Shadow AI discovery4 min readField note 10

How to build a shadow AI inventory from managed Windows endpoints

Reconcile devices, accounts and observed AI tools into a working inventory. Follow one fictional Windows fleet from the first activity chart to an actionable coverage gap.

Activity chart beside the list of observed AI destinations in the Northstar dashboard.
The idea in focus

In the fictional Northstar workspace, destination counts give the analyst a starting list of AI tools to reconcile.

Field note / 10
In brief

Pair every AI destination with a device, an observation and an owner.

  • Define which endpoints were in scope.
  • Keep browser and desktop observations distinguishable.
  • List unresolved devices alongside the observed tools.
In this article

Choose the devices your inventory covers#

An IT director asks which AI tools Finance uses before renewing an enterprise license. Start with a dated list of the Windows devices Finance owns. Compare it with enrolled agents, recent check-ins and the browsers connected during the review window. Record a reason for each missing device. A laptop offline for the entire week contributes a coverage gap to the report.

Keep your approved-tool list beside the observed destinations. Label each match approved, under review or owner unknown according to your own policy. The dashboard can supply an observation. The decision about authorization comes from your organization.

Northstar fleet counts above the table of managed Windows agents.
Use the fleet as the denominator for your observation period.

Group observations by account and hostname#

Open Individual AI usage in the demo and search Alex Morgan. A single hostname has observations for Claude, Google Gemini, ChatGPT and Kimi. Keep one inventory row per destination and device, alongside the account and last observed time. If the laptop is shared, check who held the session at that time before attaching a person’s name to a finding.

Alex Morgan’s filtered timeline shows four AI destinations on a managed endpoint.
A single person’s timeline can contribute several rows to the tool inventory.

Give desktop assistants their own evidence label#

Now search Cursor. The Windows observation identifies an editor running on Taylor Chen’s machine. Enter “process observed” in the inventory and leave the document field empty. This matters when deciding whether to buy an approved coding assistant. Knowing a tool runs on three laptops can justify talking to Engineering. Those three launches cannot quantify how much source code the team sent.

Cursor search returns a Windows process observation for Taylor Chen.
The Cursor row identifies a running editor. The document field has no observed file.

Follow a filename through to its receipt#

In Confidential documents, search for Q4_client. The matching spreadsheet has a ChatGPT receipt with a related HTTP response. Write down the filename, destination, time and exact action. A successful response could belong to another request on the page. If the decision depends on whether the provider accepted the attachment, record that as an unanswered question and identify the evidence needed to resolve it.

A search for Q4_client returns a ChatGPT file observation and related HTTP response.
Keep the response and its relationship to the selected file in the inventory record.

Prioritize destinations with sensitive paste findings#

Search API key in Sensitive pastes. The Northstar example returns observations for two destinations. Give the credential owner the device, occurrence time and detector category so they can investigate through their established process. Keep secret values out of the inventory spreadsheet. A detector match still needs review before anyone labels a credential active or exposed.

API key search results include paste observations for Claude and OpenAI Codex.
A finding category gives the reviewer a way to triage without putting the pasted value in the inventory.

Close the inventory with an owner and a coverage check#

A useful row reads “Finance pilot / NS-WIN-014 / ChatGPT / file-related response / Finance IT to review approval.” Give it a review date and link its underlying observation. For a quiet device, open agent detail and check the browser connection against the time window. Assign someone to restore missing coverage and repeat the same harmless test after the fix. Deliver the dated inventory and the list of unresolved devices together.

Agent detail for a Northstar Windows device with health, browser and capability information.
Review the device’s readiness before interpreting a quiet period.
Read the primary sources

Sources & further reading

  1. Employee AI visibility on managed endpoints
  2. Check browser and desktop coverage by tool
  3. Inspect the Northstar fleet
← Back to all articlesBack to top ↑
A CLEAR PLACE TO START

Plan your AI visibility pilot.

Create a brief for your team. Your answers stay in this browser and are included in your download.

AI tools in scope
More supported tools

This creates a planning document. It does not enroll devices or submit a request.