Shadow AI

Shadow AI Definition and the 4 Common Types

In short

Shadow AI is the unsanctioned use of AI tools by employees outside IT, security, and compliance oversight. The four most common types are public generative AI tools, AI coding assistants, AI meeting recorders, and embedded AI features inside SaaS products.

Shadow AI is the AI-era version of shadow IT. It is what happens when employees adopt AI tools faster than the organization can approve, monitor, or govern them.

A finance analyst pasting a forecast into ChatGPT, an engineer using an AI coding assistant on a customer codebase, a sales rep recording a call with an AI notetaker — all of these are shadow AI when they happen outside policy.

Shadow AI shows up in four common patterns:

• Public generative AI tools. Employees using ChatGPT, Claude, Gemini, or similar to draft documents, summarize meetings, write code, or analyze data — often pasting confidential information into prompts.
• AI coding assistants. Developers installing AI coding tools on their laptops or in their IDEs to suggest, complete, or refactor code, sometimes against an internal codebase that was never cleared for external processing.
• AI meeting recorders and notetakers. Bots that join calls, transcribe them, and store transcripts — often without the consent of all participants and outside the data residency rules the company is contractually bound to.
• AI features inside approved SaaS. Generative AI capabilities silently added to CRM, productivity, design, or analytics tools the organization already uses, switched on by default and never reviewed.

The defining characteristic is not the tool itself. It is the absence of governance: no risk assessment, no data classification check, no approval, no monitoring, no incident path.

Shadow AI is closely related to AI governance. It is the most visible symptom of a governance program that has not yet caught up to how fast generative AI is being adopted on the front line.

Why Shadow AI Exploded After 2022

In short

Shadow AI exploded because public large language models became free, instantly useful, and accessible from any browser or phone at the same moment that productivity pressure on employees increased — while IT and compliance approval cycles stayed the same length.

For most of the history of enterprise IT, adopting a new tool required someone to procure, install, and configure it. Shadow IT existed, but the friction was real.

Public LLMs removed that friction in a single quarter. From late 2022 onward, any employee with a browser could use a frontier model. There was no install, no procurement, often no cost, and no learning curve.

Three forces collided at once:

• Accessibility. Free or low-cost public LLMs available from any device, any network, in any language.
• Productivity pressure. Employees under pressure to do more with less, with leadership often signaling that AI use is encouraged.
• Slow governance cycles. Procurement, security review, and DPIA cycles measured in weeks or months while AI tools change weekly.

The early regulatory signals arrived almost immediately. In March-April 2023, the Italian Garante per la protezione dei dati personali briefly ordered a halt to ChatGPT processing of Italian users' data over GDPR concerns — one of the first public signs that public LLMs would not be left ungoverned.

Then came the Samsung incident in April 2023, when internal reports surfaced that engineers had pasted confidential source code and meeting notes into ChatGPT. The story moved shadow AI from a niche IT problem to a board topic within days.

The lesson for most CISOs was not "ban AI." It was "we no longer know what AI is being used here." That is the real definition of shadow AI risk.

The Samsung Incident (April 2023)

The Italian Garante and GDPR

The 5 Categories of Shadow AI Risk

In short

Shadow AI creates five risk categories: data leakage (confidential information leaving the organization), regulatory non-compliance (GDPR, EU AI Act, sectoral rules), IP loss (training data and ownership disputes), security exposure (malicious prompts and supply-chain risk), and unverified model accuracy (hallucinations affecting decisions).

Not every shadow AI use case is high risk. But because shadow AI is, by definition, ungoverned, the organization cannot tell the difference between the harmless and the catastrophic.

1. Data leakage. The most discussed risk and the one that triggered the Samsung incident. Confidential code, customer data, financial information, or strategy documents pasted into public LLMs may be retained, used to improve models, or exposed in ways the organization cannot control.

2. Regulatory non-compliance. GDPR applies to personal data regardless of which tool processes it. The EU AI Act (Regulation 2024/1689) applies to AI placed on the EU market or whose output is used in the EU. Sector-specific rules (finance, health, public sector) impose additional duties that shadow AI bypasses by definition.

3. Intellectual property loss. Two risks at once. Confidential IP can leak into training data. Generated output may have unclear copyright or contamination from copyrighted training data, creating disputes about ownership of the result.

4. Security exposure. Shadow AI tools expand the attack surface. Risks include prompt injection, malicious AI assistants installed on developer machines, unvetted third-party AI processing customer data, and supply-chain compromises through AI plug-ins.

5. Unverified model accuracy. Public LLMs hallucinate. Without governance, employees may rely on incorrect outputs for legal research, customer responses, regulatory filings, or operational decisions. The error compounds when the output is then cited internally as if it were authoritative.

These five categories cover most of the documented harms. The shared feature is that the organization has no visibility, no early warning, and no documented evidence trail.

Source: Alice Labs, based on EU AI Act, GDPR, NIST AI RMF

How to Audit and Manage Shadow AI

In short

The proven approach to managing shadow AI is not a ban. It is an allowlist of approved AI tools, AI literacy training under EU AI Act Article 4, network and endpoint monitoring to detect unsanctioned use, and a fast, low-friction approval path so the safe option is also the easy option.

Bans tend to fail because they do not address why shadow AI emerged: employees needed a tool, the approved path was too slow, and an unapproved one was free.

The programs that work treat shadow AI as a signal, not a sin. The goal is to surface what is already happening, give people an approved version, and govern it from there.

1. Surface what already exists. Use network monitoring, SaaS discovery tools, browser telemetry, and structured employee surveys to inventory which AI tools are actually in use, by whom, and for what.

2. Publish an allowlist. Define a short list of approved AI tools at appropriate tiers (enterprise contracts, no-training clauses, EU data residency where required). Make the allowlist obvious and easy to use.

3. Write an acceptable use policy. Cover what data can and cannot be used, which tools are approved, what review is required for new tools, and how to report concerns. Keep it short.

4. Deliver AI literacy training. Required under Article 4 of the EU AI Act since 2 February 2025. Cover prompt hygiene, data classification, hallucination, and incident reporting — not just policy slides.

5. Monitor and iterate. Track AI tool usage, run periodic audits, and feed findings back into the allowlist and policy. Shadow AI is not a one-off project. Tools and risks change quarterly.

The strategic move is to integrate this into the broader AI governance program rather than running it as a standalone CISO initiative. Governance, security, legal, and HR all have a role.

Discovery Tactics That Actually Work

Designing an Acceptable Use Policy That Sticks

The Alice Labs Shadow AI Audit Methodology

In short

The Alice Labs Shadow AI Audit is a four-step proprietary methodology — discover, classify, remediate, and govern — that surfaces unsanctioned AI use, classifies each instance against EU AI Act and GDPR exposure, replaces high-risk tools with approved alternatives, and integrates the result into the wider AI governance program.

The Alice Labs Shadow AI Audit methodology is shaped by 100+ Nordic enterprise implementations across financial services, manufacturing, public sector, and B2B SaaS. It is designed to surface shadow AI without freezing the productivity gains employees are already enjoying.

Step 1: Discover. Combine network and SaaS telemetry, endpoint inventory, browser extension audits, and structured employee surveys to build a single inventory of AI tools actually in use. Capture tool, owner, use case, and data category.

Step 2: Classify. Score each instance against EU AI Act risk categories (Unacceptable, High-risk, Limited risk, Minimal risk) and GDPR exposure. Identify which instances trigger DPIA or FRIA obligations under Article 27.

Step 3: Remediate. Replace high-risk shadow AI with approved tier alternatives (enterprise contracts, no-training clauses, EU data residency). Decommission tools that cannot be safely brought into governance.

Step 4: Govern. Integrate the result into the wider AI governance program: allowlist, acceptable use policy, AI literacy training under EU AI Act Article 4, intake for new AI tools, and quarterly re-audit.

The audit pairs with the Alice Labs EU AI Act Readiness Assessment and the Alice Labs LLMO Citation Benchmark for organizations that also need a generative-AI surface view of risk.

For broader context, see the EU AI Act compliance checklist 2026 and the Alice Labs Implementation Index 2026, which both connect shadow AI findings to the wider governance picture.

What a Shadow AI Audit Delivers