Security leaders are asking a question that shouldn’t be this difficult to answer: where is AI actually being used in our organization?
It’s a reasonable question with unreasonable complexity. Unlike software that requires installation or provisioning, many AI tools simply require a browser and a credit card. Others are embedded in existing platforms teams already use. The result is that AI usage often expands invisibly, leaving security teams to discover what’s happening through indirect signals rather than direct observation.
Why is AI usage so hard for security teams to see?
The core challenge is architectural. Traditional enterprise software operates within defined boundaries—applications require server deployments, database connections, network configurations, and identity management integration. Security teams built visibility around these patterns.
AI tools often bypass these checkpoints entirely. A product manager can start using Claude or ChatGPT through a web interface without IT involvement. A development team can integrate an AI code assistant through a browser extension. A sales team can adopt an AI note-taking tool that requires only OAuth authentication to an existing calendar.
The usage itself generates minimal network signatures that distinguish it from standard web traffic. Security tools watching for large data transfers might flag a 50MB file upload, but they won’t necessarily notice thousands of conversational queries that cumulatively share more sensitive information.
This isn’t theoretical. Organizations discover AI usage through budget reviews, compliance audits, or incident investigations far more often than through their security monitoring infrastructure.
What types of AI usage typically go unnoticed?
The AI usage security teams miss tends to fall into predictable categories.
Browser-based tools create the widest blind spot. Engineers use code completion tools. Marketing teams use content generation platforms. Customer support uses AI chat assistants. Finance teams use AI-powered data analysis tools. Each instance might be legitimate, approved, and helpful—but invisible to central security monitoring.
Embedded AI features present a different challenge. When Salesforce, Microsoft 365, or Slack add AI capabilities, the underlying platform remains familiar, but the data flow changes. Information that previously stayed within a controlled environment might now be processed by external AI models. Security teams monitoring the platform often don’t realize when AI features activate.
API integrations compound the visibility problem. Development teams integrate AI services directly into applications. These integrations look like standard API traffic. Without specific inspection, security teams see HTTPS connections, not AI interactions.
Personal accounts create another gap. An employee might use a personal ChatGPT Plus account for work tasks, operating entirely outside corporate infrastructure. From a network perspective, this appears identical to any other personal web usage.
Why traditional security tools don’t show AI usage clearly
The limitation isn’t that existing security tools are inadequate—it’s that they were designed for different problems.
Network monitoring tools excel at identifying unusual traffic patterns, unauthorized connections, and data exfiltration attempts. They struggle with AI usage because it doesn’t look unusual. A conversation with ChatGPT produces HTTPS traffic to a known domain. A code completion request looks like a standard API call. These tools can’t distinguish between an employee checking email and an employee sharing customer data with an AI model.
Endpoint detection tools monitor for malicious software and suspicious process behavior. AI tools don’t trigger these alerts because they operate as legitimate web applications. There’s no malicious process to detect.
Cloud access security brokers (CASBs) provide some visibility, particularly when AI tools use OAuth authentication. However, they typically focus on sanctioned applications. When an employee uses a personal AI account or a tool outside the CASB’s application catalog, visibility drops to zero.
Data loss prevention (DLP) systems can sometimes detect sensitive information being sent to AI platforms, but they operate reactively. By the time DLP catches a policy violation, the data has already been shared. More importantly, DLP rules require someone to first know which AI platforms to monitor—which requires knowing where AI is being used.
What does “AI usage visibility” actually mean?
Visibility means different things depending on the question being asked.
At the most basic level, it means knowing which AI tools are present in the environment. This includes both approved tools and unapproved ones. It includes SaaS platforms, API integrations, browser extensions, and embedded features in existing applications.
The next level involves understanding usage patterns. Who is using these tools? How frequently? For what purposes? This context separates security theater from security intelligence. Knowing that “ChatGPT is being used” provides minimal decision-making value. Knowing that “15 engineers in the authentication team use ChatGPT daily for code review” enables informed risk assessment.
Complete visibility includes understanding data flow. What types of information are being shared with AI systems? Are code repositories being uploaded? Are customer records being analyzed? Are internal strategic documents being summarized?
For most organizations, achieving complete visibility isn’t the immediate goal. The practical target is sufficient visibility to make informed decisions about policy, controls, and risk tolerance. This often means starting with broad discovery before attempting granular monitoring.
What’s the first step to getting visibility into AI usage?
The most effective first step is asking teams directly.
This sounds obvious to the point of being unhelpful, but it works remarkably well. Most employees aren’t trying to hide AI usage—they simply adopted tools that helped them work more efficiently. When security teams approach this as discovery rather than enforcement, people share information freely.
A structured approach helps. Send a brief survey to team leads asking which AI tools their teams use, for what purposes, and with what types of data. Frame it as a partnership: security is trying to enable productive AI use while managing risk appropriately.
Parallel to this conversation, review expense reports and procurement records. AI subscriptions create paper trails. A $20 monthly ChatGPT Plus charge might seem inconsequential individually, but if 200 employees have them, that represents significant usage that should be understood.
Technical discovery complements but shouldn’t replace human inquiry. Network logs can identify traffic to known AI platforms. Browser extension inventories can reveal installed AI tools. OAuth logs can show which employees authenticated to AI services. However, these technical methods often miss context that makes the data meaningful.
The goal isn’t to catch people. The goal is to understand the current state accurately enough to make reasonable decisions about what comes next. Organizations that start with visibility rather than restriction tend to reach workable solutions faster.