Waxell

Shadow AI: how to see which AI tools your staff use

You see shadow AI by checking the network, the browser, devices and app access together. Here is what each method shows, what it misses and where to start.

Logan Kelly12 min read

You see shadow AI by looking in four places together: your network traffic, the browser, the devices themselves, and the apps your staff have connected to company accounts. No single one shows the whole picture, because AI now arrives as websites, desktop apps, models that run on the laptop itself, phone apps and features built into tools you already allow.

Shadow AI is AI that people use for work without IT knowing or approving it. It is not a fringe habit. In Microsoft and LinkedIn's 2024 Work Trend Index, a survey of 31,000 people in 31 countries, 75% of knowledge workers said they use AI at work, and 78% of those AI users said they bring their own AI tools to work.

This guide is for IT and security leads, and for MSPs doing this across many clients. It covers what each way of looking can and cannot see, the questions IT teams and MSPs ask most, and a short plan for your first week.

What counts as shadow AI?

Shadow AI is any AI tool used for work without approval: a chat assistant on a personal account, a browser extension that summarizes pages, a coding agent a developer installed, a model running locally on a laptop, or an AI app someone connected to the company's email or files.

It is a kind of shadow IT, with one difference that matters. A shadow file-sharing app stores your data. A shadow AI tool reads it, and some tools can act on it. Microsoft's own guidance draws the same line: it lists data sprawl as a main risk of shadow IT, and data leaking to AI models and AI tools acting without oversight as risks of shadow AI.

Why is shadow AI hard to see?

Because it does not arrive in one way. Most organizations are dealing with five forms at once:

  1. AI websites in the browser, such as ChatGPT, Claude, Gemini and Perplexity, often on personal accounts.
  2. Desktop and command-line apps, including coding agents and the chat assistants' own desktop apps.
  3. Local models that run entirely on the laptop, through tools such as Ollama and LM Studio.
  4. Phone apps, on devices you may not manage at all.
  5. AI inside other software, from chat apps to office suites, where AI is a feature rather than a separate product.

Each method below sees some of these forms and misses others. That is why the answer is a combination, not one tool.

Which methods show you shadow AI, and what does each miss?

MethodWhat it showsWhat it misses
Network and DNS logs, secure web gatewayWhich AI services each user or device reaches, how often, and how much data movesWhat was typed, unless traffic is inspected; local models; devices whose traffic does not pass through it
Browser extensionWhich AI sites are used, and often which account is signed inDesktop apps, local models, browsers without the extension
Device inventory (MDM or an endpoint agent)Which AI apps and tools are installed on each deviceWebsites with nothing to install; personal devices you do not manage
App access records (OAuth)Which AI apps users have connected to company email, files and calendarsAI tools that never ask for access to company accounts
Vendor account restrictionsWhich accounts can sign in to a given AI service on your networkEvery AI service that does not offer one; anything that does not pass through your proxy

The rest of this guide takes each question in turn.

Can network logs show which AI tools staff use?

Yes, at the level of which service, which user and how much. A secure web gateway or DNS filter that recognizes AI services gives you a list of the AI sites and APIs your users reach. Microsoft's shadow AI discovery in Entra Global Secure Access is one example: it analyzes network traffic, matches it against a catalog of AI apps, and shows the users who access each app, how often, and the bytes sent and received.

What network logs do not show by default is the conversation itself. Microsoft separates the two: logging the actual prompt content requires TLS inspection, which decrypts the traffic on its way out. Network logs also miss anything that never crosses your network, such as a model running on the laptop or a phone on cellular data.

How do you see AI used in the browser, not just apps?

The browser is where you get the most detail. You have two options.

  • A browser extension that your device management installs on company browsers. It can see which AI sites a person visits and, for many services, which account they are signed in with.
  • Inspection at the network edge, where a gateway decrypts and reads web traffic to AI sites. This sees more of the content, but it means managing certificates on every device and deciding what you are comfortable reading.

Both depend on the browser or device being yours to manage. Neither sees a personal laptop at home.

Can you tell personal ChatGPT or Claude accounts from company ones?

Yes, in two ways: by seeing which account is signed in, or by only letting company accounts sign in at all.

To see it, a browser extension can read whether the signed-in account belongs to your company's domain or a personal one.

To enforce it, the larger AI vendors now let your network allow only your company's accounts:

  • ChatGPT Enterprise supports Workspace Blocking. Your proxy or secure access service adds a header listing your approved workspaces, and ChatGPT filters out every other workspace, including personal ones.
  • Claude Enterprise plans and Console organizations support Tenant Restrictions. Your proxy adds a header listing your approved organizations, and Anthropic blocks access from any organization not on the list.
  • Google Workspace lets admins block consumer accounts, such as personal Gmail, from signing in to Google services on the corporate network.

The ChatGPT and Claude controls block rather than report, and they need TLS inspection so your proxy can add the header. They also leave gaps you should know about. OpenAI notes that ChatGPT can be used without signing in at all, which you block separately. Its desktop and mobile apps use certificate pinning, so on those apps the header only works on managed devices where you have deployed an exception through device management.

Will you catch someone running a local model like Ollama?

Not reliably from the network. A local model runs on the laptop itself. Ollama, for example, listens on the computer itself on port 11434 by default, so prompts go to a program on the same machine rather than to an AI provider your gateway would see. Your network may see Ollama download a model, since it pulls models from the internet, but not what anyone types into it.

The way to find local models is on the device:

  • Device inventory. Your device management can list installed apps on company-owned computers. Look for Ollama, LM Studio and similar tools.
  • An endpoint agent that knows what AI tools look like, including command-line tools that a general app list may not name as AI.

What about AI on phones?

This is the hardest gap. On company-owned phones, device management can list installed apps, though what it sees depends on the platform and how the phone is enrolled. On a personal phone enrolled for work, it sees far less. Microsoft's Intune documentation is explicit: for personal devices, Intune never collects information on apps it does not manage.

So on personal phones your options are narrower:

  • Keep company data in managed apps, so AI apps on the personal side cannot reach it.
  • Use network controls for traffic that passes through your network.
  • Set a clear policy on which AI apps may be used for work, and say why.

Vendor controls reach phones only in some cases. OpenAI says its header-based Workspace Blocking can be enforced in its mobile apps on managed devices where the certificate exception has been deployed.

What about AI built into apps you can't block?

More AI now ships as a feature of software you already allow. Meta, for example, put its Meta AI assistant into chats on WhatsApp, Messenger and Instagram, so people can use it without leaving the app.

A list of AI websites does not catch this, because the AI is part of an app you have approved, not a separate site. Your options are the app's own admin settings where they exist, a policy that covers AI features inside approved apps, and keeping sensitive data out of apps that are not meant to hold it.

What about AI apps connected to company data?

Some of the most important shadow AI never shows up as a website visit. It is an AI app a user has connected to company email, files or calendar with a "Sign in with Google" or "Sign in with Microsoft" button.

Your identity provider keeps a record of these. In Google Workspace, the Admin console lists the apps that have accessed Google data, and details typically appear 24 to 48 hours after a user authorizes an app. Review this list regularly, and decide which AI apps may hold access to company data at all.

What about AI agents and MCP servers on developer laptops?

Developers often connect coding agents to tools and data through MCP (Model Context Protocol) servers. Some of these are hosted services, which network discovery can see. Others are configured locally inside tools such as Claude Code and Cursor, where only a check on the device will find them. The guide to MCP governance covers how to bring these under control.

What to do once you can see it

Seeing shadow AI is the first step, not the goal. Once you have a list, decide for each tool:

  1. Allow it and bring it under company accounts, if it is useful and the vendor meets your bar.
  2. Warn, if you want people to stop and think before they share data.
  3. Block it, if it is too risky or you already have an approved alternative.

Write the rules down in a policy people can follow; the guide to AI policy enforcement covers how. Then deal with the data side, since that is where most of the risk sits: see keeping sensitive data out of AI. If you are just starting, where to begin with AI governance puts this in context.

Your first week: a shadow AI checklist

  • Pull a report of AI services reached from your network or secure web gateway for the last 30 days.
  • Export installed apps from device management on company devices and search for AI tools, including Ollama and LM Studio.
  • Review which AI apps users have connected to company email, files and calendars.
  • Check whether your main AI vendors offer account restrictions, and whether your proxy can add the header.
  • List what you cannot see yet: personal devices, phones, AI inside approved apps.
  • Decide allow, warn or block for the top ten tools you found, and tell staff the outcome.

What shadow AI is not

Shadow AI is not the same as unauthorized software in general, and seeing it is not the same as stopping it. A report of AI usage tells you where to act. It does not keep data out of a model on its own. Treat visibility as the start of a policy, not a replacement for one.

How Waxell handles this

Waxell Endpoints finds the AI already running on employee laptops. You deploy it to enrolled Mac and Windows devices through the device management tool you already use. It checks each device against a catalog of AI apps, command-line tools and code-editor extensions, including local model tools such as Ollama, and it reads which MCP servers tools like Claude Code and Cursor are connected to. By default it records which app talked to which AI provider, which user, and when.

Its browser extension for Chrome, Edge and Firefox shows which AI sites are in use and whether the signed-in account is on your company's domain or a personal one. From there you can set a policy per tool, up to blocking it on the device. The extension also captures AI chats on the sites it covers by default, with secrets redacted on the device before upload; an admin can turn upload off to keep captures on the device.

Endpoints covers enrolled computers. It has no phone agent, and machines that are not enrolled do not show up. See how Endpoints works.

All posts on this topic

FAQ

What is shadow AI?

Shadow AI is any AI tool used for work without IT's knowledge or approval. It includes AI websites on personal accounts, desktop and coding apps, models running locally on a laptop, phone apps, and AI apps connected to company data. It is a form of shadow IT, with the added risk that the tool reads your data and may act on it.

What is the best way to detect shadow AI?

Combine methods, because each one misses something. Network logs show which AI services people reach, a browser extension shows which sites and accounts are used, device inventory shows installed apps and local models, and app access records show AI apps connected to company accounts. Start with the network report and the device inventory, since most companies already have both.

Can I block personal ChatGPT accounts but allow ChatGPT Enterprise?

Yes, if your network can inspect traffic and add a header. ChatGPT Enterprise's Workspace Blocking lets you list your approved workspaces, and ChatGPT filters out the rest, including personal ones. You also need to block signed-out use separately and handle the desktop and mobile apps, which need a certificate exception on managed devices.

Can I see what employees type into AI tools?

Only with tools that read the content, such as a gateway that decrypts traffic or an agent on the device that captures it. Most discovery tools show which service was used and by whom, not the conversation. Decide what you need to see before you turn on content capture, and tell staff what is recorded.

How do MSPs find shadow AI across clients?

Run the same checks in every client tenant: the network report, the device inventory and the app access review. Then report the findings per client in a consistent format. The hard part is doing it repeatedly across many tenants, which is where a tool that rolls results up per client saves time.

Sources

Want to see which AI is already running on your company's laptops? Book a 30-minute demo.