The Question
Your information security team ran a network traffic analysis as part of a broader data security review. The results showed that employees across your organization are actively submitting data to ChatGPT, Claude, Gemini, Perplexity, and at least a dozen other AI tools using personal accounts not covered by any enterprise data processing agreement. The legal team is asking about GDPR exposure. The CISO is asking about data leakage. HR is asking about the policy implications.
The immediate instinct in most organizations is to issue a policy memo reminding employees that unauthorized AI tool use is prohibited, and to add AI tool domains to the web filter blocklist. This approach will fail for the same reason it always fails: employees adopted these tools because they are genuinely useful, and prohibition without alternative drives the activity underground rather than eliminating it. The employees who stop submitting data to ChatGPT through the corporate network will start submitting it through their phones.
The more productive question is not "how do we stop employees from using unauthorized AI?" It is "why are employees using unauthorized AI tools, what use cases are they covering, and how do we build a governed path to those same capabilities?" Shadow AI is a symptom of a governance program that is not moving at the speed of the business problems employees are trying to solve.
Shadow AI governance is not about prohibition — it is about understanding why employees adopted unauthorized AI tools and building a governed path to the same capability.
Why This Matters Now
The scale of enterprise shadow AI is no longer anecdotal. Salesforce's 2025 State of IT report found that 55% of employees use AI tools their employer has not approved, up from 37% the prior year. Microsoft's Copilot usage data across M365 tenants shows consistent patterns: in organizations without structured AI access programs, employees using Teams, Outlook, and Edge are supplementing enterprise tools with consumer AI products for tasks the enterprise tooling does not cover well.
IBM's 2025 AI in Action study found that in organizations where governance programs are perceived as slow or blocking, shadow AI usage rates are significantly higher — and that the highest shadow AI adoption is concentrated in the knowledge-intensive functions where AI provides the most value: legal, finance, strategy, research, and engineering.
The data exposure consequence is not theoretical. In May 2023, Samsung employees were found to have submitted proprietary source code to ChatGPT for debugging assistance. The incident prompted Samsung to ban AI tools company-wide — a policy that was broadly reported as the kind of overreaction that reduces security theater while not addressing the underlying dynamic. By late 2023, Samsung had reversed course and begun developing authorized AI tooling for the use cases that drove the shadow AI adoption in the first place.
Several GDPR enforcement actions in 2024 and 2025 cited employee use of third-party AI tools without appropriate data processing agreements as contributing factors in data breach and data transfer violations. The regulatory exposure is not limited to large enterprises: any organization subject to GDPR that has employees submitting personal data of EU residents to consumer AI tools without DPAs has a potential compliance issue.
The threat landscape has also evolved. In 2025, security researchers documented cases of malicious AI tools designed to appear as consumer productivity AI while exfiltrating the content submitted by enterprise users. Shadow AI is not just a compliance and data leakage risk — it is an expanding attack surface.
What the CURVE™ Data Shows
The 2026 Stackcurve AI Governance CURVE™ Report evaluated tools and platforms for shadow AI discovery, sanctioned AI program management, and AI usage governance. The evaluation covered CASB-based discovery tools, AI governance platforms with shadow AI management capabilities, and purpose-built AI access management solutions.
Netskope ranked as a Leader for shadow AI discovery through its CASB platform, with AI-specific application categorization that identifies consumer AI tools in use across the enterprise network and provides usage analytics by department and data sensitivity. Zscaler provides comparable shadow AI discovery capabilities integrated with its SASE platform, with the advantage of covering mobile and off-network traffic for organizations fully deployed on Zscaler. Microsoft Defender for Cloud Apps is the default choice for M365-heavy organizations, with native AI application discovery and the ability to create shadow AI usage reports across the M365 user population.
For sanctioned AI program management — the affirmative governance work of maintaining an approved AI register and fast-tracking low-risk tool approvals — Credo AI and OneTrust AI Governance both provide workflow tools for AI intake, risk assessment, and approval management. Torchon (early stage) focuses specifically on AI access management — governing which employees can access which AI tools with what data — and has gained traction in financial services organizations with strict data segregation requirements.
The full vendor rankings are in the 2026 Stackcurve AI Governance CURVE™ Report — free to download.
The Gap Most Buyers Miss
The prohibition-without-alternative failure mode
The most common shadow AI governance response is policy restriction: issue a use policy prohibiting unauthorized AI tools, add enforcement to acceptable use training, and optionally add technical controls. This approach addresses the symptom — visible unauthorized use — without addressing the cause — the absence of authorized tools that meet the business use case.
The result is predictable: employees who were using AI productively in visible ways move to less visible methods. Consumer AI access via mobile browsers, personal laptops, home networks, or third-party contractors passes outside the visibility of corporate CASB controls. The data exposure risk may increase even as the observable policy violation decreases.
The governance approach that works starts with use case understanding: what specific tasks are employees using shadow AI to accomplish, how are those tasks currently unsupported by authorized tooling, and what is the fastest path to providing authorized tooling that covers those cases? Samsung's experience — ban, observe that the underlying use cases persisted, then develop authorized tooling — is the slow version of a journey that governance programs can compress.
The segmentation problem: not all shadow AI is equal risk
Shadow AI is not a uniform category. An employee using an AI writing assistant to draft internal communications presents different data exposure risks than an employee submitting customer PII to an AI tool for data analysis, which presents different risks than an engineer submitting proprietary source code for debugging assistance.
Effective shadow AI governance requires risk segmentation, not blanket treatment. The governance response to an employee using Grammarly's AI features for email drafting should be different from the governance response to an employee submitting financial model data to ChatGPT. Treating all shadow AI as equivalent maximum-risk behavior produces governance programs that employees experience as disproportionate and work around.
The approved AI register: governance infrastructure that prevents shadow AI
The single most effective structural intervention against shadow AI is a well-maintained, easily accessible register of approved AI tools with clear use case guidance. Employees use shadow AI largely because they do not know what is approved, do not believe the approval process is fast enough to be useful, or have a use case that the approved toolset does not cover.
An approved AI register addresses the first problem directly. A fast-track approval pathway — ideally with a 48-72 hour turnaround for low-risk tools assessed against a standard criteria checklist — addresses the second. Regular business unit engagement to understand unmet AI use cases addresses the third.
The off-boarding and contractor risk that shadow AI creates
Shadow AI governance has an often-overlooked personnel dimension. Employees using personal account AI tools to do work create a shadow knowledge repository: AI conversations, generated documents, and data submissions that exist outside enterprise systems and are not subject to enterprise data retention, discovery, or off-boarding processes. When an employee leaves, the work product they created using a personal AI account leaves with them — or more accurately, it may remain accessible in the vendor's infrastructure under the employee's personal account. Governance programs should address AI-generated work product in off-boarding procedures.
Questions Your Buying Team Should Be Asking
1. What AI applications are employees in our organization currently using, categorized by data sensitivity level and department — and does the discovery tool cover off-network and mobile traffic?
The starting point for shadow AI governance is accurate inventory. Most organizations significantly underestimate shadow AI scope until they run discovery tooling. The question about off-network and mobile coverage is important: CASB-based discovery that only covers corporate network traffic misses a substantial fraction of consumer AI usage, particularly in hybrid work environments.
2. For each sanctioned AI tool, what data is transmitted to the vendor's infrastructure, and is it covered by an enterprise DPA that meets our GDPR, CCPA, and sector-specific compliance requirements?
This question applies to both shadow AI being evaluated for sanction and to AI features embedded in existing SaaS products. Many organizations discover during this exercise that products they consider "enterprise-grade" — including Microsoft Copilot, Google Workspace AI features, and Salesforce Einstein — transmit data in ways that require updated DPAs or configuration changes to meet privacy compliance requirements.
3. What is the current approval process for a new AI tool, what is the average time from request to decision, and how does that compare to the time it takes an employee to create a personal account and start using an unauthorized alternative?
This question surfaces the governance speed gap that drives shadow AI adoption. If the answer is "4-6 weeks," and creating a free ChatGPT account takes 3 minutes, the policy will not hold against motivated employees. Fast-track pathways, pre-approved tool categories, and risk-tiered approval processes are the operational responses to this gap.
4. For business units with high shadow AI rates — which are typically those with the most genuine AI use cases — what specific tasks are employees trying to accomplish, and what authorized tools exist for each?
This question shifts the conversation from security posture to use case coverage. Shadow AI follows the value: the highest shadow AI usage is in the business units where AI provides the most genuine productivity benefit. Understanding the specific use cases is the prerequisite for building authorized coverage.
5. What DLP controls are in place for data submitted to authorized AI tools, and do those controls extend to preventing sensitive data classification from reaching unauthorized AI endpoints?
Shadow AI governance is incomplete without data controls on the authorized side as well. DLP policies that prevent high-sensitivity data classifications from being submitted to any AI tool — authorized or not — provide a backstop when policy and access controls are insufficient. Ask specifically how the DLP policy is configured for AI endpoints versus general web traffic.
The Stackcurve Take
Shadow AI is a measure of the gap between the AI governance program enterprises have built and the AI use cases employees have found valuable. Closing that gap requires two parallel tracks: reducing the demand for shadow AI by providing fast, well-supported access to AI tools that cover the use cases driving adoption, and reducing the risk surface of the shadow AI that will continue to exist regardless of policy.
Organizations that have successfully addressed shadow AI at scale share a common pattern: they ran discovery before they ran policy, they engaged business units as partners rather than as compliance subjects, and they built approval processes fast enough to compete with the alternative of just creating a personal account. The technical controls — CASB discovery, DLP for AI endpoints, access management — are enablers of that program, not substitutes for it.
The CISO who treats shadow AI as a security problem to be blocked will be in the same position in 12 months, with employees using more sophisticated workarounds. The governance leader who treats shadow AI as a signal about unmet business AI needs will be running a materially more effective program by then.
The 2026 Stackcurve AI Governance CURVE™ Report covers shadow AI discovery tools, approved AI program management platforms, and governance frameworks for employee AI access. Download it free →
Stackcurve Advisory Briefs are independent research. No vendor pays for placement, tier assignment, or editorial influence. The CURVE™ methodology is disclosed in full at stackcurve.net/research/methodology.