How companies are tackling data leakage from Bring Your Own AI (BYOAI)

TL;DR

• 78% of employees who use AI at work brought their own tools (BYOAI), and most of that usage is invisible to IT. 

• The primary leakage reason is copy-paste into browser-based AI tools that bypass traditional DLP entirely. 

• Shadow AI breaches already cost organizations an average of $670,000 more per incident (IBM¹). 

• Blocking AI tools doesn’t solve the problem. It pushes employees to personal devices and eliminates whatever visibility you had. 

• The most effective governance starts with software discovery (identifying which AI tools are actually in the environment) before layering on policy, access controls, and enforcement. 

• With the EU AI Act’s high-risk provisions enforceable by August 2026, organizations that can’t inventory their AI tools face regulatory exposure on top of data risk.

Employees aren’t waiting for IT to hand them AI tools. They’re signing up on their own, pasting company data into free-tier chatbots, and getting their work done faster than anyone expected.  

The problem: most organizations can’t see it happening.  

This article breaks down how BYOAI creates data leakage risk and why the most effective governance response starts with software discovery rather than restriction. 

What BYOAI looks like inside an enterprise

“Bring your own AI” sounds like a policy discussion. In practice, it looks like a marketing manager pasting a draft press release into ChatGPT through a personal browser tab. A developer debugging proprietary code in an AI assistant they signed up for over the weekend. A finance analyst uploading a revenue forecast to an AI summarizer that IT has never heard of. 

According to Microsoft’s Work Trend Index, 78% of employees who use AI at work brought their own tools². Personal ChatGPT accounts, browser extensions, AI-powered coding tools, image generators, meeting summarizers… These tools span every function and every seniority level. 

Most of this activity happens in the open. People aren’t sneaking around, they’re just getting their work done. An UpGuard report found that less than 11% of AI applications in the workplace are actually visible to IT teams³. That’s an entire estate of applications operating outside governance, and IT is looking at a sliver of it. 

52% of employees say they wouldn’t tell their manager they used AI for a task, according to the same Microsoft report. The incentive structure is stacked: AI makes them faster, disclosure feels risky, and nobody is checking. 

If you managed IT during the BYOD (Bring Your Own Device) wave a decade ago, the pattern feels familiar. But there’s a critical difference. When employees brought personal phones to work, the data risk was contained to a device you could wipe. With BYOAI, data flows through cloud services that retain inputs, train on them, and sometimes expose them to other users. 

Where the data leaks

Data leakage from shadow AI follows three specific patterns that most organizations aren’t equipped to detect. 

Copy-paste into browser-based AI tools

This is the most common and the hardest to catch. Employees copy proprietary code, customer records, financial data, and internal memos directly into AI chat interfaces. Cisco’s AI Readiness Index found that 48% of employees have entered non-public company information into AI tools⁴.  

Cyberhaven’s 2026 AI Adoption and Risk Report puts it even more starkly: 39.7% of all AI interactions now involve sensitive data⁵. 

Traditional DLP wasn’t built for this. It monitors file transfers and network traffic, but text pasted into a browser interface never triggers those sensors. The data leaves as a prompt, not a file, and most security stacks classify that as routine web traffic. 

Personal accounts with no enterprise controls

One-third of employees access AI tools via personal accounts, according to Cyberhaven. Personal accounts have none of the enterprise controls that matter. No SSO, no audit logging, no data retention policy, and no way for IT to revoke access if something goes wrong. Data shared through a personal ChatGPT free-tier account has fundamentally weaker protections than an enterprise deployment and most employees don’t think about the difference. 

Samsung learned this the hard way. In 2023, engineers at Samsung’s semiconductor division used personal ChatGPT accounts for code generation and debugging. They inadvertently uploaded proprietary source code and confidential meeting notes into a third-party service with no data processing agreement. Samsung banned all external AI tools⁶ shortly after. 

OAuth permissions and third-party AI integrations

This is the newest and the least understood. Employees grant AI tools access to corporate systems through OAuth permissions, often with a single click during sign-up. That one click can give a third-party AI app read access to email, documents, code repositories, or cloud storage. 

The Vercel breach in April 2026 showed exactly how dangerous this gets. A Vercel employee granted broad Google Workspace OAuth permissions to Context.ai, a third-party AI tool. Context.ai was compromised via Lumma Stealer malware⁷, and the breach went undetected for two months. Attackers gained access to employee records, API keys, GitHub tokens, and source code. Nobody at Vercel caught it. The breach surfaced when the stolen data appeared for sale on a dark web forum, listed at $2 million. 

What BYOAI and shadow AI breaches cost

IBM’s 2025 Cost of a Data Breach Report provides the clearest picture yet of the financial damage: 

• 20% of organizations have already experienced a data breach linked to shadow AI. 

• Those breaches cost an average of $670,000 more than breaches without a shadow AI component. 

• 97% of organizations that suffered AI-related breaches lacked proper AI access controls. 

• 65% of shadow AI incidents resulted in personally identifiable information (PII) exposure. 

• The average U.S. data breach now costs $10.22 million. Shadow AI doesn’t just add risk. It adds cost that compounds on an already expensive problem. 

• 63% of organizations still have no AI governance policy in place or are still developing one. 

Adoption is running years ahead of governance, and the financial consequences of that gap are measurable. 

Why blocking AI doesn’t work

The instinct is understandable. If AI tools create data risk, restrict them. Samsung did it. Many organizations have tried. 

But blanket bans don’t stop AI usage. They stop visibility into AI usage. 

89% of employees know the rules around AI at work but bypass them anyway, according to the UpGuard report. When corporate networks block ChatGPT, employees switch to personal phones. When IT restricts browser extensions, employees use personal laptops. The tools still get used. The data still flows out. You just can’t see it anymore. 

Anyone who’s tried to enforce a strict “no personal devices” policy knows the pattern. Restriction without a credible alternative drives the behavior underground. And underground is exactly where you lose the ability to manage it. 

Policy without enforcement is wishful thinking. But enforcement without visibility is guesswork. You can’t enforce DLP on applications you haven’t discovered. You can’t write acceptable-use policies for tools you don’t know exist in your environment. 

This is why the most effective AI governance strategies start with discovery, not restriction. 

The BYOAI governance framework that works

Managing BYOAI requires a layered approach. Each step builds on the one before it. Skip the foundation and the rest falls apart.

Step 1: Discover what’s already in your environment

You can’t govern what you can’t see. The first step is identifying which AI tools employees are actually using across your software estate. 

This is a software asset management problem. SaaS discovery surfaces unauthorized AI applications (browser-based chatbots, AI-powered productivity tools, third-party integrations) alongside the rest of your software inventory. Xensam’s SAM platform provides this kind of automated discovery across SaaS, on-premises, and AI applications, giving IT teams a real-time view of what’s running in the environment. 

The numbers underscore why this step comes first. Only 12% of companies can detect all shadow AI usage⁸. Shadow AI incidents are four times more likely to go undetected than traditional shadow IT. Without automated discovery, the vast majority of AI sprawl stays invisible, and so does the risk.

Step 2: Classify and risk-score every AI tool

Not all AI tools carry the same risk. A design assistant that processes only text prompts is a different threat profile than a coding tool with OAuth access to your GitHub repositories. 

Once discovery reveals the full picture, every AI tool should be categorized: 

• Approved – meets enterprise security requirements, SSO-integrated, data retention controls, SOC 2 certified 

• Restricted – acceptable for non-sensitive use cases with guardrails 

• Prohibited – unacceptable data handling practices, no enterprise tier, data residency concerns 

The classification should assess data handling policies, retention and training practices, compliance certifications, and where data is stored and processed.

Step 3: Provide secure alternatives

Employees use personal AI tools because they need to get work done. If you take those tools away without providing something better, shadow AI wins every time. 

Enterprise-tier AI deployments (managed accounts with SSO integration, audit logging, data retention controls, and contractual guarantees about how data is handled) provide the safeguards that free-tier personal accounts lack. Moving employees to IT-provisioned, governed AI accounts is the single highest-impact step most organizations can take after discovery. 

For adoption to succeed, sanctioned alternatives must be at least as capable and accessible as the tools employees already chose on their own. If the sanctioned option is slower, harder to reach, or less useful, adoption will stay in the shadows. 

Step 4: Enforce at the browser and endpoint layer

With visibility established and approved tools deployed, technical enforcement fills the remaining gaps. This means AI-aware data loss prevention that inspects prompt content at the browser layer, not just file transfers at the network layer. 

Practical controls include: 

• Conditional access policies that restrict AI tool usage to compliant, managed devices 

• Blocking personal account sign-in on corporate networks for known AI services 

• Browser-layer DLP that classifies and controls data entered into AI prompts 

• OAuth permission monitoring that flags new third-party AI integrations requesting access to corporate systems

Step 5: Monitor continuously

BYOAI isn’t a one-time audit. New AI tools appear weekly. Employees adopt them in days. The discovery process from Step 1 needs to run continuously, not annually. 

Ongoing SaaS discovery and usage monitoring detects new AI tools as employees adopt them before those tools become entrenched and before sensitive data flows through them unchecked. Continuous monitoring turns software asset management into a standing governance layer rather than a periodic exercise. 

This is where SAM earns its place in the AI governance stack. The same automated discovery that tracks license usage and identifies waste also surfaces new, unsanctioned AI applications. It keeps IT ahead of adoption rather than reacting to incidents after the damage is done.

The compliance dimension you can’t ignore

BYOAI isn’t only a security problem. It’s a regulatory exposure that’s growing fast, and the deadlines aren’t theoretical.

EU AI Act

The EU AI Act’s high-risk provisions become fully enforceable by August 2026⁹. Organizations deploying or using high-risk AI systems must maintain inventories, conduct risk assessments, and demonstrate ongoing oversight. But a majority of in-scope organizations haven’t completed an AI inventory. 

You can’t complete an AI inventory if you don’t know which AI tools your employees are using. Software visibility isn’t a nice-to-have for EU AI Act compliance. It’s a prerequisite.

GDPR and data sovereignty 

When an employee pastes customer data into a US-hosted AI tool, that data may be processed and stored outside the EU, creating a potential cross-border transfer violation under GDPR. 55% of organizations report that employees have created data sovereignty violations by routing data through offshore AI servers, according to industry research compiled by UpGuard. The European Data Protection Board issued EUR 1.2 billion in GDPR fines tied to AI data processing violations in 2023 alone. 

For organizations subject to EU regulation, knowing where data flows is a compliance requirement. Automated software discovery provides the audit trail that proves which tools are in use and where data is being processed. Xensam’s European infrastructure, built and operated in Sweden, governed by EU law, ensures that the visibility platform itself keeps data under European jurisdiction.

How Xensam helps you tackle BYOAI data leakage

Most organizations try to solve BYOAI with security tools alone. They invest in DLP, browser controls, and endpoint monitoring, all of which matter, but skip the step that makes those investments effective: knowing what’s actually running in the environment. 

Xensam’s SAM platform closes that gap. It provides automated discovery across SaaS, on-premises, and AI applications in hybrid IT environments, surfacing the tools your employees are already using before those tools become data leakage incidents.

Discover shadow AI across your entire software estate 

When employees sign up for AI tools through a browser, those applications create SaaS usage signals. Xensam detects them automatically, even when IT hasn’t heard of the tool and nobody submitted an approval request. You get a real-time view of every AI application in your environment, not just the ones you provisioned. 

This matters because you can’t write acceptable-use policies for tools you don’t know exist. And you can’t classify or risk-score applications you haven’t discovered. Xensam surfaces every AI application in your environment, including tools IT never provisioned and never approved.

Measure actual usage, not just presence 

Discovery tells you what’s installed or accessed. Xensam goes further. It measures true application usage across your estate. That means you can see which AI tools are being used daily by dozens of people versus which ones were tried once and abandoned. Usage data turns a binary “approved or not” decision into a risk-informed prioritization.

Support compliance with a verifiable AI inventory 

The EU AI Act requires organizations to maintain inventories of AI systems in use. GDPR demands that you know where personal data is being processed. Xensam’s continuous discovery provides the audit trail that satisfies both requirements, showing which tools are active, who’s using them, and where data flows. 

Because Xensam is built and operated in Sweden under EU jurisdiction, the platform itself keeps your governance data within European data sovereignty boundaries.

Feed governance into your broader security stack 

Xensam doesn’t replace your DLP or access controls, it feeds them. Through DataBridge integrations, discovery data flows into your existing security and IT service management systems. When Xensam identifies a new, unsanctioned AI tool, that signal can trigger access reviews, policy enforcement, or incident workflows in the tools your security team already uses.

Move from reactive to continuous governance 

BYOAI isn’t a problem you solve once. New AI tools appear weekly, and employees adopt them in days. Xensam’s automated, continuous discovery keeps pace with that adoption cycle, detecting new AI applications as they appear and keeping your inventory current without manual audits or periodic sweeps. 

The result: IT stays ahead of shadow AI adoption rather than reacting to breaches after the fact. 

Your software estate includes AI tools you haven’t approved and data flows you can’t see. Xensam’s SAM platform gives you the visibility to discover, assess, and govern every AI application across your environment. 

FAQ