The 90 percent that want autonomy and the 77 percent that will not grant it

Ninety percent of companies want to move toward autonomous IT operations over the next two years, with agentic AI, software that plans and carries out multi-step tasks on its own, doing the work. Yet 77% say their own organisation hesitates to let AI make an operational decision without a human approving it first (Help Net Security, 9 October 2026). Most of those companies mean autonomy with a person still at the approval step. Few mean IT that fully manages itself.

The same reporting puts numbers against the hesitation. Only 41% say their IT environment is actually ready for autonomous operations, and just 19% of IT operations are automated today (Help Net Security, 9 October 2026). Visibility is the sharper gap: 96% say unified visibility across networks, applications, endpoints and cloud matters for AI-driven IT, but only 17% have it unified across all of those areas. Just over half, 51%, say they lack enough oversight of AI agent performance and reliability. As the report puts it, businesses are planning around a technology their own environments cannot support.

Why this is not an isolated finding

The pattern holds up outside this one survey. Technical specialists inside these same organisations favour human supervision more than business and IT leaders do, and are less likely to expect autonomous IT to become the operating model (Help Net Security, 9 October 2026). The people closest to the systems trust them less than the people deciding whether to buy more of them.

That split echoes a broader problem with agent governance that shows up wherever anyone actually measures it rather than surveys intent. Reco's State of Agent Security 2026 (26 August 2026), built from anonymised enterprise telemetry and analysis of 500 public Model Context Protocol servers, found 80% of AI tools running with no oversight at all. Separately, security researchers describing how agent identities get created inside AI studios have observed the same accumulation pattern: an agent's permissions widen one unblocked run at a time until the account holds authority well past its original job, as Orchid Security co-founder and CEO Roy Katmor has put it. The agent is approved, the studio is approved, but the identities behind them sit outside the usual review process.

These three measurements use different methods but converge on the same shape: intent to automate is running well ahead of the visibility and control needed to do it safely.

What it means for a regulated enterprise

Security and compliance concerns are the top obstacle to autonomous IT operations, and leaders weigh them more heavily than any other barrier (Help Net Security, 9 October 2026). That is not caution for its own sake. Under the EU AI Act, record-keeping obligations assume an organisation can state what a system did and on what basis it acted. An agent operating with 51%-grade oversight of its own performance, inside an environment where two-thirds of the relevant systems are not visible in one place, cannot support that assumption. Under NIS2, cybersecurity risk management is a named management responsibility; handing an unsupervised agent write access to production infrastructure because the business case for speed won by a point is exactly the kind of decision that obligation is designed to make personal.

The practical exposure is narrower than "AI will make a bad call." It is that nobody can currently produce the record of which calls an agent made, under what approval, with what authority, across which systems, after the fact. Technical specialists already sense this; it is why they ask for more supervision than the leaders signing the budget.

What actually addresses it

The fix is closing the three measured gaps in order: visibility, oversight and approval, each enforced in a way that can be inspected rather than asserted.

Visibility means one record of what an agent touched, not a dashboard assembled from four tools after the fact. Oversight means comparing what an agent was approved to do against what it actually did, continuously, not at quarterly review. Approval means the consequential actions, the ones that move money, change access or alter production state, stop for a human decision by default, with that decision itself recorded next to the action it authorised.

Building the record before the autonomy is what this requires, not a choice between autonomy and control, so that when an agent does act alone the organisation can show, immediately, why that was safe to allow.

What to check on Monday

Pick one agentic workflow already running, even a modest one, and try to answer four questions about it without opening a ticket: who approved its current scope, what systems can it reach right now versus what it was originally scoped for, what did it actually do in the last seven days, and who would see it if that changed overnight. If any of those four takes more than a few minutes to answer, or needs more than one tool open to answer, that is the gap the survey above is describing, inside your own organisation rather than a sample of respondents.