspect proposed actions, manage exceptions, and decide when automation should stop. Their judgment becomes concentrated on fewer but more consequential decisions. Organizations should design workflows, staffing, and performance metrics around this division of labor rather than measure AI success primarily by full autonomy.

The evidence supports a clear operating principle: automate execution before automating authority. AI agents can take on more work as their performance improves, but businesses should keep human control where an incorrect action can materially affect security, access, employees, or operations.

The best automation targets are repeatable, reversible, and low-risk IT tasks

The main constraint on IT automation is not whether an AI agent can perform an action. It is whether the organization can tolerate the agent being wrong. Fixify found that current agents work best when requests are well understood, repeat frequently, and can be reversed without serious consequences.

Software, application, collaboration, and some security operations fit this profile. An agent can process standard requests, gather information, send messages, and execute established procedures. These tasks provide clear rules and observable outcomes. If something fails, the organization can often correct it without creating a major business impact.

Higher-stakes work is different. Human analysts remain closely involved in identity verification, onboarding and offboarding, access provisioning, and hardware environments. These activities may look routine, but they can grant a person access to sensitive systems, remove access needed for work, or leave former employees with permissions they should no longer have. The cost of an incorrect action therefore matters as much as its technical complexity.

Security illustrates this distinction. Agents participated in only about 6% of the security work examined by Fixify. Their activities included adding or removing people from groups and channels, unlocking accounts, resetting passwords, analyzing multi-factor authentication issues, provisioning or deprovisioning accounts, and assigning software licenses. These are technically automatable operations, but many change an organization’s security state.

Identity-lifecycle work was also where agents struggled most. Fixify found that identity-related changes failed three to nine times as often as hardware and connectivity changes. This reinforces why automation scope should be determined by both reliability and consequence.

The executive decision is therefore not simply where AI can be deployed. It is where errors can be detected, contained, and reversed at acceptable cost. Low-risk and reversible work should move first. Higher-risk processes should use agents for preparation, information gathering, and recommendations while preserving explicit human approval for consequential changes.

This approach still creates substantial room for expansion. As data improves and organizations learn which actions agents perform reliably, more tasks can move into supervised automation. The objective is controlled delegation: increase the amount of work AI executes while keeping authority aligned with business risk.

Most AI agent activity is coordination and planning

Only 1.1% of the actions Fixify analyzed involved running an entire workflow. That figure sets an important expectation for executives. Enterprise AI agents are already useful, but their value today comes mainly from handling individual steps, coordinating work, and deciding what should happen next.

Fixify divides agentic work into four stages: planning, proposing an action, receiving human approval or rejection, and executing an approved action. This structure allows an agent to assess a request before changing a production system. It also gives human analysts clear points where they can intervene.

The study shows how much work happens around execution. Running a specific skill, meaning actually performing a task, accounted for 39.4% of all agent actions. Sending a message to the person making the request represented 27.7%. Initial comments accounted for 13.2%, instructions to human analysts for 9.8%, and waiting for 8.8%. Complete workflow execution was just 1.1%.

Planning is similarly extensive. Fixify found that agents typically mapped around 15 possible actions but ultimately executed only two. This is deliberate. IT requests often change as new information becomes available. An agent may discover that an account does not exist, a request requires approval, or another team owns the necessary system. It must then select a different action rather than blindly complete a predefined sequence.

Fixify describes this process as the agent mapping possible paths and then choosing the most appropriate one as it encounters real operating conditions. The study also found that more sophisticated agents eventually produce leaner plans and re-plan when necessary instead of predicting every contingency in advance.

This matters when companies measure returns. A narrow focus on completed autonomous workflows will understate what agents are doing. An agent that collects missing information, checks system state, communicates with an employee, and prepares the correct action for analyst approval can reduce significant manual work without ever owning the complete process.

Executives should therefore measure individual work removed as well as workflows automated. Useful metrics include analyst time saved, number of manual steps eliminated, approval rates, rejection rates, execution success, and the frequency of replanning. These measures provide a clearer picture of operational value than the percentage of processes that run without human involvement.

The near-term opportunity is not maximum autonomy. It is reducing the coordination and routine execution that consume analyst capacity. End-to-end automation can expand where evidence supports it, but it should be an outcome of proven reliability rather than the starting requirement.

Selective automation can improve IT service without handing complete control to AI

AI does not need to complete an entire IT process to produce measurable value. It can handle the early stages of a request, gather the required information, and send the case to the right expert. This reduces manual work while preserving human control over decisions that need business context or carry higher risk.

Fixify described a high-volume ticket process that demonstrates this model. An AI agent determined which team needed access to handle the request. It performed initial triage, asked follow-up questions, and routed the ticket to the team with the information and authority required to act. The agent did not attempt to automate the full process.

The important result was better preparation. The receiving team could act sooner because basic discovery and routing had already happened. This addresses a common source of IT service delay: not the technical work itself, but the time spent collecting information, clarifying a request, finding ownership, and moving tickets between teams.

Matt Peters, co-founder and CEO of Fixify, summarized the case in simpler terms: “We didn’t need a world-ending hive mind,” he said. “We just needed to point a little conversational intelligence in the right direction.” His point is operational. Useful automation does not require a system capable of independently managing every stage of a request.

For business leaders, this creates a practical way to prioritize investment. Companies can automate high-volume steps where the rules are clear while keeping approval gates around sensitive changes. Initial triage, information gathering, classification, communication, and routing are strong candidates because they can consume substantial analyst time without necessarily requiring final decision authority.

This approach also limits the cost of errors. If an agent makes an incorrect recommendation during preparation, a human can reject or correct it before a consequential action occurs. Fixify’s wider findings support this design: human analysts remain most involved in higher-stakes areas, while agent activity is concentrated in repeatable work that is easier to reverse.

The business objective should be less unnecessary human work. Selective automation can shorten queues, improve the information available to analysts, and allow skilled staff to concentrate on exceptions and consequential decisions. As agent reliability and operational data improve, organizations can expand the boundary of automated execution based on observed performance rather than assumptions.

Identity and access work is where AI agents fail most often

Identity-lifecycle changes fail three to nine times as often as hardware and connectivity changes, according to Fixify. This makes identity and access management (IAM) one of the clearest areas where enterprises should retain strong human controls.

The problem is not simply task complexity. Identity operations depend on accurate relationships between employees, accounts, groups, applications, permissions, and organizational structures. Those relationships change constantly. Employees join or leave. People move between teams. Groups are reorganized. Accounts are renamed. Access requirements change. An agent acting on outdated information can make a technically valid decision based on an incorrect view of the organization.

This matters because identity errors can have significant consequences. Giving access to the wrong account can create a security exposure. Removing legitimate access can disrupt work. An incomplete offboarding process can leave former employees with permissions that should have been revoked. The action itself may be simple, but its authorization and business context require greater confidence.

Fixify found agents participating in security work, although security represented only about 6% of their activity. Tasks included adding or removing users from groups or channels, unlocking accounts, resetting passwords, analyzing multi-factor authentication issues, provisioning and deprovisioning accounts, and assigning software licenses.

These activities should not all receive the same automation policy. Password resets and routine account operations may be suitable for higher levels of automation when identity can be verified reliably. Privileged access, sensitive group membership, and employee offboarding require stricter controls because an incorrect decision has a larger security or compliance impact.

For executives, the main constraint is the quality of the organization’s identity system. Increasing the capability of the AI model will not correct inaccurate account ownership, obsolete groups, inconsistent employee records, or poorly documented access policies. Automating those processes faster can simply expose existing weaknesses more frequently.

The practical priority is therefore to improve identity governance before expanding autonomous access changes. Enterprises need current employee records, clear ownership of groups and systems, defined access policies, dependable approval rules, and reliable connections between identity platforms and other enterprise systems.

Human review should remain concentrated on actions with the highest impact. As organizations improve identity data and demonstrate consistent agent performance, they can automate more routine changes. The evidence from Fixify suggests that expansion should be based on measured reliability by task type, not on a broad assumption that all IT operations are equally suitable for AI.

Bad operational data is the largest source of AI agent failure

Nearly half of the failures identified by Fixify were classified as “target not found.” This is a direct warning for executives investing in agentic AI: automation quality depends on the state of the systems and data the agent is expected to use.

Fixify could examine this problem because its workflows kept humans in the approval process. Agent recommendations diverged from human judgment about 23% of the time. The company then analyzed why proposed actions failed or were inappropriate.

The largest category, representing nearly 50% of failures, occurred when an agent could not locate the user, account, group, or other resource it expected to find. Fixify linked these failures largely to operational data quality. Employees may have changed teams, accounts may have been renamed, groups may have been restructured, or work may already have been completed without the relevant system being updated.

Invalid inputs represented about 29% of failures. Other problems included unhandled errors, denied permissions, invalid operations, and incorrect configurations. Fixify said these latter categories can indicate genuine failures in integrations rather than poor agent reasoning.

This distinction is important for technology investment. A company can deploy a more capable AI model and still see poor results if the agent receives stale records or calls unreliable enterprise systems. The limiting factor in these cases sits below the AI layer. The underlying operational environment does not provide a sufficiently accurate and consistent state for automated action.

Leaders should therefore separate agent errors from infrastructure and data errors when reviewing performance. A rejected proposal caused by incorrect reasoning requires a different response from a failed action caused by a missing account, obsolete identity record, denied API permission, or broken integration. Combining these failures into one AI accuracy metric makes it harder to identify where investment is actually needed.

The Fixify findings also change the business case for data quality. Cleaning identity records and maintaining system integrations are not merely administrative tasks. They become prerequisites for scaling automation. An agent must know which resource exists, who owns it, what state it is in, and what actions are permitted before it can operate reliably.

Executives should require operational dashboards that expose failure causes rather than only overall success rates. “Target not found” errors should trigger work on identity hygiene and system records. Invalid inputs should lead to better validation and workflow design. Permission or configuration failures should be traced to integrations and access controls. Reasoning errors should inform agent instructions and review policies.

The priority is to fix the underlying constraint before expanding automation. Better models can improve planning and judgment, but they cannot reliably compensate for enterprise systems that describe the business incorrectly. Clean, current operational data is therefore part of the AI infrastructure itself, even when it sits outside the AI platform.

Human feedback is what allows AI agents to take on more work safely

Human approval of AI-proposed actions rose from 23% to 41% during Fixify’s three-month study. Rejection fell from 27% to 16%. The direction is important. Agents became more aligned with the actions analysts were willing to approve, while humans became more comfortable delegating work.

Fixify links this improvement to feedback loops built into hybrid human-AI workflows. Agents plan and propose actions. Analysts then approve or reject those proposals before consequential changes occur. These decisions provide signals that can be used to improve agent behavior and the surrounding automation system.

The study also identified a change in how agents plan. Less mature behavior involves preparing many possible actions in advance, including scenarios that may never occur. More sophisticated agents develop leaner plans and re-plan when conditions actually change. Fixify described this ability to adapt during execution as a sign of greater sophistication.

Replanning is important because enterprise IT environments are not static. The state of an account, permission, application, or request can change between the initial plan and execution. An effective agent needs to detect those changes and adjust rather than continue with an action that is no longer appropriate.

However, rising approval rates should not be interpreted as evidence that human review can automatically be removed. The study covered three months, and approval can improve for several reasons. Agents may become better aligned with analyst preferences, workflows may improve, organizations may assign agents more suitable tasks, and users may become more familiar with the system. Fixify’s results demonstrate improvement within the environments studied, not a universal rate of AI learning that every enterprise should expect.

The hardest requests also remain dependent on people. Fixify found that cases requiring repeated replanning or contextual judgment continue to be human-heavy. As agents absorb routine execution, analysts increasingly focus on exceptions, high-impact decisions, and controlling how the agents behave.

This changes how executives should think about human-in-the-loop systems. Review is part of the operating model, not simply a temporary cost to be eliminated. Analyst decisions generate information about where automation works, where policies are unclear, and where an agent lacks sufficient context.

Companies should measure this process carefully. Approval and rejection rates are useful, but they need context. Leaders should also track why proposals are rejected, which tasks require repeated replanning, how often humans modify proposed actions, and whether the same errors recur. These measures reveal whether the agent is actually improving or whether employees are repeatedly correcting the same weaknesses.

The objective is progressive delegation backed by evidence. When a specific action achieves consistently high reliability and low business risk, organizations can reduce the amount of human intervention it requires. High-impact and ambiguous work should retain stronger controls. This allows automation to expand without treating autonomy itself as the measure of success.

Scaling AI agents requires better operating systems

Enterprises cannot scale agentic IT automation by deploying more agents alone. Fixify identifies clean identity data, strong playbooks, reliable integrations, and effective human review as core requirements. These operational capabilities determine whether agents can turn their reasoning into correct actions.

The first requirement is accurate data. Agents need current information about employees, accounts, groups, applications, permissions, and resources. Fixify’s finding that nearly half of failures involved a target that could not be found shows the cost of weak operational records. An agent cannot execute reliably when the organization itself lacks a consistent record of what exists and who owns it.

The second requirement is clear process design. A playbook defines what actions are permitted, when approval is required, and how exceptions should be handled. Without explicit rules, an agent may need to re-plan repeatedly or escalate decisions that the organization has never clearly defined.

The third requirement is reliable integration. Agents act through existing enterprise applications, identity platforms, ticketing tools, APIs, and automation systems. Permission failures, invalid configurations, and broken integrations can stop a valid AI recommendation from being executed. Model quality cannot resolve an API permission that the organization configured incorrectly.

Human review is the fourth requirement. Fixify recommends designing analyst queues, workflows, and metrics specifically around reviewing agent proposals. The review interface should make proposed actions easy to understand so analysts can assess them quickly. The analyst’s limited attention should be concentrated on decisions where human judgment adds the most value.

Replanning provides another useful management signal. Fixify suggests that a single replan can indicate healthy adaptation to new information. Repeated replanning can indicate ambiguity, irrelevant information, missing context, or unclear policies. Executives can therefore use replanning frequency to identify operational weaknesses instead of treating every replan as an AI failure.

This has direct implications for investment. AI budgets should include identity-data maintenance, integration reliability, process redesign, controls, observability, and analyst review capacity. Spending heavily on models while neglecting these areas is unlikely to produce dependable automation at scale.

Governance should also become more granular as adoption grows. Organizations should define acceptable autonomy by action type and business impact. A routine, reversible action can operate with fewer controls after it demonstrates reliable performance. A change involving sensitive access, identity, or substantial operational impact should face a higher approval threshold.

The strongest measure of progress is therefore not the number of deployed agents. It is the amount of useful work they can complete reliably with appropriate supervision. Fixify’s findings point toward an operating model in which agents execute more routine work over time while humans set boundaries, approve consequential decisions, and handle exceptions. For executives, building that management system is as important as selecting the AI technology itself.

Concluding thoughts

AI agents already perform roughly one-third of enterprise IT actions. The next question is not whether that share will grow. It is whether companies can expand it without increasing operational, security, and access risk.

Fixify’s findings point to the main constraint. Reliable automation depends as much on the enterprise environment as on the AI itself. Clean identity data, stable integrations, explicit policies, and fast human review determine whether an agent can turn a reasonable plan into a correct action. Better models cannot compensate for systems that contain stale data or unclear permissions.

Executives should therefore resist making full autonomy the target. Start with repeatable, reversible work. Measure approval, rejection, failure causes, and repeated replanning. Expand automation when the evidence supports it. Keep human authority around actions where errors have material consequences.

The companies that gain the most from AI agents will not be those that remove people from IT workflows fastest. They will be those that redesign those workflows so AI handles routine execution while skilled employees focus on exceptions, judgment, and control. That is a practical path to scaling automation while maintaining accountability.