Corporate policies struggle to stop employees from using unapproved AI tools

The rise of shadow AI

Many organizations are currently struggling to manage what experts call 'shadow AI.' This happens when employees use artificial intelligence tools that have not been vetted or approved by their company's IT department. According to a report by Cybersecurity Insiders, 65 percent of security and compliance professionals have caught staff using these unapproved tools with company data.

The primary driver for this behavior is often productivity. Employees are frequently trying to meet deadlines and find that AI tools are the most efficient way to get work done faster. When a company bans a tool, the work does not stop; it simply moves to personal accounts on corporate laptops, which hides the activity from security teams.

This shift creates a significant visibility gap for security leaders. While a company might have a policy in place, those policies often fail to account for the reality of daily workflows. The Verizon 2026 Data Breach Investigations Report notes that 45 percent of employees are now regular AI users on corporate devices, a sharp increase from 15 percent just one year ago.

Why policies are not enough

A major issue with the current approach is that a policy is merely a statement of intent. It tells employees what they should do and tells auditors what the company intended, but it does not create a record of actual behavior. If a regulator asks which AI tool touched a specific customer file on a specific date, saying "we told people not to" does not provide evidence of control.

The data suggests that the quality of a policy matters, but even the best ones have limits. In a survey of 209 security professionals, 72 percent of organizations had an AI use policy, but only 44 percent called it comprehensive. Organizations with comprehensive policies reported a 17 percent rate of widespread shadow AI, while those with none reported a 37 percent rate.

Furthermore, 67 percent of regular AI users reach these tools through non-corporate accounts. Because these accounts sit outside the organization's logging systems and policy engines, the security team loses oversight at the exact moment data is most sensitive. This creates a situation where the company is essentially flying blind regarding where its information is being processed.

The challenge of AI agents

As companies begin to move toward using AI agents—software that can perform tasks on behalf of humans—the security risks become more complex. Security professionals are worried about excessive permissions, unclear ownership, and the inability to verify that an agent stayed within its assigned role. These are not problems with the AI models themselves, but rather problems of identity and authorization.

To manage these risks, experts suggest treating agents as distinct identities rather than just features. Each agent should have a named human owner, permissions that are strictly scoped to a specific task, and a clear path for revocation. This means holding an AI agent to the same review cadence as a privileged human user who has access to sensitive systems.

The goal is to ensure that every request for sensitive data, whether made by a human or an agent, passes through a verification layer. This layer should check the user's identity, evaluate the specific request against company policy, and write a tamper-evident record. By building an approved path that is easy to use, companies can encourage employees to stay within the governed system.

Moving toward evidence

Security leaders are encouraged to move away from perception-based reporting and toward telemetry-based evidence. Instead of citing general surveys, they should count AI use by account type, data class, and department. This provides a clearer picture of where the actual risks lie and allows for more accurate reporting to company boards.

One practical way to test current readiness is to perform a manual drill. A security team could choose a date from the previous month and attempt to assemble a complete record of every AI interaction with sensitive data. If this process takes weeks to complete, it reveals a significant exposure that could be costly to address during a real audit or legal deposition.

Finally, organizations need to close the side doors that allow for data leakage. This involves using AI-aware data loss prevention tools, browser controls, and endpoint security. The most effective programs pair a governed, easy-to-use path for employees with strict controls on the ungoverned tools they might still try to use. This approach addresses the reality of human behavior while maintaining a record of control.