A salesperson pastes a client's pricing quote into ChatGPT to tighten the wording. A contract manager drops a vendor agreement into Claude to summarize the termination clause before a call. An account executive uploads a client's financial statement to ask an AI tool to draft a proposal. None of this happens with malicious intent. Each person is trying to work faster. Each one is also moving confidential company data outside the company's control, with no record of what left, where it went, or who can see it now.
This is shadow AI: the use of consumer AI tools for work tasks without sanction, oversight, or an audit trail. It has already become the default way many employees use AI at work, and the data on how much confidential material is moving through it is now well documented.
What does shadow AI actually look like inside a business?
It rarely looks like a policy violation. It looks like ordinary work, done with a free tool that happens to be very good at the task. A staff member has a contract to review, a client email to answer, or a quote to draft, and a consumer AI tool gets it done in minutes instead of an hour. Nobody sits down and decides to expose company data. The exposure is a side effect of speed.
The tools involved are almost always public, consumer-facing products: ChatGPT, Claude, Gemini, and similar services accessed through a personal or free-tier account. These accounts sit outside any company-managed environment. There is no login tied to a company identity provider, no data-retention agreement negotiated by the business, and no log a company can pull if it later needs to know what was shared.
How much confidential data is actually leaving through these tools?
Cyberhaven Labs analyzed ChatGPT usage across 1.6 million workers and found that 11% of what employees paste into ChatGPT is confidential company data, and that 4.7% of employees have pasted confidential data into the tool at least once (Cyberhaven, "11% of Data Employees Paste Into ChatGPT Is Confidential," updated May 2025).
A 2026 survey of 1,250 office professionals at companies with at least $500 million in annual revenue found the exposure is broader than a single tool. Eighty-eight percent of respondents had shared work-related information with a public AI system such as ChatGPT, Claude, or Gemini. Of that group, 43% had shared emails and correspondence, 40% had shared meeting notes, 34% had entered customer data, and 31% had shared financial information or confidential company documents and strategies (PagerDuty, "Shadow AI Workplace Survey," June 2026).
The pattern is routine, not a handful of careless employees: ordinary work tasks, run through tools that were never built to hold a company's confidential information.
Why do employees turn to consumer AI instead of a sanctioned tool?
The most direct answer is that most companies have not given them another option. IBM's 2025 Cost of a Data Breach Report found that only 37% of organizations have a policy to manage or detect shadow AI use (IBM, "Cost of a Data Breach Report 2025"). Without a sanctioned tool and a clear policy, an employee facing a deadline reaches for whatever is fastest and free.
Weak enforcement compounds the gap. In the PagerDuty survey, 66% of office professionals said they had used an AI tool at work despite believing it was not permitted under company policy. Among those who were confronted about it, 53% received only informal feedback and continued using the tool anyway, and 48% faced formal disciplinary action and persisted regardless. A rule with no sanctioned alternative behind it does not change behavior; it just pushes the behavior further out of sight.
This sits inside a larger pattern. Deloitte surveyed 3,235 IT and business leaders across 24 countries and found that only 21% of organizations have a mature governance model for agentic AI, meaning roughly 79% do not (Deloitte, "Agentic AI Is Scaling Faster Than Guardrails," January 2026). Shadow AI is what that governance gap looks like on an ordinary Tuesday, at the level of a single employee with a contract open in one tab and ChatGPT open in another.
What does a leak like this actually cost?
IBM's 2025 Cost of a Data Breach Report found that organizations with high levels of shadow AI use incurred an average of $670,000 in additional breach costs compared to organizations with low or no shadow AI use. The same report found that one in five organizations had already experienced a breach involving shadow AI (IBM, "Cost of a Data Breach Report 2025"). Globally, the average cost of a data breach in 2025 was $4.44 million (IBM, "Cost of a Data Breach Report 2025").
A pasted contract or a pasted pricing quote does not need to trigger a headline-grabbing breach to cause damage. A client's confidential terms sitting inside a third-party AI provider's systems, with no data-residency guarantee and no record of who accessed it, is itself a breach of the confidentiality a client was promised. The cost is not only financial; it is the loss of the authority boundary a business is supposed to hold around its own client relationships.
What does controlled deployment actually look like?
Employees already want to use AI, and in large numbers, because it makes them faster at real work, as the data above shows. Banning the behavior and hoping enforcement holds has not worked. Giving staff a sanctioned path that does the same job, without the data leaving the company's control, does.
That means a private or managed deployment, or a self-hosted model, built with three things a consumer chatbot does not have: an authority boundary that defines what the tool can see and act on, data-residency control so confidential material stays inside an environment the company controls, and a full audit trail of every document and prompt that passed through it. Staff still get the speed of AI assistance for drafting, summarizing, and reviewing. The company still knows exactly what was shared, by whom, and where it went.
What is the practical first step?
A policy memo is not the first step. A policy without a sanctioned alternative is what produced the current numbers. The first step is standing up one controlled tool, scoped to the tasks employees are already doing in ChatGPT and Claude today: contract review, quote drafting, correspondence summarization. Once staff have a sanctioned option that is as fast as the consumer tools they are already using, the incentive to paste confidential material into a public chatbot goes away on its own.
Where Mirai360 AI fits
Mirai360 AI builds controlled AI deployments across three paths: self-hosted, managed, and custom-built, each configured with authority boundaries, data-residency control, and full audit trails from day one. The goal is to make sure the speed employees already want from AI does not come at the cost of a client's confidential contract, quote, or data sitting inside a system nobody at the company can see into, not to slow a team down with restrictions.
If your team is already pasting confidential material into consumer AI tools, and most teams are, the question is not whether to give them a faster tool. It is whether that tool is one your business controls.
Book a free 30-minute call to talk through what a controlled AI deployment would look like for your business.
Sources
- Cyberhaven, "11% of Data Employees Paste Into ChatGPT Is Confidential," updated May 2025: cyberhaven.com
- PagerDuty, "Shadow AI Workplace Survey," June 2026: pagerduty.com
- IBM, "Cost of a Data Breach Report 2025": ibm.com
- Deloitte, "Agentic AI Is Scaling Faster Than Guardrails," January 2026
FAQ
- How much confidential data is actually leaving through these tools?
- Cyberhaven found 11% of what employees paste into ChatGPT is confidential company data, and 4.7% of employees have pasted confidential data at least once (updated May 2025).
- Why do employees turn to consumer AI instead of a sanctioned tool?
- Only 37% of organizations have a policy to manage or detect shadow AI use (IBM, 2025), so employees default to whatever free tool is fastest.
- What does a leak like this actually cost?
- IBM found organizations with high shadow AI use incurred $670,000 in additional breach costs on average, and 1 in 5 organizations had a breach involving shadow AI (2025).