AI Agents Handle 1 in 3 IT Tasks — But Human Oversight Drives Quality

What happened

A new operational study by Fixify, an automation platform provider, has found that AI agents are now handling roughly one in three actions within enterprise IT workflows — a share that is climbing steadily. The research, drawn from tens of thousands of recorded human-AI interactions, maps how agentic systems and human analysts divide labour across four distinct stages: planning, proposing, approving or declining, and executing approved steps.

The data reveals that human analysts are currently rejecting approximately one in four AI-proposed actions, though that rejection rate is falling as the systems mature. Crucially, Fixify found that when things go wrong, the root cause is more often poor operational data than flaws in the underlying AI infrastructure itself. Human analysts retain control over high-stakes decisions and exception handling, while AI agents absorb the routine, repeatable execution work.

Fixify's co-founder and CEO Matt Peters described the arrangement not as a replacement of the help desk, but as a more credible and durable model for transforming how IT work actually gets done — one built on structured human oversight rather than wholesale automation.

Why it matters

For anyone designing service operations — whether in IT, customer support or broader shared services — this study offers a rare, data-grounded picture of how human-AI collaboration actually performs in production, rather than in controlled pilots. The finding that data quality, not AI capability, is the primary failure point is significant: it shifts the design challenge away from model selection and toward information architecture and knowledge management. In behavioral terms, the human analysts in this model are not simply approvers; they are active supervisors who shape and calibrate agentic behaviour over time, which means their judgment and feedback loops are themselves a core service-design input.

The declining rejection rate is equally telling. It suggests that trust between human operators and AI agents is being built incrementally through repeated, low-stakes interactions — a pattern consistent with behavioral research on automation adoption. Service designers and CX leaders building hybrid human-AI teams should pay close attention to how that trust is scaffolded, because the rate at which humans cede oversight is a leading indicator of both efficiency gains and risk exposure.

By the numbers

• ~1 in 3 enterprise IT workflow actions are currently executed by AI agents, with that proportion rising.
• ~1 in 4 AI-proposed actions are rejected by human analysts, though the rejection rate is trending downward.
• 4 stages of agentic work identified by Fixify: planning, proposing, approving or declining, and acting on approved steps.
• Tens of thousands of human-AI interactions were analysed to produce the findings.

The Renascence take

Most commentary on AI in service operations fixates on the automation rate — how much the machine can do without a human. Fixify's data suggests that framing is the wrong one entirely. The more strategically important metric is the quality of the handoff points: where humans intervene, why they decline, and what those decisions teach the system over time.

The real service-design work here is not configuring the AI — it is designing the human review experience so that analysts give rich, consistent signals rather than binary thumbs-up or thumbs-down responses. Behaviorally, a declining rejection rate can mean the system is improving, or it can mean analysts are fatiguing into rubber-stamping; operators need instrumentation to tell the difference. Customer-obsessed organisations should treat the human-in-the-loop not as a compliance checkbox but as their most valuable source of ground-truth data — and invest in the interface, the incentives and the cognitive load management that makes that role sustainable.