What Is Shadow AI? Causes, Compliance Gaps, and How to Manage It

Shadow AI refers to any generative AI tool, chatbot, browser extension, or AI-powered application that employees use to conduct company business without explicit approval, oversight, or governance from IT, security, legal, or compliance teams. 

It’s the natural evolution of shadow IT. But it moves faster, touches more sensitive data, and leaves behind almost no trace unless an organization has specifically planned for it.

Common Shadow AI Tools

The shadow AI compliance risks facing most organizations center on a small set of widely available consumer tools, including:

• ChatGPT — free and personal-account versions, used for drafting, summarizing, and analyzing internal documents
• Google Gemini — embedded in personal Gmail and Google Workspace accounts
• Microsoft Copilot — used outside of a licensed, governed enterprise deployment
• Claude, Perplexity, and other consumer chatbots
• AI-powered browser extensions  — summarize webpages, emails, or documents
• AI meeting transcription tools — that record and process conversations, sometimes including confidential discussions
• AI coding assistants — connected to production systems or proprietary codebases

Shadow IT governance frameworks were built around discoverable assets. Security teams could run a network scan, review firewall logs, or audit SaaS subscriptions to find unauthorized AI tools. Shadow AI evades this approach. An employee doesn’t need to install anything. They can open any AI tool in a browser tab on a personal or work device, and there is often no log, no admin console entry, and no IT ticket to flag it.

Shadow IT is where an employee would install an unapproved app or store files in a personal cloud account, shadow AI is where an employee is having a live, interactive conversation with a third-party model, one that can ingest, process, and sometimes retain whatever information is typed into it. The tool doesn’t just store the data. In many cases, it actively processes and learns from it.

A recent study by BlackFog found that 49% of employees use AI tools their employer hasn’t approved. Of those, 58% rely on free versions that lack enterprise-grade data governance or security controls. That same research found 60% of employees said they would accept the risk of using an unapproved AI tool if it meant finishing a project on time, and 63% considered it acceptable to use AI without IT oversight when no approved alternative existed.

Employees aren’t deliberately circumventing policy. They’re filling a gap that compliance and IT teams haven’t closed yet.

Shadow AI doesn’t fail in one place. It creates exposure across multiple fronts. Some of the risks that comes with Shadow AI include:

Unauthorized processing of sensitive data.

Employees submit proprietary or regulated information into external AI systems without fully knowing where it goes. Most assume the interaction is private and temporary, like a search query. In reality, the data may be logged, stored, or reused by the provider in ways the company never agreed to. Once that submission happens, the organization loses certainty over who has access to it.

Data breaches

Breaches involving AI models or applications were reported by 13% of organizations. Every shadow AI tool an employee adopts is an unvetted vendor: no security agreement, no place in incident response planning. If that vendor is breached, the organization’s data goes down with it, often with no one on the compliance or security team aware of the exposure. There’s no notification process, because there was never a contract requiring one. By the time a data breach becomes public, the company often learns about its own exposure from a headline, not a vendor disclosure.

Data leakage through storage and retention.

Many AI tools retain the inputs and outputs of every conversation on servers the organization can’t inspect. That includes customer records, financial information, and internal strategy documents. Retention policies for these tools are often vague, undisclosed, or entirely absent, leaving data sitting on third-party servers well beyond the interaction itself. Because there’s no internal record of what was submitted, compliance teams can’t even scope the potential leakage if a concern is later raised.

Regulatory Exposure Across Industries

Shadow AI data exposure intersects with existing, industry-specific compliance obligations that most organizations already spend significant resources managing:

• Financial services firms subject to SOX, FINRA, and SEC rules must produce unaltered records of any AI-assisted communication tied to financial reporting or internal investigations.
• K-12 districts and higher institutions face overlapping FERPA, FOIA, and sunshine law requirements. Pasting a student’s disciplinary record into a consumer AI tool can trigger the same liability as any other unauthorized disclosure
• Healthcare organizations governed by HIPAA face similar exposure whenever patient information is entered into an AI tool not covered by a signed Business Associate Agreement.
• Government agencies are subject to FOIA and state public records laws, which apply to AI-assisted communications the same as any other electronic record.
• Multinational organizations must navigate overlapping data sovereignty rules under GDPR, CCPA, PIPEDA, and more

These risks rarely show up on their own. A single unapproved AI conversation can trigger a data breach, a regulatory violation, and a records gap all at once. Most compliance teams won’t know it happened until it’s too late to fix.

Shadow AI in the Workplace:

• 60% of AI security incidents exposed data, 31% caused operational disruption.
• Shadow AI has contributed to 1 in 5 breaches, but only 37% of orgs have policies to manage it.
• Heavy shadow AI use added $670K to breach costs on average.
• 63% of breached organizations have no AI governance policy and even fewer audits for unsanctioned AI.

The numbers make it clear. Most of the workforce is bringing AI tools into work without IT approval, with adoption climbing sharply over the past two years. A large share of employees already admit to typing confidential company information, customer records, financial data, internal strategy notes, into public AI tools they were never authorized to use. Much of that activity happens on personal accounts, on personal devices, or through browser extensions no security team ever reviewed.

The financial stakes have grown alongside the behavior. Data breach costs hit record highs in 2024, and a meaningful share of organizations have already experienced an AI-related security incident. Yet governance hasn’t kept pace. Most organizations still lack a formal way to detect unsanctioned AI use, can’t produce a full inventory of the AI tools running across their workforce, and have no dedicated policy addressing what employees can and cannot do with these tools.

The pattern across every data point is the same. Adoption is outpacing governance, and the gap widens every quarter instead of closing.

A governance framework isn’t complete without a system for capturing the records it produces. Closing the records capture responsibility gap requires the same foundational discipline that email archiving, social media archiving, and text message archiving already apply, extended to cover AI-assisted communication:

• Real-time capture of communications as they happen, rather than relying on employees to self-report or manually save AI conversations.
• Tamper-evident storage that preserves records in an unalterable, evidentiary-quality format, so a record produced during an audit or litigation hold can be trusted as authentic.
• Searchable indexing so compliance and legal teams can locate relevant records quickly during an investigation, audit, or public records request, rather than manually searching through disconnected tools.
• Defined retention policies that apply consistently across every communication channel, including AI-assisted ones, rather than treating AI as an exception to the retention schedule.
• Role-based access controls that determine who can view, export, or delete records, closing off the kind of unsupervised deletion authority that regulators specifically flag as a governance failure.

Compliance teams don’t need to solve shadow AI compliance risks overnight, but they do need a structured plan. The organizations can follow a framework to

Step 1: Discover What’s Already Happening

Most companies significantly underestimate the scale of unauthorized AI use inside their own organization. Before writing a single policy, compliance and IT teams need a realistic picture of which AI tools employees are already using, informally survey department heads, and review network traffic for known AI domains.

Step 2: Provide Approved, Governed Alternatives

If employees are using ChatGPT because there’s no sanctioned option, provide an enterprise version with data protections and clear data-handling terms. If developers rely on AI coding assistants, license a version with appropriate security controls rather than leaving them to find their own. Meeting the underlying need safely is far more effective than removing the option entirely.

Step 3: Write Specific, Actionable AI Policies

Vague guidance doesn’t change behavior. Define what data categories can never be entered into any AI tool, whether that’s customer PII, student records, patient information, or unreleased financial data, and specify which sanctioned tools are approved for which use cases. Review and update these policies at least twice a year. AI tools change fast, and a policy written in the beginning of the year can be outdated by summer.

Step 4: Train Employees on the Reasoning, Not Just the Rule

Most shadow AI usage stems from a lack of understanding about how these tools store and process data, not deliberate policy violations. Training that explains why a “deleted” AI conversation may still exist somewhere, and why that matters for a records request or litigation hold, changes behavior more effectively than a policy document employees skim once and forget.

Step 5: Archive, Monitor, and Adapt Continuously

This is the step compliance teams most often skip, and it’s the one that closes the records capture responsibility gap discussed above. AI tools evolve quickly, and a governance framework built once and left untouched will fall behind within a year. Ongoing monitoring, paired with a records archiving system that extends across every communication channel employees actually use, keeps the framework current as new tools and use cases emerge.

• Shadow AI moves faster than shadow IT and leaves almost no trace. Employees don’t need IT approval to open a browser tab and start typing.
• Nearly half of employees already use unapproved AI tools. Most rely on free versions with no enterprise-grade data controls.
• Every ungoverned AI tool is an unvetted vendor. A breach on their end becomes your exposure, with no contract and no notification process.
• FERPA, HIPAA, SOX, and FOIA apply to AI conversations the same way they apply to any other record. The tool doesn’t change the obligation.
• Banning AI tools doesn’t stop usage. It pushes usage onto personal devices and networks where visibility drops to zero.
• Providing a sanctioned, governed AI tool works better than prohibition alone. Employees need a legitimate alternative, not just a rule.

What is shadow AI?

Shadow AI is the use of AI tools like ChatGPT, Gemini, or Copilot by employees to conduct business tasks without formal approval, visibility, or governance from IT, security, legal, or compliance teams.

Is shadow AI the same as shadow IT?

They’re related but not identical. Shadow IT typically involves unapproved software or storage tools. Shadow AI involves interactive systems that actively process, analyze, and sometimes retain the data entered into them, creating a broader and faster-moving set of data exposure risks.

Why doesn’t banning AI tools solve the problem?

Blanket bans push usage onto personal devices and unmonitored networks, where visibility drops to zero. Employees who need AI to stay competitive will find a way to use it regardless of network-level restrictions, so governance and approved alternatives are more effective than prohibition alone.

Does FERPA, HIPAA, or SOX apply to AI conversations?

Yes, these regulations apply based on the content and context of a communication, not the tool used to create it. If a conversation involves a protected education record, patient information, or financial reporting data, existing regulatory obligations apply regardless of whether AI was involved in creating it.