How Can an MSP Find Shadow AI on Client Devices?

Four sources to combine, what each one misses, what to record for every client, and a two-week plan MSPs can run to see which AI tools staff actually use.

An MSP finds shadow AI on client devices by combining four sources: what is installed and running on each device, the AI apps connected to each client's Microsoft 365 or Google Workspace accounts, network or DNS records of the AI services reached, and a short conversation with staff. No single source sees everything, so run them together for two weeks and record what each one finds, client by client.

Shadow AI is any AI tool used for work without the organization knowing about it or approving it. It is shadow IT with one difference: people paste in whatever they need help with, and that is sometimes the most sensitive material they have. Finding it is the first step of governing it, and the first step is mostly an inventory problem.

The Short Answer

  1. Check the devices. List installed AI apps, browser extensions, local model tools and the AI tools that run from a command line.
  2. Check the account connections. List the AI apps that staff signed in to with a work account and that now hold access to mail and files. (In ProctoredAI this connection is in Preview and available on request.)
  3. Check the network. See which AI services the office network reaches, from DNS or firewall logs.
  4. Ask people. A short, non-punitive conversation surfaces what the first three miss.
  5. Record it per client, then decide each app: approve it, approve it with work accounts only, or block it.

Where Shadow AI Hides

Many clients can name one or two AI tools. A real look can find more, in places a browser-only check never sees:

  • Browser chat apps on personal accounts.
  • Desktop AI apps, which a browser control cannot see.
  • Command-line and coding tools that read files and run commands.
  • Browser extensions that read the page in front of the user.
  • AI built into software the client already pays for, such as note-taking, meeting and e-signature apps.
  • AI apps connected to the client's accounts through Microsoft 365 or Google Workspace sign-in.
  • Local models run on the workstation, where nothing leaves the device to be seen.
  • MCP servers that give AI tools access to repositories, databases and the file system.

What Each Source Can and Cannot See

SourceWhat It FindsWhat It Misses
Device inventoryInstalled AI apps, extensions, local runtimes and command-line toolsWeb apps used in a browser, and apps that are installed but never used
Account connectionsAI apps connected to Microsoft 365 or Google Workspace, from any device, phones includedEverything used without that sign-in
Network or DNS recordsAI domains reached from the office networkRemote devices, which product on a shared domain, which account, and any content
Conversations with staffWhat people remember and are willing to sayMost of it, if people fear a penalty

Each source is a useful first look, and what each report includes depends on the client's licences and settings. Check what a report actually shows before you rely on it for a client.

What to Record for Every Finding

  • The app, and which products share its domain.
  • How many people and which devices.
  • Whether work accounts or personal accounts were used.
  • What kinds of sensitive data reached it, if any.
  • Whether it is installed on devices, connected to the client's accounts, or both.
  • The decision: approved, approved with work accounts only, or blocked.

Keep one record per client. A finding with no decision is not finished.

A Two-Week Plan

  1. Day 1: agree the scope. Agree with the client what you will look at, including whether prompts or only app names are recorded, and how staff will be told. Have the client confirm with its own counsel what notice or consent applies. The visibility you gain should be visibility they know about.
  2. Day 1: start the device inventory in observe-only mode. Nothing is blocked while you look.
  3. Day 1: with the client administrator's approval, connect the account sources. List the AI apps that hold access through Microsoft 365 or Google Workspace.
  4. Days 2 to 14: let it run. Two weeks covers month-end work, meetings and the tools people use only occasionally.
  5. Day 14: build the findings. Which apps, how many people, personal or work accounts, and which are installed, connected or both.
  6. Day 15: walk the client through it. Decide each app together. This meeting is where the managed service starts.

What Detection Cannot Tell You

  • A connection is not a leak. A hostname in a DNS log shows that an AI service was reached, not what was sent. Seeing the content takes inspection on the device.
  • A clean scan is not a clean bill. AI built into approved software, devices without your agent, personal devices and phones can all sit outside it.
  • Some apps cannot be inspected. Apps that pin their certificates can be seen, but an on-device proxy cannot read what they send.
  • Lists go stale. New AI tools appear often, so a list of known tools needs a way to flag the ones it does not know yet.

Questions MSPs Ask

Should we block everything we find? Blanket blocks can push AI use onto personal phones and laptops, where nothing is visible. Approving one or two tools with work accounts, blocking the risky ones and stopping sensitive data everywhere usually works better.

Is a DNS filter enough? It is a good first look at which AI domains the office network reaches. It does not see remote devices, desktop apps, the account used or any content.

How often should we repeat it? Run a full look when you take on a client, then review monthly. New tools show up constantly.

Can we do this with scripts instead of an agent? For a snapshot, yes: a script from the RMM can list installed apps and extensions, and it can be scheduled to repeat. A script does not see what is sent to AI services, report AI-looking services it does not know, or bring every client into one console.

Where to Go Next

Our shadow AI guide covers each place AI hides in more detail, and ways to control AI use compares what a policy, an extension, a network filter, account connections and an endpoint agent each see and miss. We built ProctoredAI so an MSP can run the device side of this look from the RMM in observe-only mode, with Microsoft 365 and Google Workspace account connections (Preview, available on request), and review the results with each client.

Governance starts with visibility. You can't govern what you can't see.

This article is general information, not legal, compliance or security advice, and does not create a professional relationship. Product details are accurate as of the publish date; see ProctoredAI.com/terms for what the service does.

See Every AI Tool on a Real Client in Two Weeks

Deploy from your RMM in observe-only mode, then walk the client through what their staff actually use.

Running your own IT? See ProctoredAI for your business