Shadow AI Discovery: How to Find the AI Tools Your Employees Already Use

Somewhere in your organization today, a developer is pasting an error log into a chatbot to fix a bug faster. A sales manager is asking an AI assistant to rewrite a client proposal, and someone in HR is summarizing interview notes with a browser extension nobody in IT has heard of. None of these people think of themselves as a security risk. They are trying to get their work done. That gap between what employees actually use and what IT and security teams know about is why shadow AI discovery has become a practical priority for CISOs, CIOs, and IT managers. Before you can write a sensible AI policy, you need an honest picture of which AI services are in use, by whom, and from which machines.

What Shadow AI Means

Shadow AI is the use of artificial intelligence tools by employees without the knowledge, approval, or oversight of IT and security teams. The most common examples are public chatbots such as ChatGPT, coding assistants that suggest or generate code inside a developer’s editor, AI writing tools, transcription services, and browser extensions that read page content to summarize or rewrite it.

It also includes AI features that appear inside software your company already pays for. A note-taking app or CRM may switch on an AI assistant in a routine update, and suddenly customer data is flowing to a model provider that was never part of your vendor review.

How It Differs From Shadow IT

Shadow IT is an old problem. Employees have long signed up for file-sharing sites or project tools without asking. Shadow AI carries a sharper risk for one reason: the whole point of the tool is to send it your information. A person doesn’t use a chatbot without typing something into it, and what they type is often the very material you most want to protect. Source code, contract terms, customer records, and financial forecasts make useful prompts precisely because they are sensitive and specific.

Why Employees Turn to Unapproved AI Tools

Most shadow AI comes from good intentions. People find a tool that saves them an hour a day, and they don’t wait for a procurement cycle that might take months.

The scale is large. The 2024 Work Trend Index from Microsoft and LinkedIn found that 75% of knowledge workers were using generative AI at work, and that 78% of those AI users were bringing their own AI tools rather than relying on ones supplied by their employer. Whatever your official position on AI, that research suggests a large share of your staff has already made their own decisions.

Banning these tools outright rarely solves the problem. When access is blocked on the corporate network, people often switch to personal phones or home laptops, where you have even less visibility. The demand doesn’t go away. It just moves out of sight.

The Risks for Sensitive Data

Once information is pasted into an external AI service, your control over it weakens or ends. Depending on the provider and the account type, prompts may be stored, reviewed by staff, retained for a set period, or used to improve future models. Consumer accounts and free tiers often carry different terms from enterprise agreements, and employees rarely read either.

The data most often exposed through unsanctioned AI use tends to fall into a few categories:

• Source code and technical details, including API keys, credentials, and internal architecture pasted in while debugging
• Customer and employee personal data, such as names, contact details, health information, or performance reviews
• Confidential business information, including pricing, merger plans, board materials, and unreleased product details
• Legal and contractual documents shared for summaries or redrafting
• Regulated financial data that falls under specific handling and retention rules

A Well-Known Example

In 2023, Samsung restricted employee use of generative AI tools after engineers reportedly entered internal source code into ChatGPT, according to reporting by Bloomberg. The engineers weren’t acting maliciously. They wanted help with their code. The incident showed how quickly sensitive material can leave a company through an everyday productivity habit.

The cost of these exposures is starting to show up in breach data. IBM’s Cost of a Data Breach Report 2025 found that one in five organizations in its study had experienced a breach involving shadow AI, and that high levels of shadow AI were linked to higher breach costs. The same report found that most organizations hit by AI-related breaches lacked proper access controls for AI.

The Compliance Exposure

Data leaving the company is a security problem. It can also be a legal one, and regulators don’t accept “we didn’t know” as an explanation.

Several frameworks come into play when employees feed data into unapproved AI services:

• GDPR requires a lawful basis for processing personal data and controls on transfers outside the European Economic Area. Fines can reach €20 million or 4% of global annual turnover, whichever is higher.
• The EU AI Act places obligations on organizations that deploy AI systems, including AI literacy requirements for staff that began applying in February 2025.
• Sector rules such as HIPAA in US healthcare or financial services regulations set strict limits on where protected data can go and who can process it.
• Contractual commitments to customers often restrict sharing their data with third parties, and an AI provider counts as one.
• Data residency requirements may be broken if a tool processes data in a region your contracts or local laws don’t allow.

Each of these depends on knowing where data goes. If you can’t show auditors which processors handle personal data, you can’t show compliance. Shadow AI creates processors you never recorded.

Why Discovery Has to Come First

Many organizations respond to shadow AI by drafting a policy first. The policy lists approved tools, bans others, and sets rules for what data can be shared. It sounds sensible, but a policy written without usage data is built on guesses.

You might ban a tool that half the engineering team depends on, which pushes them toward workarounds. You might approve a tool nobody uses while missing the five that are actually in play. You might focus on chatbots and overlook the AI features buried in your existing SaaS stack, which may carry more risk.

Established frameworks put inventory at the start of AI governance for this reason. The NIST AI Risk Management Framework includes, under its Govern function, the expectation that organizations maintain an inventory of their AI systems, and NIST’s Generative AI Profile (NIST AI 600-1) builds on that framework with guidance specific to generative AI, including risks around data privacy, information security, and third-party components. You can’t assess, approve, or restrict what you haven’t found.

Discovery also changes the conversation with the business. When you can show leaders real numbers on which tools are used and by which teams, the discussion moves from abstract fear to concrete decisions about cost, risk, and value.

What Good Discovery Should Tell You

A useful discovery effort answers specific questions. A vague sense that “people use ChatGPT” isn’t enough to act on. Aim to learn:

• Which AI services are in use, including chatbots, coding assistants, image and transcription tools, and AI features inside approved apps
• Which workstations and users are connecting to them, so you can see whether use is spread across the company or concentrated in certain teams
• How often each service is used, since a tool visited once is a different case from one used hundreds of times a day
• Whether access comes through personal or corporate accounts, because the data handling terms often differ
• Which browser extensions and plugins have AI capabilities and what page data they can read
• What kinds of data may be involved, based on the teams and systems where use is heaviest

Knowing the workstation matters more than it might seem. A coding assistant on a developer laptop with access to production repositories carries a very different risk from the same tool on a marketing intern’s machine. Location tells you where to focus first.

Practical Ways to Start

You probably already hold more evidence of AI use than you realize. Most of it sits in systems your team manages today.

Review Network and DNS Logs

Proxy, firewall, and DNS logs show which domains workstations reach. Filtering for known AI service domains gives you a first map of usage by device. Keep the domain list current, since new AI services launch constantly.

Check Endpoints and Browser Extensions

Endpoint management tools can list installed applications and browser extensions across your fleet. Look for desktop AI clients, IDE plugins for code generation, and extensions that request permission to read all site data.

Audit OAuth Grants and SaaS Settings

Employees often connect AI tools to corporate email, calendars, or file storage through OAuth sign-ins. Your identity provider can show which third-party apps have been granted access. Review the admin settings of your main SaaS platforms too, since AI features may have been switched on by default.

Look at Expense Reports

Individual subscriptions to AI tools frequently show up as small recurring charges on corporate cards. Finance data can reveal paid tools that network logs alone might miss.

Ask People Directly

A short, non-punitive survey often surfaces tools and use cases that technical scans miss, especially on personal devices. Make it clear the goal is to understand needs, not to discipline anyone. Employees who feel safe will tell you far more.

Turning Findings Into Policy

Once you have a clear picture, policy decisions become much easier to make and defend. Tools with heavy, legitimate use and acceptable risk can move into an approved list with enterprise agreements that offer better data protection. High-risk tools can be restricted, ideally with an approved alternative offered at the same time so people aren’t left without help. Training can then focus on the real behaviors you’ve observed rather than generic warnings.

Treat discovery as a recurring process, not a one-time audit. New AI tools appear every month, and existing software keeps adding AI features. A quarterly review of logs, extensions, and SaaS settings keeps your inventory accurate and gives you early warning when usage shifts.

Shadow AI is a sign that your employees see real value in these tools. The organizations that handle it well don’t start by saying no. They start by finding out what’s actually happening, then build rules that protect data while letting people keep the tools that genuinely help them work.