Agentic AI governance: more autonomy, more control

The adoption of artificial intelligence agents is advancing faster than many companies’ ability to establish how they should operate. In Brazil, this gap already shows up quite clearly. According to Deloitte’s State of AI in the Enterprise 2026, 95% of Brazilian respondents say their organizations plan to adopt Agentic AI in the next two years, while only 27% say they have mature governance models for this technology [1]. In the survey’s global results, the percentages are 97% and 21%, respectively [1].

This gap matters because agents introduce a characteristic that changes the nature of the problem. In many traditional artificial intelligence applications, a person still receives an answer, evaluates the result, and decides what to do. An agent can go further: it receives a goal, accesses data, uses tools, interacts with systems and, depending on the permissions it receives, executes actions.

The greater this ability to act, the greater the consequences of an inappropriate decision can be. At the same time, excessively restricting the activity of all agents may eliminate precisely part of the speed and automation gains that justify adopting them.

It is no longer enough to define whether an agent can or cannot be used. You need to establish how much autonomy it can receive, under what circumstances, and which controls must accompany that autonomy.

Not all agents represent the same risk

An agent that queries a knowledge base and prepares a recommendation should not be treated the same way as another capable of changing a configuration, modifying a record, moving resources, or executing an action that directly affects a customer.

This difference seems simple, but it has important implications. During the first years of corporate adoption of artificial intelligence, most policies were built for a reality in which the person remained clearly at the center of execution. There were already relevant concerns about privacy, security, intellectual property, answer quality, and data use, but the final decision normally remained a human responsibility.

With agents, part of that execution can be delegated to the AI itself. An agent can, for example, query information, call an Application Programming Interface (API), an interface that allows programmatic communication between systems, select a tool, and perform a sequence of tasks without a person needing to determine each step individually.

That is why applying the same control model to everyone can lead to errors in opposite directions. We may grant excessive freedom to agents capable of producing relevant consequences, or impose so many controls on low-risk activities that automation stops making sense.

Gartner draws attention to exactly this problem by advocating governance models proportional to the level of autonomy. The firm predicts that, by 2027, 40% of companies may downgrade or deactivate autonomous agents due to governance failures identified only after incidents in production [2].

PwC reaches a similar conclusion when discussing the entry of agents into the workforce: they need enough access to be useful, but their autonomy must remain compatible with the risks the organization can manage. As the agent’s independence and the potential consequences of its decisions increase, oversight mechanisms also need to evolve [3].

The level of control needs to follow not only the technology used, but mainly what the agent can do and the consequences its actions can produce.

In practice, this means replacing a binary view, in which the agent is simply authorized or prohibited, with a more contextual assessment. What matters is understanding what it can access, which tools it can use, what kind of decision it is authorized to make, and what the impact of an inappropriate action would be.

Autonomy needs to be accompanied by governance during execution

This logic also suggests that autonomy does not need to be granted all at once. An agent can start by analyzing information and recommending actions. At another stage, certain tasks can be executed with human approval. With greater operational maturity and adequate control mechanisms, some of these actions can gain more independence, while higher-impact decisions continue to require a person’s intervention.

Gartner works with different levels of autonomy and recommends that governance requirements follow that evolution [2]. This approach allows the organization to gradually increase agents’ capacity to act as it develops enough experience, trust, and controls to handle the risks involved.

In this context, human-in-the-loop, an expression used to describe situations in which a person takes part in the AI’s decision or execution process, also no longer necessarily means human approval at every step. Depending on criticality, intervention may be required only for certain actions, in exceptional situations, or when a previously defined threshold is exceeded.

Human oversight needs to contribute to safety and accountability without eliminating the operational gains that led the company to use agents. That requires the degree of intervention to be defined according to the criticality of the activity and not simply applied uniformly to every situation.

Autonomy can be expanded progressively as the organization develops trust, observability, and the ability to intervene when necessary.

There is, however, another relevant change. Corporate governance usually starts with policies, defining permitted technologies, data that can be accessed, responsibilities, and situations in which approval is mandatory. In agentic environments, these rules need to be able to reach the moment of execution.

A written policy does not, by itself, prevent an agent authorized to use a given system from performing an inappropriate action. Gartner warns about this difference and predicts that, by 2029, at least 70% of organizations with Agentic AI in production in infrastructure and technology operations may experience some material service, security, or cost incident related, in part, to insufficient runtime controls [4].

That requires turning governance principles into operational mechanisms capable of verifying identities, applying permissions and limits, recording activity, tracking behavior, controlling costs, and requesting approvals when the situation demands. The goal is to make the rules defined by the organization remain valid when the agent is actually working, including at speeds and volumes that make permanent manual monitoring unfeasible.

In agentic environments, governance needs to stop existing only in policy and also start working during execution.

Boston Consulting Group (BCG) adds another dimension to this discussion. In its research, 35% of organizations already had Agentic AI in production, 44% planned new deployments, and more than a third expected to grant agents independent decision rights within three years [5].

As this autonomy grows, controlling only which resources an agent is authorized to access may leave gaps. An agent may have legitimate permissions to use different systems and still choose an inappropriate sequence of actions to achieve a given goal. BCG calls this problem the authorization gap, highlighting that governance needs to consider not only the access granted, but also how that autonomy is exercised [5].

This point becomes particularly important when agents begin to plan tasks, choose tools, and interact with other agents. Authorization stops being only a check made before execution and becomes part of a continuous governance process.

Governance may be what makes it possible to expand autonomy

There is an apparent contradiction in the discussion about agents. The more autonomy we seek, the more control mechanisms seem to be needed. Seen only as restriction, this relationship may give the impression that governance and innovation move in opposite directions.

In practice, an organization that cannot identify its agents, understand their access, track their actions, or intervene in the face of inappropriate behavior will have legitimate reasons to limit what they can do. The absence of these mechanisms ends up setting a ceiling on the very autonomy the company can safely grant.

That is why the goal should not be to maximize the number of controls or apply the same set of rules to every agent. The challenge lies in establishing mechanisms proportional to the capacity to act and the potential consequences of each case, allowing lower-risk activities to flow more freely while more sensitive decisions remain subject to adequate limits and oversight.

This approach also changes how we think about trust. Instead of assuming an agent is permanently trustworthy, the organization can expand its autonomy as it accumulates evidence about its behavior, improves its observability, and develops the ability to detect deviations and intervene when necessary.

Governance thus becomes part of the infrastructure that makes it possible to increase autonomy responsibly, instead of working only as a mechanism to restrict it.

This may be one of the main challenges of the next stage of Agentic AI adoption. Companies that want to capture the productivity and scale gains agents provide will need to develop, at the same time, the ability to grant autonomy and the mechanisms needed to govern it.

From vision to practice: ARGUS

This relationship between autonomy, context, and control is directly connected to the proposal of ARGUS as an Operational and Agentic Intelligence platform.

With AgentOps (Agent Operations) and AI Gateway capabilities, ARGUS seeks to support organizations in applying policies and tracking agent activity as they begin to access models, tools, and corporate resources. In this environment, governance involves controlling permissions and forms of access, preserving traceability, and tracking aspects such as executions, decisions, tool calls, costs, latency, and results.

This structure also favors an incremental approach to autonomy. Depending on the context and the risk, certain actions may require human approval, while others can be executed within previously established limits. As the organization develops trust and matures its controls, these limits can evolve without losing oversight and auditability.

There is also a dimension we consider particularly important in ARGUS: an agent’s activity does not happen in isolation. A decision may be related to the state of an application, a recent change, an infrastructure event, a human action, or activities previously carried out by other agents and automations.

That is why agentic intelligence connects to shared operational context. By correlating information from multiple sources, ARGUS seeks to allow what the agent did to be understood together with what was happening in the environment when the decision was made. This systemic view broadens the ability to investigate incidents, reconstruct decisions, and understand the relationships among people, agents, automations, applications, and infrastructure.

As agents gain autonomy, governing them will require more than tracking their individual behavior. It will be necessary to understand their actions within the environment in which they produce consequences.

Learn about ARGUS and see how it works in practice.

About the author

Paulo Florêncio de Menezes is an entrepreneur, partner, and co-founder of companies such as Target Solutions, BiPTT, and ARGUS.

He holds a degree in Data Processing from the Federal University of Amazonas (UFAM), an MBA in Business Management from Fundação Dom Cabral (FDC), and a Professional Master’s in Economics from Fundação Getulio Vargas (FGV).

In his Master’s research, he developed the IBEOP-IA (Brazilian Index of Potential Occupational Exposure to Artificial Intelligence), investigating the potential exposure of Brazilian occupations to AI capabilities by combining occupational data, artificial intelligence patents, and natural language processing techniques.

References

[1] Deloitte Brasil. State of AI in the Enterprise 2026. Brazilian results of the survey on enterprise adoption of artificial intelligence and Agentic AI.

[2] Gartner. Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure. May 26, 2026.

[3] PwC. AI agents are joining the workforce. Is your governance ready?

[4] Gartner. Govern Agentic AI Infrastructure at Machine Speed. 2026.

[5] Boston Consulting Group. The Authorization Gap: Why AI Agent Governance Needs a New Approach. 2026.