Enterprise adoption of agentic AI has moved rapidly into production while the governance and security controls needed to manage those systems have struggled to keep pace, according to research from Enterprise Management Associates prepared for Cequence Security.

The August 2026 report found that 65% of enterprises surveyed have experienced an AI agent acting outside its intended scope. Nearly 29% reported measurable organizational impact, including data exposure, operational disruption, financial consequences or reputational damage, while 36% experienced a near-miss that was caught before causing damage.

The findings suggest the challenge is no longer whether enterprises will deploy agentic AI but whether they can effectively govern agents once they are operating in production.

"This research demonstrates that enterprises have moved well past experimentation with agentic AI right into production, and governance has not kept pace with that shift," said Christopher M. Steffen, vice president of research at EMA and author of the report.

Nearly 46% of respondents said they are already scaling agentic AI across multiple departments and production workflows while more than 31% are moving prototypes into production. Nearly 79% said they are running generative and agentic AI simultaneously.

Agents are increasingly being deployed in operational functions, with IT help desk and service desk automation leading at 56.4%, followed by software development and engineering at 37.6%, security operations at 29.2% and customer support at 25.2%.

Governance Gaps Create Security Exposure

The research found that enterprises often lack the ability to identify and stop unauthorized agent activity quickly.

Only 32.2% of respondents said they can detect and contain an out-of-scope agent action within minutes using automated mechanisms. Nearly 55% require hours and manual steps, while 8% would identify the issue during a scheduled review. Four percent would only know after measurable impact had occurred.

The ability to reconstruct what happened after an incident is also limited. Forty-six percent of respondents said they cannot easily and completely produce an audit trail of all actions taken by a specific agent during the previous 30 days.

Steffen said the problem is not necessarily a lack of awareness but a disconnect between policies and enforcement.

"The gap is between what’s written down and what’s enforced when an agent takes an action nobody approved," Steffen said.

That disconnect is reflected in the report's findings on agent access. Although 94% of respondents expressed at least moderate confidence that their agents are not overprovisioned, only 32.7% said they provision agents using least-privilege access.

Randolph Barr, chief information security officer at Cequence Security, said the disparity is significant.

"What jumps out most is that 94% confidence sitting right next to only 33% of agents provisioned with least privilege," Barr said.

Only 34.2% of respondents evaluate an agent's authorization at the moment it attempts to take a specific action. The majority instead rely on standing permissions or periodic policy reviews.

Barr said organizations should begin by determining exactly which agents are operating in their environments and then evaluating their actual activity and permissions.

"As a CISO, not knowing what you don't know is what keeps me up at night," Barr said.

Inventory and Decommissioning Add to the Risk

Visibility into deployed agents remains another challenge. While 53% of respondents have an automated, centralized agent inventory process, 18% rely on manual inventory and 28.2% have only partial visibility. The report found that 47% therefore cannot reliably inventory all deployed agents.

The research also identified potential exposure from abandoned AI pilots. Thirty percent of agentic AI pilots have been paused indefinitely, discontinued or abandoned, and the report said many were not fully cleaned up after being granted production access and credentials.

EMA identified three priorities for addressing the governance gap: governing agent actions in real time, establishing detection and containment capabilities before expanding deployments and treating agent decommissioning as a security discipline.

The report recommends defining what agents are permitted to do on a task-by-task basis and enforcing those boundaries at the time of action rather than relying primarily on permissions established during provisioning.

The full EMA research report is available here.