Shadow AI is the use of artificial intelligence tools by employees without formal approval or oversight from IT. It's not malicious. Most of the time, people are just trying to get their work done faster. But the gap between good intentions and serious risk is smaller than you'd think.
The Samsung incident
In 2023, Samsung discovered that engineers had uploaded sensitive internal source code to ChatGPT. They were using the AI to help debug and optimise their work. Reasonable motivation, catastrophic outcome.
Samsung's response was an immediate company-wide ban on generative AI tools. But here's the thing: bans rarely work. A survey found that even in organisations with explicit AI prohibitions, half still have unauthorised usage happening.
The Samsung case became a cautionary tale, but it's far from unique. Similar incidents happen daily at companies that never make the news.
Engineering and development teams
Software engineers were early adopters of AI tools, and they remain the heaviest users.
Code generation: Developers paste existing code into ChatGPT or Claude to generate new functions, refactor legacy systems, or write test cases. The problem? That "existing code" is often proprietary. Once it's in the prompt, it's left your network.
Debugging help: When code breaks, the fastest path to a fix is often asking an AI. But the context required to debug effectively usually includes sensitive implementation details.
Documentation: Writing technical documentation is tedious. AI makes it faster. But feeding your entire codebase into an external tool to generate docs means exposing your intellectual property.
A recent Replit incident showed another risk: an AI coding agent deleted a production database during what was supposed to be routine automation. AI-generated code that hasn't been properly reviewed can introduce bugs, security vulnerabilities, or catastrophic failures.
Customer support and service teams
Support staff face constant pressure to resolve tickets quickly. AI offers an obvious shortcut.
The most common pattern is drafting responses. Copy a customer's question into ChatGPT, get a suggested reply, clean it up slightly, and send. Fast and effective, except that customer questions often contain personal information, account details, or specifics about their situation. All of that goes into the AI.
Summarising long email threads before escalation is another common use. A supervisor doesn't want to read fifty back-and-forth messages, so a support agent pastes the whole thread into AI for a summary. That's potentially sensitive customer data entering an external system.
Then there's knowledge lookup. Instead of searching internal documentation, it's faster to ask an AI. But if the AI doesn't have access to your knowledge base, the answers might be wrong. Or the employee starts feeding internal documents into the AI to give it context, which creates an entirely different problem.
According to research from Zendesk, shadow AI usage in customer service has increased by up to 250% year over year in some industries.
Marketing and communications
Marketing teams have embraced AI enthusiastically. Drafting blog posts, social media copy, email campaigns. This is generally lower risk unless the prompts include confidential strategy documents or unreleased product information.
The real problems show up in editing and refinement. A communications specialist at one company uploaded a confidential strategy memo to get AI polish before summarising it for a customer-facing message. That document now sits on third-party servers. Similar patterns appear when teams upload draft press releases or internal memos for wordsmithing.
Competitive analysis creates its own risks. Asking AI to analyse competitor positioning is fine if you're pasting competitor documents you've obtained publicly. But if you're pasting your own strategy documents for comparison, that's data exposure.
Finance and legal teams
These departments handle some of the most sensitive information in any organisation.
Financial modelling: Using AI to build or refine spreadsheet models. The data required to make these models useful is often confidential: revenue figures, projections, client information.
Contract review: Legal teams use AI to speed up document review. But uploading client contracts to external AI tools may violate confidentiality obligations and data protection regulations.
Report generation: Finance teams create reports constantly. AI can draft narratives, summarise data, and format outputs. But the underlying data has to go into the AI to make that happen.
One finance associate was found using an LLM to forecast revenue. Helpful for their workload, problematic for data governance.
HR and people operations
HR teams handle personal employee data, making shadow AI particularly risky.
Job descriptions seem harmless. Using AI to write or polish job postings is common. But if your prompts include salary bands, internal structures, or confidential hiring strategies, you've just shared sensitive compensation data with an external system.
Performance reviews are worse. Getting AI help to draft feedback requires context: specific performance data, compensation details, sometimes sensitive interpersonal situations. That's exactly the kind of information employees expect to remain confidential.
Even policy drafting creates exposure. If you're feeding existing policies into AI for reference while writing new ones, you're sharing internal governance documents with external systems.
The embedded AI problem
Not all shadow AI involves employees actively choosing to use external tools.
AI features are now embedded in software your organisation has already approved. Microsoft Copilot in Office 365. Google's Gemini features in Workspace. Slack AI. Notion AI. Salesforce Einstein.
These features are often enabled by default. Your team might be using AI without consciously deciding to, simply because it's integrated into their daily tools.
This creates a different detection challenge. You're not looking for visits to chatgpt.com. You're trying to understand which AI features are active across your entire software stack.
Why this matters for your business
The common thread across all these examples: data leaving your control.
Once information goes into an external AI system, you lose control. You don't know whether it's being stored. You don't know if it's being used to train future models. You don't know who else might eventually access it, or whether you can ever get it deleted.
For businesses with regulatory obligations (GDPR, HIPAA, industry-specific compliance), this creates immediate legal exposure. For businesses with client confidentiality requirements (law firms, consultancies, financial services), it's a potential breach of trust.
And for any business with competitors, it's intellectual property walking out the door.
What to do about it
Awareness is the first step. If these examples sound familiar, you probably have shadow AI happening in your organisation right now.
The response isn't to ban AI. That approach has been tried, and it fails more often than it succeeds. People find workarounds, use personal devices, or simply ignore the rules.
The better approach is visibility. Know which AI tools your team is using. Understand what data is going into them. Set policies based on evidence rather than guesswork.
Then provide sanctioned alternatives. If your team needs AI to be productive, give them options that come with appropriate governance. Enterprise AI tools with data protection guarantees, internal AI assistants, approved workflows.
Shadow AI isn't going away. The question is whether you'll manage it proactively or find out about it after something goes wrong.