How agentic AI is changing IT

Automation with agentic AI is reshaping the world of IT in profound ways. To understand the changes, Fixify dug into the data. We found that a new division of labor is emerging among early IT automation adopters: AI agents are handling more and more of the routine execution of IT tasks, while analysts are supervising, approving, rejecting, redirecting, and taking over when requests require judgment. AI is emerging, not as an autonomous replacement for IT work, but as a structured coordination and execution layer for getting work done.

Our findings are based on 17,929 agentic plans (the plans an AI agent produces for a request), 147,351 plan actions (the individual steps inside a plan), and 52,689 skill executions (the discrete actions an agent carries out) from March through June 2026. The data comes from more than 40 companies ranging in size from fewer than 100 employees to more than 2,000 across a range of industries including technology, financial services, and healthcare. All of these organizations are early adopters of agentic AI for IT, so their experience gives us signals, but we know there’s a lot of change still ahead. What we can see most clearly now is where the shifts are happening quickly, when agentic AI works and when it breaks, and how the role of IT analysts is evolving — and why it will remain critical.

Here’s the big picture: Our research shows that agentic AI is becoming a new operating layer for IT. It plans, proposes, communicates, waits, branches, executes, and replans. But the analyst remains the governor of the system. It’s a human-supervised model where AI agents handle repeatable work while people focus on the decisions that matter most.

Let’s take a deeper look at some of the key findings and what they may mean for the future. For the most curious among us, we’ve also included a methodology note and a glossary.

AI agents handle a substantial share of the execution when work is repeatable, structured, lower risk, and easiest to reverse, including software and application requests, security tasks, and collaboration changes. But even in these categories, agents still handle well under half of the actions.

Meanwhile, the work that stays mostly in people’s hands is onboarding and offboarding (23% AI-executed), IAM (29%), and hardware (27%). These involve identity-lifecycle and physical work where a wrong action is hardest to undo.

Here’s one big change on the horizon for IT analysts: supervising AI agents will become a big part of the job. When an agent proposes an action, an analyst reviews the proposal and decides whether it runs. Across our dataset, analysts approved most proposals and declined about 23% of them.

Staying in the approval loop lets IT analysts accomplish two important things. First, it keeps the most consequential changes under human control. For example, you probably don’t want AI automatically resetting the CEO’s password. Second, the IT analyst approvals help train AI agents: each approval or rejection teaches the model which actions fit a given situation, so the agents get better at proposing actions a supervising analyst will accept. Our data shows the impact of that learning curve. Over four months, the rejection rate declined from 27% to 16%.

Open up the work an agent does on a request and you’ll find six kinds of actions. Only one of them — running a skill — actually changes a system. (You can think of a skill as a discrete action the agent can carry out in a connected tool, such as resetting a user’s multi-factor authentication in Okta, adding someone to a Google group, assigning a software license, or unlocking an account.) The other five action types allow the agent to coordinate that work: messaging the requester, leaving notes for the team, instructing an analyst, waiting, and occasionally running a packaged workflow.

Running a skill accounts for about 39% of all actions. The rest is communication (41%, split between messages to the requester and internal notes), analyst hand-offs (10%), waiting (9%), and workflows (1%). The system-touching work is the minority; most of an agentic plan is the coordination and communication around it.

Like all of us, when an AI agent builds a plan it doesn’t know what it doesn’t know. To navigate the unknown, AI agents build in far more scenarios than they execute. A typical plan lays out 15 possible actions but runs only 2 — because most of those actions are conditional branches for situations that might arise. 72% of plans include at least one branch, and 40% of all blueprinted actions are conditional.

The gap between possible scenarios and executed ones is a key strength for automation with AI. The agent maps the paths a request could take, then walks down the path that makes the most sense as it meets reality. Most of the plan stays unused on any given run, just as the untaken branches of a decision tree go unused.

Structuring work into plans pays off in reliability. In our dataset, skills that ran inside an agentic plan succeeded 93.7% of the time; skills that ran as one-off, ad-hoc actions succeeded 87.5% of the time.

Note that the success of agentic plans versus one-off approaches is impacted by the types of problems handled by each. Structured agentic automation plans are usually tackling simpler problems, while harder or more novel work is more likely to be handled with one-off skill executions. Even so, the difference in success rates is consistent, and it holds across the window.

In our study, there were a few key areas in identity and access management where AI agents were most productive. Group and channel membership is by far the largest piece — about 77% of all changes — followed by credential resets like MFA and passwords (15%) and account lifecycle work (9%). And these changes succeed far more often inside a structured plan (84% success rate) than when run as one-off skills (71% success rate) (see Section 2 for details about agentic AI automation plans and why they work).

They’re also highly concentrated. About 84% of all changes in our data ran through Okta, the identity provider, with most of the remainder in Google Workspace.

When we look at these changes by type, most fall into three categories. Group and channel membership is the bulk of the work — adding and removing people from the groups and channels that grant access. Credential resets come next: MFA resets, password resets, and account unlocks. Account and app lifecycle, including provisioning and deprovisioning accounts and assigning application licenses, makes up the remainder.

Across all skills, about 1 in 10 executions in our data failed (9.6%). The failure rates were highest in identity-lifecycle work — onboarding and offboarding (18.5%) and IAM (13.5%).

Hardware and connectivity changes rarely fail; identity-lifecycle changes fail three-to-nine times as often. The pattern isn’t random. It reflects issues with data AI agents depend on.

The most common reason for a skill failure is relatively simple. The thing an AI agent is trying to act on — the user, group, or account — isn’t where the agent expects it. Target-not-found accounts for 48.5% of all failures.

When we also take the second most common failure, input that doesn’t match what the system expects (29%), into account, we can see that roughly three-quarters of all failures come down to one challenge. The agent’s picture of identity doesn’t match reality. People change teams, accounts get renamed, groups get restructured, and the directory the agent reads lags behind. This is an identity-hygiene problem more than an AI problem: the cleaner and more current the identity data, the fewer of these failures occur.

Conventional wisdom says automation creates complexity: more rules, more edge cases, more scaffolding over time. But our data showed the opposite. Over time, agentic plans got leaner, and the AI was allowed to execute more of them. The chart below puts both trajectories side by side.

Plans created later in our study’s window were measurably simpler. The median plan went from 18 actions to 13, and the share of actions that were conditional — the “if/then” branches a plan builds in to hedge — fell from 48% to 25%, with the sharpest drop at the end of the window. Over time, we saw fewer steps and far less branching.

The obvious question is whether the mix of work simply changed — maybe later weeks just had easier requests. This wasn’t the case. The decline holds within the same use cases (login troubleshooting, app assignment, and the rest) and when we hold the overall work mix fixed. Plans aren’t getting simpler because the work gets easier. They’re getting simpler because the agent is improving.

At the same time, the division of labor shifted. Of the actions actually carried out, the share AI executed climbed from 23% to 41%. In parallel, the rate at which analysts rejected an AI-proposed action fell from 27% to 16%. The system was handed more, and humans second-guessed it less.

Leaner plans are exactly what you’d expect as the underlying technology matures. Early on, an agentic planner isn’t sure which procedure will work, so it applies hedges. It writes branches for cases that may never occur. As the system grounds more reliably in an organization’s real procedures, pulls in more context about the environment it’s working in, and gets better at judging when to run a known step rather than plan around it, that hedging falls away. The plan converges on what the task actually requires.

We observed this trajectory across the time of our study. The whole population of plans got leaner over 13 weeks as the technology improved. Since our window was short, we read it as a strong early signal rather than a settled law.

Before agentic automation, much of IT work was step-by-step execution: look up the requester, check the policy, find the right group or app, make the change, update the ticket, message the requester. Agentic automation changes that sequence. The agent can assemble the plan, propose the next action, communicate along the way, and carry out routine steps after approval.

In this model, the IT analyst takes an active role in controlling how the IT agent behaves. The analyst decides whether the agent has the right target, whether the request fits policy, whether the proposed action is reversible, and whether the situation has moved outside the pattern. These approve-and-decline decisions help get tickets closed and they provide a signal that teaches the system how to handle the same situation next time.

The result is a role shift: less time spent repeating known steps, more time spent reviewing proposed actions, handling exceptions, and improving the workflows agents rely on. Routine execution moves toward the agent. Judgment, escalation, and system improvement stay with people.

The requests that need the most analyst attention aren’t necessarily the least automated ones. They’re the requests where reality keeps diverging from the plan.

Section 4 showed that as agentic AI’s plans got leaner, replanning rose: replans per plan increased from about 0.34 to 0.57 over the window, while roughly 7 in 10 plans never replanned at all. That combination matters. Replanning isn’t constant churn. Most plans still run without it. But when a request keeps forcing course-corrections, it’s a useful signal that the work may require human judgment.

For IT teams, that makes replanning an operational marker. A single replan can indicate healthy adaptation. Repeated replanning can mean a request was ambiguous, the underlying data is stale, the policy is unclear, or the situation simply does not fit the standard path. This is where the analyst’s attention belongs.

There was one pattern we expected to find, but didn’t. When we looked at the number of end users per IT staff member, the most stretched IT teams weren’t necessarily the most automated.

What we found instead was a relationship between organizational size and AI automation. Larger organizations automated a higher share of their performed actions. Grouping organizations into three batches by employee count, the median AI-executed share was 26.4% in the smallest third and 41.9% in the largest. The mean moved in the same direction, from 27.3% to 37.5%.

While we found this relationship between the size of an organization and its level of automation notable, there were exceptions. The largest organization in our dataset automated less than the overall size trend would predict. But the direction is consistent, and the explanation is likely more about operating maturity than staffing pressure.

Larger organizations tend to have an ideal foundation for AI automation. They have more request volume for agents to learn from, more documented processes for agents to follow, and more resources to invest in integration, configuration, and tuning. Automation maturity appears to grow less from being understaffed and more from having the foundations that let agents operate effectively at scale.

Three further data tables appear in sidebars: agentic versus ad-hoc reliability by month (Section 2), the signature-stat summary (Section 4), and conditional actions versus replans by month (Section 4).