All Articles

When Employees Treat Their AI Tools as Confidants, How Much Sensitive Data Already Leaves Through a Chatbot?

When Employees Treat Their AI Tools as Confidants, How Much Sensitive Data Already Leaves Through a Chatbot?

When the journalist Helen Li spent a month talking to daters in San Francisco in 2026, she found people already treating AI chatbots as confidants, at least about their feelings. In that personal realm the numbers are already striking. According to an Elon University study described by The Verge, 27 percent of Americans turn to chatbots for personal, emotional, or social questions, and 38 percent disclose things they would not tell their romantic partners. People who confide in a chatbot are not looking for a therapist so much as a listener who never talks back. That habit is now walking through the office door, and it has left executives with a problem that reads more like data governance than romance.

Where the confidant habit shows up at work

The same instinct to tell a chatbot things you would keep quiet makes for a different kind of exposure once employees are on the job, because the data in question is no longer a secret or a grievance. It is customer lists, financial figures, sales pipelines, and proprietary project detail. A global study of 32,352 workers across 47 countries by researchers at Melbourne Business School, published with KPMG in April 2025, found that nearly half of all employees had uploaded sensitive company or customer information into public generative AI tools. Another 44 percent admitted to using those tools in ways that went against company policy, and 63 percent said they had seen colleagues doing the same. The most widely used products were the free public tools, used by about 70 percent of employees, rather than the versions their employers had approved.

What actually leaves the building

So how much sensitive data already leaks through a chatbot, and how quickly. The blunt answer is that a large share is already moving, before any deal is struck or breach is discovered. In the Melbourne and KPMG study, 66 percent of employees relied on AI output without evaluating it, 56 percent said they had made mistakes at work because of it, and a third said AI use had raised privacy and compliance risks. The exposure comes as much from how people talk to the tools as from what they type. A support agent, rushed and looking for a cleverer reply, might paste a client complaint, the account history, and the billing dispute into a consumer chatbot, not imagining that this conversation is now stored on a vendor's servers and possibly fed into a model trained on other companies' data.

That is where the personal analogy runs its course. When a person tells a chatbot what they would not tell a partner, the downside is a broken heart or an awkward meeting. When an employee tells one what the company needs kept quiet, the downside is a reportable breach. The surveys behind these numbers did not document named-company breaches tied to the behavior, so it is worth being careful about what we can claim. What the data do show is the pattern at scale, and the pattern is that sensitive data moves early, through channels nobody is watching.

Why nobody is watching closely

The reason governance trails behavior is that most of this activity happens in plain sight and in private. In the same study, 66 percent of employees used AI tools without knowing if the company even allowed it, 61 percent avoided revealing when they use it, and 55 percent presented AI-generated content as their own work. Only a third of organizations had a policy guiding the use of generative AI at all, and 6 percent had banned it. The result is a shadow layer of use that makes it nearly impossible for an IT team to know what is leaving the building, let alone where it lands. When people treat the tool as a confidant, they also keep the conversation quiet from their managers and security teams.

Who owns the data that leaves

For executives the real problem is not that employees are lazy or reckless, but that the question of ownership was never really answered. An employee using a public chatbot to get help is creating a data relationship with a third party, one that bypasses the vendor contracts, security reviews, and data agreements the company usually negotiates. What lands where depends on the tool's training policies, retention rules, and whether the account is set to share data by default, none of which the user is expected to understand. The Elon University research makes the behavioral case plain: people will confide in the tool and will not treat it as a channel whose inputs and outputs matter. The governance question is simply how you keep up with that once the tool has a job to do.

The answer

Back to the title question, and the answer is this. Quite a lot, and it is already happening faster than any policy cycle can track. It is plausible that the same instinct to confide can carry over into the office, since a tool that feels like a confidant also feels like someone who will keep a secret, and people may then hand sensitive company or customer data to it without pausing. The Melbourne and KPMG study suggests the scale is near half of all AI users uploading sensitive data, and that the rest of the workforce is watching and following along. Nothing about this is mysterious, and it is a data governance gap that opens the moment you let people talk freely to a machine and assume the data will stay put. The fix is not to stop employees from using AI, which is impossible, but to assume whatever they say to a chatbot can leave the building, and to build the controls around that reality.

Sources

Ready to Apply This?

Let's Put This Into Practice

Reading about automation is useful. Having us implement it for your specific business is transformative.