Shadow AI: which AIs and agents operate in your company?

Perhaps your company knows exactly which corporate artificial intelligence platforms were contracted, which licenses were purchased, which projects were approved, and which tools are being officially made available to employees.

But that does not necessarily mean it knows which artificial intelligences are actually being used inside the organization.

An employee may use a personal generative AI account to analyze a document. Another may install a browser extension. A team may contract a tool on its own. A developer may embed a model or an agent in a workflow. And a business area may start automating activities before IT, security, or governance even know that solution exists.

This phenomenon already has a name: Shadow AI, or AI used, contracted, or implemented outside the organization’s formal mechanisms of awareness, approval, and governance.

The problem is not entirely new, but AI is changing its scale and, above all, its consequences.

Shadow IT did not start with artificial intelligence

For decades, companies have lived with Shadow IT, the term used to describe technologies adopted by employees or business areas without the knowledge or formal approval of the area responsible for technology.

First came locally installed applications. Then cloud services and Software as a Service (SaaS) platforms, a model in which applications are offered as a service over the internet. With a credit card and a few minutes, an area could contract a solution without depending on a corporate technology project.

Artificial intelligence takes this behavior to another level.

The ease of access is enormous. Many tools have free versions, others cost a few dozen dollars per month, and some are already built into applications that companies use daily.

Moreover, unlike many traditional software products, the perception of value can appear immediately. A professional tries a tool to summarize a document, prepare a presentation, analyze a spreadsheet, or write code. If it works, it quickly becomes part of their routine.

Adoption can therefore begin before the company has time to decide how it intends to govern it.

The problem is already inside companies

A recent Deloitte survey helps size this gap between formal and actual adoption.

The GenAI Workforce Survey 2026, conducted with 25,000 workers in the United Kingdom, found that 63% of the workers surveyed said they had already used generative artificial intelligence at work. Among those using the technology, 31% said they did so without their employer’s knowledge. Deloitte itself characterizes this behavior as Shadow AI [1].

Another data point helps understand the dynamic: 17% of users paid out of their own pocket for at least one AI tool used professionally. Deloitte estimates that British workers together spent approximately £958 million a year on these tools [1].

In other words, part of the AI infrastructure used by companies may not appear in the technology budget simply because it is being funded by the employees themselves.

This also shows why looking only at contracts, corporate licenses, or formally approved projects offers an incomplete view of adoption.

Adoption often starts in the activity itself

There is one dimension of this phenomenon that I find especially interesting.

In my master’s research, when developing the Brazilian Index of Potential Occupational Exposure to Artificial Intelligence (IBEOP-IA) to analyze the exposure of Brazilian occupations to artificial intelligence, one of the perspectives adopted was to look at the activities that make up each occupation and their relationship with the technological capabilities of AI [6].

This same perspective helps understand Shadow AI. Often, adoption does not start from a centralized corporate decision. It starts when someone realizes, in their everyday work, that a given activity can be performed or expanded with the use of the technology.

One professional finds a better way to research, another discovers how to automate an analysis, and another starts using AI to produce documents or code. The technology enters through the activity before entering through the corporate architecture.

This is a later reflection based on the perspective used in the research, and not a result of the dissertation on Shadow AI. Even so, it helps explain why this phenomenon is so hard to control through centralized decisions alone.

Why do employees turn to their own tools?

The simplest answer would be to attribute the problem to lack of discipline or to non-compliance with corporate policies. The data show a more complex reality.

Among the participants in the Deloitte survey who paid for their own generative artificial intelligence tools or used potentially unapproved solutions, 21% said those tools were better than the ones provided by the company, while 14% said they were essential for work but were not funded by the employer [1].

There is also a guidance gap. Only 28% of the generative AI users surveyed by Deloitte said their organization had a policy on the topic that they knew about, understood, and considered clear and up to date. And 50% had not received formal training [2].

These numbers help show that Shadow AI cannot be treated only as a disciplinary problem. There is a real tension between productivity and control.

People are finding utility in the technology faster than many organizations can incorporate it into their evaluation, procurement, security, and governance processes.

Blocking everything probably does not solve it

The initial reaction of some organizations may be simply to restrict access to unapproved tools. There are situations in which this will be necessary, but making blocking the main strategy can produce exactly the opposite effect.

When discussing the proliferation of AI agents, Gartner warns that, when employees cannot work with the authorized tools, they may look for alternatives outside corporate controls, expanding Shadow AI [3].

The challenge, therefore, is not to choose between total freedom and prohibition. It is to create a corporate alternative useful enough that people have less incentive to bypass governance mechanisms, while the organization keeps the ability to define policies and limits.

This balance will become even harder with the next stage of this evolution.

From Shadow AI to Shadow Agents

So far, much of the discussion about Shadow AI involves people using artificial intelligence tools without the company’s knowledge. With agents, the problem changes in nature.

A chatbot normally receives a request and produces a response. An agent can receive a goal, access data, use tools, call Application Programming Interfaces (APIs), interfaces that allow programmatic communication between systems, interact with applications, and execute a sequence of actions to complete a task.

This means AI stops only producing information and also starts taking part in execution.

Gartner already uses the expression shadow AI agents, or ungoverned AI agents, to describe agents that may be embedded in enterprise software, consumed directly over the internet, or created by employees themselves [4].

According to Gartner, 75% of organizations reported unauthorized use of AI-based coding assistants, while 50% reported employees accessing public generative AI agent platforms offered as SaaS [4].

The question, therefore, is no longer only which tool the employee is using. It also includes which agent is acting and what data it can access, which tools it can use, and which actions it can execute.

The first problem is finding out what exists

It is hard to govern what the organization cannot see.

That is why, even before discussing sophisticated governance policies, there is a more elementary problem: visibility.

The organization needs to be able to identify which tools, models, and agents are being used, who created and answers for each agent, which applications and data are within its reach, which tools can be triggered, and which actions can be executed.

This need tends to grow quickly.

Gartner projects that an average company on the Global Fortune 500 list could go from fewer than 15 AI agents in 2025 to more than 150,000 in 2028 [3].

It is a projection, not a current count. Even so, the order of magnitude helps show the potential size of the problem.

In the same survey, only 13% of organizations said they believed they had adequate governance to manage AI agents [3].

If this expansion happens, relying on manual controls or on inventories kept in isolation by the different areas will simply not be a sustainable strategy.

The risk is not only data leakage

Information leakage is probably the most intuitive risk of Shadow AI. An employee may enter into an external tool customer information, internal documents, source code, or commercial data without understanding how that information will be stored or processed.

But the problem is broader.

There are risks related to security, compliance, intellectual property, duplicated costs, lack of auditing, use of inadequate models, and difficulty knowing where corporate information is being processed.

When agents enter the equation, an additional dimension appears: ungoverned actions. An agent with inadequate access may query data it should not, use a credential, trigger another system, or execute an action outside the intended limits.

The 2026 Boston Consulting Group (BCG) survey of cybersecurity leaders shows how controls are still at an early stage: fewer than 20% of respondents had adopted measures such as Shadow AI monitoring, detection of prompt injection attacks, or governance of non-human identities [5].

BCG also found an association between greater adoption of these controls and a lower reported incidence of significant impacts related to AI vulnerabilities, data leakage to AI tools, and Shadow AI [5].

This does not mean a single control solves the problem. It shows, however, that visibility and governance are starting to become part of the infrastructure needed to use AI safely.

From policy to the ability to control

A corporate AI policy remains necessary. It can establish which tools are allowed, which data can be used, which practices are prohibited, and who is responsible for approving new use cases.

But a policy, on its own, does not necessarily show what is happening.

As the number of models, tools, integrations, and agents grows, the organization will need to combine policy with operational capability to discover and inventory what exists, identify owners, authorize access, monitor activities, record actions, maintain audit trails, and, when necessary, apply blocks.

This transition is important because Shadow AI reveals a fundamental difference between defining governance and having the capacity to exercise it. And this problem will become even more important when agents stop being isolated experiments and start taking part in corporate processes at scale.

From vision to practice: ARGUS

This scenario connects directly to the proposal of ARGUS as an Operational and Agentic Intelligence platform, aimed at environments in which people, AI agents, automations, systems, and infrastructure come to share the operation.

With AgentOps (Agent Operations) and AI Gateway capabilities, ARGUS expands visibility and governance over the agentic environment, allowing you to apply policies, control how agents access models, tools, and AI resources, and track their executions, actions, costs, and results.

One of the challenges of this new reality is precisely to prevent the expansion of AI from creating new silos of intelligence inside the organization. As the number of agents grows, it becomes increasingly important to know which agents exist, who is responsible for them, what they can access, which tools they use, and what they are doing.

This is where this agentic dimension connects to operational intelligence. ARGUS also correlates events from multiple sources, reduces noise, and builds shared operational context, making it possible to relate what agents do to what is happening in the infrastructure, applications, and services when their decisions and actions take place.

This way, we do not observe agents in isolation on one side and the operation on the other. People, agents, automations, and systems can share a systemic view, with policies, traceability, and oversight compatible with the level of autonomy granted.

The goal of ARGUS is to create conditions for AI adoption to grow in a controlled, observable, and scalable way, without the company progressively losing the ability to understand and govern the environment.

Get to know 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 UK. GenAI Workforce Survey 2026 — What we do in the AI shadows. Survey of 25,000 UK workers, conducted in May and June 2026.

[2] Deloitte UK. GenAI Workforce Survey 2026 — Transform your workforce and your organisation will follow.

[3] Gartner. Gartner Identifies Six Steps to Manage AI Agent Sprawl. Apr. 28, 2026.

[4] Gartner. The AI Agents You Can’t See Can Still Create Risk. Sep. 30, 2026.

[5] Boston Consulting Group. Cybersecurity Budgets Are Growing Fast. AI Threats Are Growing Faster. Aug. 27, 2026.

[6] Menezes, Paulo Florêncio de. IBEOP-IA — Índice Brasileiro de Exposição Ocupacional Potencial à Inteligência Artificial (Brazilian Index of Potential Occupational Exposure to Artificial Intelligence). Professional Master’s dissertation in Economics, Fundação Getulio Vargas, 2026.