AI agent governance becomes a critical challenge for businesses

Why has governance become a problem for AI agents?

The expansion of artificial intelligence systems capable of performing tasks previously done by humans presents a new dilemma for organizations: how to monitor and control programs that can make decisions and act with varying degrees of autonomy.

Gartner predicts that by 2027, 40% of companies may reduce or disable autonomous AI agents due to governance failures identified after the systems have been deployed.

This figure does not mean that four out of ten agents will be shut down. The estimate refers to companies that may limit autonomy or withdraw these agents from use due to control, security, compliance, or management issues.

AI agents are programs designed to perform task steps with greater independence than a conventional chatbot. Depending on their configuration, they can consult data, suggest actions, modify information, send messages, or execute processes in other systems.

This capability increases the potential for technology application in the workplace, but it also increases risks. For example, an agent with access to internal systems can cause damage on a larger scale than a tool used only for text generation.

According to Gartner, a common mistake is to treat governance as a binary choice: either block the agent completely or trust it without establishing adequate controls for its level of autonomy.

The recommendation is to adopt different levels of governance. An agent that only reads documents does not pose the same danger as one that can alter data or execute actions without human approval.

The more autonomy, the greater the need for control

Gartner classifies agents into several levels. The simplest ones can only observe information or produce recommendations. At higher levels, the system can execute an action after human approval or operate autonomously within predefined limits.

In practice, this requires companies to know exactly what data each agent can access and what operations it can perform.

For example, a program used to summarize documents may have read-only permissions. However, an agent responsible for updating financial records would need more restrictive permissions, approval mechanisms, action logging, and ways to interrupt its operation.

For more autonomous agents, Gartner recommends controls such as continuous monitoring, interruption mechanisms, the possibility of reversing actions, and a clear definition of who is responsible for the system’s behavior.

The risk of creating agents without knowing where they are

The problem is not limited to agents formally developed by technology departments. The spread of artificial intelligence tools can also lead employees to create solutions on their own, a phenomenon known as shadow AI.

In April 2026, Gartner stated that only 13% of organizations considered themselves to have adequate governance for their AI agents. The consultancy also projected that an average Fortune 500 company could have more than 150,000 agents in use by 2028.

In this scenario, maintaining an inventory of agents becomes an essential control step. Companies need to identify which systems exist, who created them, what data they access, what actions they can perform, and when they should be reviewed or deactivated.

AI agents can also be discontinued due to lack of results

Governance is not the only reason that can lead to an agent project being terminated. In 2025, Gartner had already predicted that more than 40% of agentive artificial intelligence projects would be canceled by the end of 2027 due to high costs, lack of clear value, or insufficient risk controls.

The forecast shows that deploying an agent does not guarantee its continuity. The project needs to demonstrate results, have controlled costs, and operate within defined limits.

For professionals working directly with artificial intelligence, this also changes the profile of the skills required. In addition to mastering the tools, those involved in implementing these systems need to understand permissions, data, human supervision, security, and criteria for evaluating agent performance.

The trend indicates that the debate about AI agents should shift from technical capability to issues of responsibility, control, and reliability. The greater the autonomy granted to a system, the greater the need to define in advance what it can do, what data it can access, and when a person should intervene.

For companies, governance becomes part of the artificial intelligence implementation process itself, not just a step after the technology has been developed.