Español
Artificial intelligence has entered everyday work at a pace that is difficult to compare with most other technologies. Today, it is common for someone to use ChatGPT to summarize a document, prepare a presentation, analyze information, or complete a task that once took hours. Often, this happens before anyone in the company has formally decided to adopt that tool.
For a manager, this can look like good news. The team saves time, moves faster, and finds new ways to get work done. And often, it is.
But faster does not always mean better.
Consider Argentina. In the local findings of a global study by KPMG and the University of Melbourne, 80% of respondents said they intentionally use AI at work. Yet only 50% said their organization has policies and practices in place to support its responsible use.
The gap between those two numbers captures something larger about where companies are today. AI is already being used across organizations, even as many of them are still deciding how they want it to be used.
That creates a new set of questions for managers. What information are employees sharing with these tools? Are they using personal or company-managed accounts? What happens to the data they enter? How are those accounts configured? And who checks an AI-generated answer before it becomes part of a decision, document, or business process?
This phenomenon has a name: Shadow AI.
What is shadow AI?
Shadow AI is the use of artificial intelligence tools, models, or applications within an organization without the approval, oversight, or knowledge of the teams responsible for managing them. It can be as simple as using a personal ChatGPT account to work with company information, but it can also include AI-enabled applications, models, or agents introduced by a team outside the organization’s established controls.
The term is related to Shadow IT, which has long described technology used outside a company’s officially managed systems. AI, however, introduces an important difference. These tools do more than store or move information. They can analyze it, transform it, generate new content from it and, increasingly, take action.
Your company may be using more AI than you think
Another set of data from Argentina helps illustrate how this adoption is taking place. The National Survey on AI Adoption in Argentina, conducted by the Center for Evidence-Based Policy at Universidad Torcuato Di Tella and Fundar, found that 82.2% of people who use AI for work rely on free versions of these tools. Only 6.2% said their organization pays for them.
Using a free version does not necessarily mean that a tool is being misused or that it is unsafe. But it raises a question many organizations still cannot answer clearly: Under what accounts, configurations, and conditions are our people using AI?
Using artificial intelligence inside a company can mean very different things.
One employee may use an AI tool through a personal account, while another accesses it through a company-managed environment. Someone else may rely on AI features embedded in software the organization already uses. In each case, the available controls, data retention policies, enabled integrations, and treatment of company information may differ.
The issue, then, is not simply which application an employee is using. It also matters what information they share with it, who controls the account, and what configuration sits behind it.
What data are we sharing with AI?
To use AI at work, we need to give it context. And much of that context comes from the company itself.
It might be a contract we want to summarize, a customer database we need to analyze, a sales proposal we want to improve, code containing proprietary logic, or an internal document we want to turn into a presentation.
The ease with which we can now do these things makes it equally easy to overlook something fundamental: for a tool to process that information, we first have to share it with the tool.
Not every tool or license handles that data in the same way. Nor do they necessarily provide the same controls over retention, access, administration, or use of information. Even within the same product, the conditions can differ between a personal account and an enterprise environment.
So the question should not stop at, “Can we use this tool?” A manager also needs to ask what data the team can use, under what type of account, and with what configuration.
This matters because most people using AI are not thinking about infrastructure, retention policies, or data governance. They are trying to do their jobs better.
That is precisely what makes Shadow AI difficult to manage: an apparently small decision, such as pasting information into a prompt to save half an hour, can introduce company data into a system the organization has never evaluated.
Can we trust what AI gives us back?
So far, we have focused mainly on the information we share with AI tools. But there is another side to the problem: what we do with their answers.
AI can significantly accelerate work. A study published in The Quarterly Journal of Economics analyzed 5,172 customer support agents and found an average 15% increase in productivity among workers who had access to an AI assistant.
But the effect was not the same for everyone. Less experienced workers improved both the speed and quality of their work, while more experienced workers saw modest gains in speed accompanied by a slight decline in quality.
The distinction matters. Introducing AI does not automatically produce a better outcome. The result depends on the task, the person using it and, importantly, what happens with the answer the system produces.
The Argentine findings mentioned earlier reveal a similar tension. The KPMG and University of Melbourne study found that 68% of respondents had used AI-generated output at work without evaluating its accuracy, while 53% said they had made mistakes in their work because of AI.
A well-written answer can look correct without being correct. The same is true of a convincing analysis, a formula, a translation, or a piece of code. The fluency with which an AI system presents an answer is not a measure of its accuracy.
For managers, this introduces a different question. It is not enough to know whether a team uses AI. Leaders also need to understand what they use it for and what level of review each use case requires.
Using AI to organize ideas before a meeting is very different from using it to analyze a contract, evaluate financial information, or generate code that will reach production. The level of oversight should reflect the potential impact of an error.
Faster can be better. But only when we know where we can rely on that speed and where we need to stop and verify.
Shadow AI can also reveal unmet demand
When someone uses a tool their company has not approved, the first reaction may be to ask how to stop it.
But there is another question worth asking first: Why are they using it?
In 2024, Microsoft and LinkedIn found that 78% of people using AI at work were bringing their own AI tools. The phenomenon, known as Bring Your Own AI, spanned generations and was slightly more common in small and medium-sized companies.
Some of this adoption also remains outside the formal conversation inside organizations. Among those surveyed, 52% were reluctant to admit using AI for their most important tasks, while 53% worried that doing so might make them appear replaceable.
This helps explain why Shadow AI should not be viewed solely as a compliance failure.
A 2026 analysis published by Harvard Business Review makes precisely this argument: unauthorized AI use can also signal internal demand that the organization has not yet addressed.
The BBVA case examined by HBR is particularly instructive. The bank chose to observe how its own employees wanted to use AI and channel that demand into a secure environment. The program grew to more than 11,000 active users and led employees to create 4,800 internal tools, with reported time savings of between two and five hours per person each week.
The lesson is not that every tool an employee chooses should be approved. It is that behind an unauthorized use case, there may be a real business problem worth understanding.
Perhaps someone is manually summarizing dozens of documents every week. Perhaps a team spends hours consolidating information for a recurring report. Perhaps there is a repetitive task that has never been important enough to become a formal technology project.
Shadow AI can make those points of friction visible.
For managers, the conversation then changes. Instead of asking only, “What tools is my team using?” a more useful question emerges: “What are they trying to solve with them?”
Banning AI does not make the need disappear
The response to Shadow AI cannot be reduced to blocking tools.
A 2025 Microsoft study found that 71% of UK employees surveyed had used unapproved consumer AI tools at work, and 51% continued to do so weekly. Among the reasons, 41% said they turned to tools they already knew from their personal lives, while 28% said their company did not provide an approved alternative.
A ban may remove access to a particular tool. It does not necessarily remove the problem that tool was solving.
That does not mean companies should allow any use of AI. It means they need to provide a better path.
The first step is visibility: understanding which tools teams are using, for what tasks, and with what information. Then come rules that people can actually apply in their day-to-day work. Telling employees to “use AI responsibly” has limited value if they do not know whether they can analyze customer information, upload a contract, or use a personal account for work.
Providing alternatives matters too. If organizations expect their teams to take advantage of AI, they need to provide appropriate tools, accounts, and configurations. Otherwise, a company can find itself encouraging AI adoption while leaving each employee to decide how to implement it.
Education is equally important. Not everyone needs to understand how a language model works, but employees should understand what information they can share, when an answer needs to be verified, and why a tool accessed through a personal account may operate under different conditions from a company-managed environment.
The goal should not be to stop experimentation that is creating value. It should be to create a path for those uses to come out of the shadows.
That is where the conversation moves from Shadow AI to AI governance.
AI governance allows organizations to establish responsibilities, criteria, and controls by design, rather than trying to add them after a tool or process has already become part of everyday work.
Artificial intelligence allows us to do many things faster. The challenge for companies is making sure that speed does not outpace their ability to understand what they are doing with it.