Shadow AI: How to Stop Data Leaks Without Banning AI
Shadow AI is the use of AI tools your company has not approved, usually a public chatbot on a personal account, with company data. It is a risk because whatever an employee pastes in leaves your control: you do not know where it is stored, for how long, or who can read it, and you have no record that it happened. Banning the tools rarely fixes it. What works is giving people an approved tool that is easier to use than the unapproved one, then backing it with a short policy and some detection.
That order matters. Detection and policy without a good alternative just push the behavior onto phones and home laptops, where you see even less of it. This post covers how big the problem usually is, why bans fail, and a rollout plan that replaces shadow AI with something you control.
How common shadow AI is
The best recent numbers come from employee surveys, so treat them as indicative rather than exact. In a January 2025 survey of 1,000 US employees at companies with at least 5,000 staff, TELUS Digital found that among employees who use generative AI at work:
- 68% reach public assistants such as ChatGPT, Copilot or Gemini through personal accounts.
- 57% admit entering sensitive information into them. The most common types were personal data (31%), unreleased product or project details (29%) and customer information (21%).
- 29% say their company has a policy that forbids exactly this, and they do it anyway.
- Only 24% say their company requires AI training, and 44% say their company has no AI guidelines or they do not know whether it does.
Two more figures from the same survey explain why bans fail. 84% want to keep using AI assistants at work, and 22% of employees who already have a company-provided assistant still use personal accounts. People are not using shadow AI to be reckless. They use it because it saves them time, and the approved option, if there is one, is slower, weaker or harder to reach.
Why shadow AI is a risk
The risk is not that a chatbot is evil. It is that data moves somewhere you cannot see or govern. The UK National Cyber Security Centre made the same point in its guidance on ChatGPT and large language models: do not include sensitive information in queries to public LLMs, because stored queries can be leaked or hacked, and queries made under one login can be combined into a fuller picture than any single one reveals.
In practice the exposure falls into four buckets:
| Risk | What it looks like |
|---|---|
| Data leaving your control | Source code, contracts or customer records stored under a provider's consumer terms, on an account the company does not own |
| No audit trail | You cannot answer "who sent what, where, and when" if a customer or regulator asks |
| Contract and regulatory exposure | Customer data sent to a processor your contracts or data protection rules do not allow |
| Unreviewed output | Generated code, figures or legal text reaching production or customers with nobody checking it |
The best-known example is Samsung. In 2023, after engineers pasted internal source code into ChatGPT, the company banned generative AI tools on company devices and networks and started building internal tools instead. The ban was the emergency response. The internal tools were the fix.
Why banning AI does not stop shadow AI
A ban removes the tool from the places you can see: company laptops and the company network. It does nothing about a personal phone on mobile data, which is where the use goes next. You end up with the same data leaving the company, less visibility, and staff who have learned that the policy and the work are in conflict.
Blocking also has a maintenance cost. New AI tools and AI features inside existing SaaS products appear every month, so a blocklist is never complete. Keep blocking for the highest-risk cases, but do not mistake it for the plan.
The fix: make the sanctioned path the easy path
People choose the unapproved tool for three reasons: it is good, it is fast to reach, and it does not ask permission. A sanctioned alternative has to match all three.
- Quality close enough that nobody feels punished. Current open-weight models in the 14B to 70B range handle drafting, summarizing, extraction and most coding help well. Our guide to choosing an open-source LLM for enterprise use covers which families to test and how to read their licenses.
- One bookmark, signed in through SSO, available on day one. If getting access takes a ticket and a week, people stay on their personal accounts.
- Clear rules about what data is allowed, so people do not have to guess.
There are two ways to provide it. If your data rules allow a hosted provider under an enterprise contract, that is a legitimate sanctioned option and often the fastest. If your most sensitive data, the data people are actually pasting, must not leave your network, the alternative has to run inside it. The trade-offs between the two are in private LLM vs OpenAI.
A self-hosted setup has three parts: a model server inside your network, an OpenAI-compatible endpoint in front of it, and a chat interface. Open-source chat UIs such as Open WebUI and LibreChat can be pointed at any OpenAI-compatible endpoint and support SSO through your identity provider; check each project's current documentation for the exact configuration. The self-hosted ChatGPT alternative guide walks through the pairing, and the self-hosted Copilot alternative does the same for developers, who are often the heaviest shadow AI users because source code is what they paste.
How to detect shadow AI
Detection tells you how big the problem is and whether the sanctioned tool is working. You do not need a dedicated product to start:
- Web proxy, secure web gateway or DNS logs. Count requests to the main AI domains per week and per department. The absolute number matters less than the trend after launch.
- Expense reports and corporate cards. Personal AI subscriptions claimed as expenses are a direct signal of unmet demand.
- Browser extension inventory on managed devices. AI writing and meeting extensions often read whole pages.
- OAuth grants in your identity provider and SaaS admin consoles, for AI apps that employees connected to email, calendar or drive.
- Ask. A short anonymous survey on which tools people use and what for gives you the requirements for the sanctioned tool.
Use what you find to decide what to build, not to discipline people. If staff think the survey is a trap, you lose your best source of information.
A 60-day rollout plan
Weeks 1-2: measure and decide. Pull the proxy and expense data above, run the survey, and pick the three most common use cases (for most companies: drafting and rewriting, summarizing documents, and code help). Decide which data classes the sanctioned tool may receive.
Weeks 3-4: stand up the sanctioned tool. Deploy a model sized for those use cases, put a chat UI with SSO in front of it, and give developers an editor integration on the same endpoint. Start with one general model; add more once you see real usage.
Weeks 5-6: pilot and publish the policy. Launch with two or three departments that showed the most shadow AI use. Publish a one-page acceptable-use policy that names the approved tools, links the data rules, and requires human review of output before it reaches customers or production. The LLM governance guide has a data-class table and policy outline you can adapt.
Weeks 7-8: expand, then restrict. Open access to everyone and announce it widely. Only after the sanctioned tool is live and adopted, block the highest-risk unapproved tools on company networks. Restricting before you have an alternative is the step that drives use onto personal devices.
After that, compare monthly: traffic to public AI domains, active users of the sanctioned tool, and the top requests people still take elsewhere. A falling first number and a rising second one is the result you want. A request type that keeps going to public tools tells you what to add next, usually a larger model, a vision model for screenshots and scanned documents, or an embedding model for search over internal documents.
Where AI Refinery fits
AI Refinery, a self-hosted LLM platform, covers the model side of the sanctioned alternative inside your own infrastructure: your cloud, your VPC, on premises or an air-gapped network, with nothing phoning home. You pick a model from a curated catalog of text, embedding and vision-language models, choose a quantization level and launch an inference server with one click. Each deployed model exposes an OpenAI-compatible /v1/chat/completions endpoint, so the chat UI and editor plugins above connect to it by changing a base URL. RBAC, SSO, approval workflows and a full audit log give you the "who used what, and when" record that shadow AI never leaves.
It does not include a chat interface; you pair it with an open-source UI such as Open WebUI or LibreChat. It does not detect shadow AI on your network or write your policy. And it is not the right fit if a hosted provider's enterprise plan already satisfies your data rules, or if you have no GPUs and no one to run them, since you provide the hardware. If the data people paste into public chatbots is data that should never leave your network, a private LLM platform you run yourself is the alternative that removes the reason to paste it elsewhere.
This post is general information, not legal advice. Check your data protection and contractual obligations with your own counsel.