The obvious response to shadow AI is to block it. Restrict access to ChatGPT, Claude, and other AI platforms. Problem solved. Except it doesn't work.
Samsung banned generative AI tools after discovering sensitive code had been uploaded to ChatGPT. Several banks and enterprises followed with similar restrictions. The result? According to industry surveys, even in organisations with explicit AI bans, around half still have unauthorised AI usage happening.
Blocking AI is a losing battle. Here's what works instead.
Why bans fail
There are practical reasons and human reasons.
The practical problems
AI tools are everywhere. They're in browsers, on phones, embedded in software you've already approved. You can block chatgpt.com on your corporate network, but employees can use mobile data. You can restrict desktop applications, but browser-based tools don't require installation.
New AI tools launch constantly. By the time you've added one to your blocklist, three more have appeared.
The human problems
People use AI because it makes them more productive. A Gartner study found that employees who use AI tools report significant time savings on routine tasks. When you ban something useful, people find workarounds. That's not defiance. It's problem-solving.
Bans also signal distrust. Your team starts to feel like they're being treated as a risk rather than an asset. That creates resentment and, ironically, less willingness to flag potential issues when they arise.
The visibility-first alternative
Instead of trying to prevent AI usage, focus on understanding it.
When you have visibility into how your team uses AI, you can identify which tools are actually being used and see what types of data are being shared. You can spot high-risk behaviours before they become incidents. And you can make policy decisions based on evidence rather than assumptions.
This is governance, not surveillance. The goal isn't to catch people doing something wrong. It's to understand what's happening so you can manage risk appropriately.
How visibility works in practice
Browser-based monitoring tools show you which AI platforms your team accesses. Not just ChatGPT, but Claude, Perplexity, Gemini, and the dozens of other tools that have launched in the past year. They detect when sensitive data patterns appear in prompts, things like credit card numbers, personal identifiers, code snippets, or specific keywords you define.
You also get usage trends over time. Which teams use AI most? Has usage increased after you rolled out new policies? Are people moving to different tools? And you can see file uploads, which often represent higher-risk actions than typing a question.
Setting policies that work
Visibility gives you the information you need to create realistic policies.
Risk-based categories
Not all AI usage carries the same risk. A marketing coordinator using AI to brainstorm headline options is different from a lawyer uploading client contracts.
Think in tiers. At the low end, you have general research, brainstorming, and writing assistance where no sensitive data is involved. In the middle, there's AI usage with company information that isn't confidential, like publicly available content or general business processes. At the high end sits anything involving customer data, financial information, proprietary code, legal documents, or strategic materials.
Your policies can then be proportionate. Low-risk usage might require no oversight. Medium-risk might require approved tools only. High-risk might require specific protocols or approval.
Clear data guidelines
Most shadow AI problems come from data exposure. Help your team understand what counts as sensitive.
Be specific. "Confidential information" is vague. Instead, list categories: customer names, email addresses, phone numbers; financial data including revenue, costs, pricing; employee personal information; proprietary code or technical documentation; legal correspondence and contracts; strategic plans and unreleased product information.
When people know exactly what not to share, they're much more likely to comply.
Approved alternatives
If you want people to stop using unauthorised tools, give them something better.
Enterprise versions of AI platforms come with data protection guarantees. Microsoft Copilot for Business, ChatGPT Enterprise, and Claude for Business all include provisions that your data won't be used for model training.
Internal AI assistants can be deployed on your own infrastructure, keeping data entirely within your control.
Approved workflows can include AI at specific steps. For example: "You can use AI to draft customer responses, but only through [approved tool], and always review before sending."
The key is making the approved option as easy as the shadow option. If your sanctioned AI tool requires a five-step login process while ChatGPT is one click away, people will keep using ChatGPT.
The intervention approach
Even with good policies, there will be moments when someone is about to do something risky. How you handle those moments matters.
Real-time warnings
The most effective intervention happens at the point of risk.
When someone is about to paste sensitive data into an AI prompt, showing them a warning at that moment is far more effective than sending them an email three days later.
Good intervention systems alert users when they're about to share potentially sensitive information. They give options: cancel, redact the sensitive parts, or proceed with acknowledgment. The decision gets logged for audit purposes. And unless the risk is severe, they don't block workflows entirely.
This approach treats employees as adults who can make informed decisions when given the right information.
Education over enforcement
When someone triggers a warning or violates a policy, the response shouldn't be punitive. It should be educational.
Most shadow AI happens because people don't understand the risks. They're not trying to harm the company. They're trying to do their jobs.
Explaining why a particular action is risky, and showing them how to accomplish the same goal safely, is more effective than issuing warnings or restricting access.
Building an AI-positive culture
The companies managing shadow AI best are the ones that have embraced AI, not fought against it.
This means acknowledging that AI is useful. Your team isn't wrong to want these tools. It means investing in approved solutions, because if you expect people to stop using free tools, you need to provide alternatives that are at least as good. It means creating clear guidance so people can do the right thing easily. And it means measuring and iterating, using the visibility you've built to understand what's working and what isn't.
The bottom line
Shadow AI isn't a technology problem you can solve with blocking rules. It's a governance challenge that requires visibility, clear policies, and respect for why people turn to these tools in the first place.
Banning AI pushes the behaviour underground. Visibility brings it into the open where you can actually manage it.
The goal isn't to stop your team from using AI. It's to make sure they use it safely.