Finding Shadow AI Without a Witch Hunt

What is shadow AI, and why does it appear?

People adopt AI tools because they are useful and easy to reach. A marketing analyst tries a writing assistant, an engineer adds an AI coding extension, a vendor switches on a new generative feature in a product you already pay for. None of this is malicious. The risk is that sensitive data leaves your control, outputs are used without review, and your inventory misses the exposure.

Where can you look for signals?

No single source finds everything, so combine several:

• Network and identity logs. Traffic to known AI services, single sign-on grants and OAuth consents.
• Procurement and expense data. Subscriptions, card spend and vendor onboarding records.
• Software inventories. Browser extensions, desktop applications and plugins.
• Cloud and code. AI service usage in cloud billing, API keys and AI libraries in repositories.
• Vendor release notes. New AI features in existing SaaS products, which often arrive switched on.
• People. Short declaration surveys and conversations with team leads.

Why lead with amnesty and a fast lane?

Technical detection tells you that something is in use. Only people can tell you why, with which data, and whether the output is relied on. A time-limited amnesty, where declared use is assessed rather than punished, makes that conversation possible. Pair it with an intake route that returns a decision quickly. If approval takes months, people will return to workarounds.

How do you triage what you find?

Assess each finding on a small number of questions:

• What data goes in: public, internal, confidential, personal or regulated?
• Where does it go, and does the provider retain or train on it?
• Does the output feed a decision about people, money or safety?
• Who is the business owner, and how many people use it?
• Is an approved alternative available?

The answers place it in a tier. Low-exposure tools may be approved with usage guidance. Higher tiers go through fuller assessment. Some uses are stopped, with a reason and a supported alternative. Everything lands in the AI inventory.

A hypothetical walk-through

A health insurer runs a six-week discovery. This example is hypothetical.

1. Single sign-on logs and expense data show several AI subscriptions that procurement never reviewed.
2. A declaration survey, with amnesty, adds uses that logs could not see, including an embedded summarization feature inside a claims platform.
3. Triage finds one tool receiving member information; it moves to an approved enterprise plan with data controls, and the original use is stopped.
4. The rest are approved with guidance or scheduled for assessment, and all are recorded.
5. Discovery signals are wired to alert the governance team when a new AI service appears.

How do you keep it from recurring?

Make the approved path easier than the unapproved one. Publish a short list of sanctioned tools, say what data each may handle, and give teams a place to ask. Add AI checks to procurement and vendor onboarding, so new AI features are asked about before contracts renew. Agents and retrieval systems raise the stakes because they act on data and systems; see governing agents and RAG.

What does good communication look like?

Say plainly why discovery is happening: to protect data and to help people use AI safely, not to count violations. Publish what will and will not lead to action. Give examples of acceptable use, such as drafting from public material, and examples that need approval, such as entering customer or patient data. Short, specific guidance beats a long policy that nobody reads. Share what you learn, too; if many teams want the same capability, that is a signal to provide an approved tool.

Track a few indicators over time, such as the number of declared tools, the time from request to decision, and the share of tools with an assigned owner. They show whether discovery is turning into governance or just into a longer list.

What should you avoid?

• Naming and shaming individuals for early adoption.
• Relying on one signal, such as network logs, and assuming the picture is complete.
• Collecting more employee data than discovery needs; apply the same privacy discipline you expect elsewhere.

Revisit the approved list every quarter. Tools that nobody uses can be removed, and tools that people keep asking for can be assessed and added. Over time the declared list should come closer to the real one, and the gap between them is a useful measure of how well governance is working.

Where aerais fits

aerais runs AI inventory and shadow-AI discovery as part of our AI governance practice: connecting detection signals, designing the intake and triage process, and operating it within the method we use to build and run governance. The aim is visibility that stays current, without turning governance into surveillance.

This article is for general information only and is not legal advice.