Nidhi Jain is CEO & Founder of CloudEagle.ai, a platform helping enterprises govern SaaS, identities, and AI agents at scale.gettyMost CISOs discover their shadow AI problem through a procurement report. Someone runs a query expecting one or two approved AI subscriptions. Instead, they find 40. Then, expense reports double the number. That moment signals how badly traditional security tooling has fallen behind employee behavior.Silent Adoption: How Shadow AI Went Mainstream Without ITMicrosoft's 2024 Work Trend Index found that 78% of surveyed AI-using employees brought their own tools to work. Microsoft's October 2025 U.K. research is sharper, finding that 71% of surveyed employees had used unapproved consumer AI tools at work and that 51% were continuing to do so every week.Salesforce's 2023 Generative AI Snapshot Research, covering 14,000 global workers, found that more than half of responding AI adopters used unapproved tools at work without employer approval.Employees didn't wait for the policy. Policy is catching up to behavior that has been normalized for two years.Security Stacks Built For A Different ProblemCASB looks for unsanctioned SaaS. DLP catches structured data egress. Endpoint agents flag suspicious processes. None of them were built to detect a marketing analyst pasting customer contract terms into a browser AI tool that returns a summary in three seconds.Cisco's 2025 Cybersecurity Readiness Index makes the visibility gap explicit:• 60% of IT teams can't see the prompts employees make to GenAI tools.• 60% lack confidence detecting unregulated AI deployments.• 86% experienced an AI-related security incident in the previous year.Data leaving the enterprise through an AI tool doesn't look like a breach. It looks like a browser tab.Every Ban You Issue Buys An Attacker Another WeekThe instinct most security teams have is to block. It backfires. When sanctioned tools are slower than what an employee can open in a browser tab, policy becomes a suggestion. Bans push usage to personal accounts, where IT can't see it and the company can't capture the value.Enterprises that reduced shadow AI stopped banning. Instead, they deployed sanctioned alternatives matching consumer tools and then enforced data boundaries inside them. Governance over prohibition is the principle that works.The Three Layers Of Shadow AI Most Security Teams MissShadow AI is hard to govern because most CISOs are still looking at one layer of it. There are three:1. Standalone ToolsChatGPT, Claude, Gemini and browser-based AI assistants make up the layer that security teams focus on because it's the easiest to name. It also consumes the least sensitive data because employees know it's unsanctioned, so they self-censor.2. Embedded AI Inside The Tools You Already ApprovedCopilot inside Microsoft 365. Einstein inside Salesforce. AI features inside Zoom, ServiceNow, Slack and Atlassian. Your vendor shipped these into your existing contracts, often without notice. The vendor was reviewed, but the AI inside the vendor wasn't.3. Agents And Integrations Employees Are BuildingA workflow pulling customer data from Salesforce, running it through an LLM and posting the summary to Slack. A custom GPT trained on internal documents. An MCP server connecting an AI assistant to a production database. These are invisible to every security tool not built to see them, and they carry persistent credentials.Most enterprises are still fighting the first layer. The second layer generates most of the actual data exposure. With Gartner, Inc. forecasting that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, the third layer is about to become the largest.The First 90 Days Of An AI Visibility ProgramThe most common mistake I see enterprises make is starting with a policy. Policies without visibility are just words on a page. The order that works is discovery first, classification second and enforcement third:1. DiscoveryDiscovery requires pulling from sources that most security teams haven't connected: SSO and OAuth grant logs, browser extension audits, corporate card and expense reports, help desk tickets and a candid employee survey. Together, these five inputs typically surface the vast majority of AI usage within 30 days. The gap between what leadership assumes and what those inputs show is where strategy turns operational.2. ClassificationWhat data is flowing through the approved AI tools? Group usage by data sensitivity, not by tool brand. A junior employee using ChatGPT for grammar checks is a different risk than a finance lead uploading forecasts.3. EnforcementThis only works if sanctioned alternatives already exist. The most effective sequence I've seen is deploying an enterprise AI tool with data boundaries built in before restricting personal accounts. When reversed, employees route around it.The Concerns That Slow Programs DownThree objections come up in almost every conversation with a security leader starting this work:1. Visibility Feeling Like SurveillanceEmployees are already using these tools. A visibility program is about understanding usage. The companies that communicate this get cooperation. The ones that lead with restrictions get workarounds.2. BudgetMany security teams assume AI visibility requires a new platform. Most of the discovery work uses tools already in the environment. The investment is analyst time.3. OwnershipAI visibility doesn't fit cleanly under IT, security, legal or procurement, so no one owns it. The companies moving fastest name a cross-functional AI governance lead reporting to the CISO across legal and procurement.What The Industry Owes EnterprisesThe pace of AI adoption has outrun the guardrails that vendors and standards bodies have shipped, and enterprises are absorbing the cost.SaaS vendors should disclose embedded AI features by default, in plain contract language, before they ship. Right now, most updates arrive silently. Frameworks such as NIST AI RMF are moving in the right direction, but they still treat AI visibility as an implementation detail. AI providers should build enterprise-grade audit logging in by default.The companies leading on AI visibility are effectively subsidizing the ecosystem's immaturity. That won't scale. The AI is already inside the enterprise. Visibility is the only precondition for control, and the enterprises that build it now will define what governance looks like for everyone else.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
Why Shadow AI Has Already Outpaced Your Security Stack
The pace of AI adoption has outrun the guardrails that vendors and standards bodies have shipped, and enterprises are absorbing the cost.
TL;DRAI
71% of employees use unapproved AI tools weekly, yet traditional security stacks can't detect shadow AI because it flows through ChatGPT or embedded in approved vendors (Salesforce, M365). Governance over bans works—enterprises deployed sanctioned alternatives with data boundaries before restricting personal accounts, as 40% of enterprise apps will ship task-specific agents by 2026.
987 words~4 min read
