A new report from IT automation firm Fixify found that AI agents completed about one in three actions in the IT workflows it studied, though analysts still decided whether each proposed action would run.

Among early adopters of agentic AI for IT, agents handled more routine execution, while analysts reviewed plans, approved or rejected proposed steps, redirected workflows and took over when a request required expert judgment. For IT leaders, the report offers a window into where agentic automation is working, where it breaks down and how much human oversight it still requires.

Fixify analyzed about 13 weeks of production activity from more than 40 companies using its agentic automation platform in their IT operations. The dataset covered 17,929 plans generated by AI agents, 147,351 individual steps within those plans and 52,689 executions of automation skills from March through June 2026. The companies ranged from fewer than 100 employees to more than 2,000 and included organizations in technology, financial services and health care.

The results suggest that companies are not using the same level of automation across every part of IT. Agents performed the largest share of actions involving software and application requests, security tasks and collaboration tools. Analysts were more involved in onboarding and offboarding, identity and access management, and hardware, where mistakes may be harder to undo or the work could require physical intervention.

That division of labor also changed over the course of the study. Analysts rejected about 23% of proposed actions overall, but the monthly rejection rate fell from 27% in March to 16% in June. Over the same period, the share of actions carried out by AI rose from 23% to 41%. Fixify said the two trends indicate that agents were taking on more execution as their recommendations became more aligned with what analysts would approve. The agents’ plans also became more focused, with the median falling from 18 steps in March to 13 in June and conditional steps dropping from 48% to 25%. Fixify said the pattern persisted even after accounting for changes in the types of requests.

Even as agents took on more of the work, some parts of IT were harder to reliably automate. Across the dataset, 9.6% of skill executions were recorded as failures, rising to 18.5% for onboarding and offboarding and 13.5% for identity and access management. Nearly half of the failures were classified as “target not found,” meaning a user, account or other resource was not where the agent expected it, while 29.3% involved invalid input. Fixify said these results often reflected stale or mismatched identity data, although some “target not found” results meant the requested change had already been made.

The report’s findings represent a narrow group of companies already using Fixify, not the average IT department. But it suggests there is still a practical limit to how much work companies can hand off to agents. For now, their role can expand only as far as the data, integrations and human oversight around them will allow. As agents take on more work, IT teams will need clarity around which actions can be delegated, what conditions must be met before they run and when a human in the loop is required.