Shadow AI discovery for MSPs: what the browser shows that DNS and Microsoft 365 logs miss

Good shadow AI discovery tells an MSP which AI tools a client's end users rely on, who uses them, with which accounts, what data goes in, and how much time they take up. DNS logs show which AI domains devices reach, and Microsoft 365 logs show which apps end users signed in to with work accounts. The browser shows the rest, including personal accounts, what end users paste and upload, and how much time they spend in each tool. For most MSPs, two weeks of browser-level discovery is the fastest route to an approved AI list a client will actually agree to.

What does a client actually need to know about AI use?

A list of AI domains is a start, but it doesn't help a client decide anything. The decisions behind an approved AI list need answers to five questions:

1. Which AI tools are in use, and how widely? One end user trying a tool once is different from half the office using it daily.
2. Who uses them? Usage clusters by team. Marketing, finance, and operations often reach for different tools.
3. Which accounts? A business account on an approved tool is governed. A personal account on the same tool isn't, even though the domain is identical.
4. What goes in? Pastes, uploads, and prompts that carry client data matter far more than a visit to a homepage.
5. How much time is spent? Time spent shows which tools end users depend on, and which ones can be retired without a fight.

The answers become the client's approved list, the warn list, and the block list. Without them, the MSP is guessing, and the client signs off on a policy that doesn't match how its end users work.

What does each layer show?

Each layer answers some of the five questions well. Most MSPs already have DNS filtering and Microsoft 365 logs, so the practical question is what the browser adds on top.

The layers complement each other. DNS covers traffic from apps outside the browser, and Microsoft 365 is the best place to find AI apps that end users granted access to mailboxes and files. With licensing beyond Business Premium, Microsoft's cloud app discovery tools catalog more AI apps as well. The browser fills in the three questions the client cares about most, namely accounts, data, and time.

What does browser-level discovery add?

It turns a list of domains into a picture of how end users actually work with AI. Four findings matter most to small clients.

Personal accounts on approved tools. The client pays for a business AI plan, yet a share of end users still sign in with personal accounts on the same site. To DNS, both look the same. The browser can tell which account an end user enters at sign-in, and that difference decides whether the data stays under the client's control.

Uploads and pastes, not just visits. A visit to an AI site is a weak signal. A spreadsheet uploaded to one is a strong one. Browser-level discovery shows which tools receive files and pasted text, so the MSP can rank risk by what leaves rather than by how often a domain appears.

Sensitive data in prompts. Detection on major AI chat apps can flag prompts that contain data types such as card numbers, health identifiers, and phone numbers. That gives the client a concrete reason to approve a business plan or tighten a rule, without the prompts being stored.

Time spent. Active time per app separates the tools end users depend on from the ones they tried once. That matters at decision time. Blocking a tool someone spends an hour a day in needs an approved alternative ready on day one. Blocking a tool nobody has opened in a month needs only a notice.

Together, these findings let the MSP walk into the client meeting with a recommendation instead of a list.

How should an MSP run discovery for a new client?

Two weeks of watching, one meeting, and a short list of decisions. A plan that works for most 10- to 100-seat clients:

1. Days 1 to 2, turn on visibility. Deploy browser monitoring in report-only mode, without blocking anything new. Confirm DNS filtering is active, and pull the list of apps with consent to Microsoft 365 data.
2. Days 3 to 10, let normal work happen. A full working week captures the routine tools. Month-end or a busy season captures more, so note what the period didn't cover.
3. Days 11 to 12, build the picture. Rank AI tools by active users and time spent. Flag personal-account sign-ins, uploads to unapproved tools, and prompts with sensitive data types. Check the consented apps list for AI tools with mailbox or file access.
4. Day 13, meet the client. Bring one page. Show the top tools, the personal-account share, and the riskiest data flows, then recommend which tools to approve, which to warn on, and which to block.
5. Day 14, set the policy. Approve the agreed tools on business plans, set unreviewed AI sites to warn, block pastes and uploads to unapproved AI, and warn on personal logins to approved tools.

Discovery doesn't end there. Keep the reporting running and review it monthly, since new AI tools and features appear faster than any approved list can be written. For turning the findings into an enforceable policy, see how to make a client's AI policy enforceable.

How does this work in Atakama?

In Atakama, the Discovery policy is the usual starting point for a new tenant. Web app usage appears under Insights, with top apps and usage over time. Prompts that contain sensitive data types appear on the Sensitive AI prompts tab of the Data Loss Prevention Insights page, without the prompts being stored. SaaS login controls can warn on or block personal-account logins to approved tools, and each warn, block, and proceed event is logged under Credentials Analysis, so personal-account use shows up as a count rather than a guess. Atakama also includes DNS filtering, so the DNS view sits in the same console as the browser view.

For the client meeting, the Client Health Report produces a shareable PDF with an AI activity page covering approved and unapproved AI usage, top AI apps and users, and block and warn counts, along with recommendations. It can be sent to the client automatically on the first of each month with the prior month's data. For MSPs, that means discovery runs the same way in every tenant, and the monthly review is ready before anyone builds a slide.

FAQ

What is shadow AI?

AI tools and features that end users adopt for work without IT approval or oversight. That includes free AI chat tools, personal accounts on approved tools, browser extensions with AI features, and desktop AI apps.

How do MSPs discover shadow AI at a client?

Combine the layers they already have. DNS filtering shows which AI domains devices reach, Microsoft 365 shows apps that end users signed in to or granted access to company data, and browser-level monitoring shows accounts, pastes, uploads, and time spent. Two weeks of data is usually enough for a first approved list.

Can DNS filtering detect shadow AI?

It detects which AI domains devices reach, from browsers and other apps alike. It can't tell a personal account from a business one on the same domain, and it doesn't show what end users send. Use it as the first layer, not the only one.

Should MSPs block shadow AI right away?

Not on day one. Watch first, agree on an approved list with the client, then block pastes and uploads to unapproved AI and warn on everything nobody has reviewed. Blocking before an approved tool is ready pushes end users toward personal devices.

How often should shadow AI discovery run?

Continuously, with a monthly review. New AI tools and features appear every month, so a one-time audit can quickly become out of date.

What should a shadow AI report show the client?

The top AI tools by users and time, the share of personal-account sign-ins, blocked pastes and uploads, prompts flagged for sensitive data types, and a short list of recommendations. Lead with patterns, and use names only where a decision needs them.

Keep reading

For next steps after discovery, see how to make a client's AI policy enforceable, why personal AI accounts are the real leak, how MSP policy inheritance should work, and our page on shadow AI governance for MSPs.