Shadow AI is employees using AI tools at work without their company's knowledge or approval. It's already happening at your organization — research puts the rate at over 47% of employees who use AI tools at all. Here's what it means, why it matters, and what HR can actually do about it.
Shadow AI (definition): the use of artificial intelligence tools — chatbots, AI writing and coding assistants, AI note-takers, and AI features inside other apps — by employees for work, without the knowledge, approval, or oversight of their employer's IT or security function. The term is the AI counterpart of "shadow IT."
Most conversations about shadow AI are written for CISOs, IT directors, and security teams. This one isn't. If you're in HR, operations, or people management and you've been asked about your company's AI policy — or you know you should have one but haven't gotten there yet — this is for you.
No acronyms, no technical jargon. Just a clear explanation of what shadow AI is, what the real risks look like in practice, and what response actually works.
The meaning of shadow AI is straightforward: it refers to the use of artificial intelligence tools by employees without the knowledge or approval of their employer. It's the AI version of "shadow IT" — the decades-old problem of employees using personal Dropbox accounts, installing unapproved software, or using consumer tools for work purposes without IT oversight.
The most common shadow AI scenario today looks like this: an employee discovers that ChatGPT, Claude, Gemini, or another AI tool would save them hours of work per week. They sign up with their personal email. They use their own free-tier account. They start doing their job better and faster — and they never mention it to anyone, because no one asked, and there's no policy that says they should.
That employee is not being malicious. In most cases, they're doing their job well. The problem is structural: the company has no visibility into what data went into those sessions, what the AI tool's terms say about data use and training, or what would happen if something went wrong.
Shadow AI shows up differently across roles. Here are the patterns that HR and operations teams encounter most often:
Shadow AI is rarely one rogue app — it's a handful of genuinely useful tools that employees adopt on their own because they make work faster. These are the categories that show up most often, and the data risk each one tends to carry when used on a personal or free-tier account:
| Tool / category | What employees use it for | Typical shadow AI risk |
|---|---|---|
| ChatGPT, Claude, Gemini | Drafting, summarizing, analysis, Q&A | Confidential text pasted into a free-tier account with no data-processing agreement |
| Microsoft Copilot / Google Gemini in apps | AI features inside email, docs, and spreadsheets | AI acting on company data through personal logins or unmanaged tenants |
| AI note-takers (Otter.ai, Fireflies, Read) | Recording and transcribing meetings | Recording-consent, privilege, and retention issues; transcripts stored off-platform |
| AI coding assistants (Copilot, Cursor) | Writing and reviewing code | Source code and secrets sent to a tool with broad access |
| AI writing / grammar tools (Grammarly, Jasper) | Editing and rewriting copy | Customer and internal text processed by an unvetted third party |
| Research assistants (Perplexity) | Answering questions, gathering sources | Internal questions and context revealing company plans |
The point isn't to ban these tools — most are legitimately useful. It's to decide which ones are approved for which data, and to write that decision down. Our deeper guide to the AI tools employees use without telling IT covers each tool in detail, and the free AI acceptable use policy template includes a ready-made tool tier list you can adapt.
When employees use AI tools outside any policy, the risks cluster into four areas that HR, legal, and security all share:
Data leakage. Confidential customer, employee, or financial information pasted into a free-tier tool can be retained, used for model training, or exposed — outside any contract or oversight.
Compliance exposure. Feeding regulated data (PII, PHI, payment data) into unapproved tools can trigger GDPR, CCPA, or HIPAA obligations, and increasingly AI-specific state laws.
Unreviewed output. AI can produce confident, wrong, or biased content. Without a human-review rule, that output can reach clients, regulators, or production systems unchecked.
Invisible incidents. Because unsanctioned AI use is unlogged, security teams can't see what happened when something goes wrong — which is exactly why IBM found shadow AI adds an average of $670,000 to a breach.
Each of these is addressable with a written policy that sorts tools by tier and names the data that can never be shared. Our free AI acceptable use policy template gives you that structure, and the AI policy generator builds a tailored version in minutes.
Shadow AI has been a concern since ChatGPT launched in late 2022. What's changed in 2026 is scale, sophistication, and consequences.
Scale: The volume of data being entered into AI tools has grown dramatically. Netskope's 2026 research found that prompt volume sent to AI tools inside organizations grew sixfold in a single year — from an average of 3,000 prompts per month to 18,000. The same research found organizations now detect an average of 223 data policy violations per month tied to AI usage.
Sophistication: Employees aren't just using consumer chatbots. They're building workflows, connecting AI tools to company data via integrations, and using AI coding assistants that have access to source code and internal systems. The risk surface is broader than it was.
Consequences: IBM's 2025 Cost of a Data Breach Report found that organizations with high shadow AI involvement in a breach paid an average of $670,000 more per incident than those with low or no shadow AI exposure. The primary driver is detection time: because security teams have no visibility into what happened in unsanctioned AI sessions, containing those breaches takes significantly longer.
The instinct when something feels risky is often to ban it. Shadow AI is no exception. Many companies' initial response has been a blanket "no AI tools unless IT approves them" policy, communicated via all-hands email and enforced by... hoping people comply.
The evidence on this approach is clear, and not encouraging. Research consistently shows that nearly half of employees would continue using personal AI accounts even after an organizational ban. Prohibition drives shadow AI underground rather than eliminating it — and underground shadow AI is worse than visible shadow AI because it means incidents are even less likely to be reported.
The response that works is governance, not prohibition. Clear rules about which tools are approved, what data can and can't be shared with AI tools, and a visible, approved alternative for common use cases. When employees have an approved path, most of them take it.
You don't need enterprise-grade technical tooling to address shadow AI effectively at a 50–300 person company. You need four things:
Before you build any of it, it helps to know where you stand today. Our free Shadow AI exposure scorer runs through ten quick questions and returns a 0–100 exposure score plus the specific governance gaps to close first — a useful baseline before you write a policy.
1. A written policy. Not a suggestion or a memo — a formal AI acceptable use policy that employees can be held accountable to. It should cover which tools are approved, what data can be shared with AI, and what employees are expected to do before using AI-generated output in their work.
2. A tool tier list. A specific list — not a vague category — of which AI tools fall into which permission level. Employees shouldn't have to guess whether the tool they're considering using is approved. They should be able to look it up.
3. Acknowledgment and communication. Employees should formally acknowledge that they've received and read the policy. And that acknowledgment process should involve some actual communication — a team walkthrough or Q&A, not just an email with a PDF attached.
4. A clear escalation path. When an employee encounters a situation the policy doesn't cover — a new AI tool they want to use, an AI output that seems problematic, a colleague doing something that feels off — they need to know exactly who to talk to and what to expect.
Shadow AI Policy generates a tailored acceptable use policy, tool tier list, employee acknowledgment form, and manager FAQ based on your industry, team size, and the AI tools your employees actually use. No legal team required.
Generate my policy kit →Shadow AI is the use of artificial intelligence tools by employees for work without their employer's knowledge or approval. It includes consumer chatbots like ChatGPT, Claude, and Gemini, AI note-takers, AI coding assistants, and AI features inside other apps — used through personal or free-tier accounts that IT and security have no visibility into. It is the AI version of "shadow IT."
The meaning of shadow AI mirrors "shadow IT": technology adopted by employees outside official channels. As generative AI tools became widely available after 2022, employees began using them for work faster than companies could govern them — so "shadow AI" became the shorthand for AI use that happens in the shadows, outside any policy or approval process.
Common examples include a recruiter pasting candidate notes into ChatGPT to draft a summary, a salesperson dropping customer emails into an AI tool to write a proposal, a finance analyst summarizing non-public numbers with a free chatbot, and an engineer using a personal AI coding assistant on company code. In each case the tool helps the employee — but company data leaves controlled systems without oversight.
The tools that show up most often are general chatbots (ChatGPT, Claude, Gemini), AI features in everyday apps (Microsoft Copilot, Google Gemini), AI meeting note-takers (Otter.ai, Fireflies), AI coding assistants (GitHub Copilot, Cursor), AI writing tools (Grammarly, Jasper), and research assistants (Perplexity). Most are genuinely useful; the goal of a policy is to sort them into approved, limited, and prohibited tiers rather than ban them outright.
The honest answer is: more than you think. You can get a rough sense by running an anonymous employee survey asking which AI tools people use for work and through what accounts. Some IT teams can also pull network traffic data that shows connections to AI tool endpoints. But the more useful framing is: assume it's happening at scale and govern accordingly, rather than waiting to quantify it before acting.
Not inherently — the risk depends on what data is being shared. Using a free AI tool to brainstorm marketing taglines with no specific company data is low-risk. Using the same tool to summarize customer complaint emails or analyze employee performance data is high-risk. The policy should make this distinction clear, so employees can self-calibrate rather than either avoiding AI entirely or using it without any thought about data.
The key difference is the data processing agreement. Most paid enterprise tiers of major AI tools — ChatGPT Enterprise, Claude Teams, Microsoft Copilot for Business — include contractual commitments that your data won't be used to train the model and will be handled according to specific security standards. Free tiers typically don't include these commitments, and their terms often allow data to be used for model improvement. This distinction should be reflected directly in your tool tier list.
Both — and increasingly, an HR issue. On the legal side, sharing customer PII with unauthorized AI tools can create GDPR, CCPA, or HIPAA exposure depending on your industry and customer base. On the security side, it expands your breach risk and increases remediation costs when incidents occur. On the HR side, if an employee violates a policy that doesn't yet exist, enforcement is complicated. Having a written policy with employee acknowledgment is the precondition for everything else.