SMSMC Blog

Seeing and controlling staff use of AI chatbots: Microsoft Purview, Chrome Enterprise Premium and Cloudflare One

Three ways to find out which AI tools staff use and to stop sensitive data being pasted into them: endpoint DLP, browser controls and a secure web gateway. What each sees and where each fits.

By Jim Venuto

The policy is written. Now what?

A 60-person accounting firm adopted an AI acceptable-use policy last spring. It says staff may use the firm’s approved AI assistant, and must not paste client tax data into public chatbots. The managing partner now asks two reasonable questions: which AI tools are people actually using, and would we know if someone pasted a client’s Social Security number into one?

A policy cannot answer either question. Answering them takes a technical control, and there are three broad places to put one: on the computer, in the browser, or on the network path to the internet. This post compares one well-documented product of each kind. They are not the only choices, but they show the trade-offs clearly.

Three places to put the control

ApproachWhere it runsSeesTypical fit
Endpoint DLPAgent on the managed computerPastes and uploads in supported browsers on that deviceOrganizations already on Microsoft 365 compliance tools
Browser controlsThe managed browser itselfWhat happens inside that browser, on any networkOrganizations that can standardize on one browser
Secure web gatewayTraffic routed through a cloud proxyWeb traffic from enrolled devices, including AI app use and prompts the gateway can inspectOrganizations that also want web filtering and remote access

All three need managed devices or a managed browser. None of them sees a personal phone on home Wi-Fi. That limit is worth saying out loud to leadership before anyone buys anything.

Microsoft Purview: endpoint DLP and DSPM for AI

Microsoft’s Purview documentation groups AI apps into three categories: Microsoft’s own Copilot experiences, enterprise AI apps connected to the organization (it lists ChatGPT Enterprise and Anthropic Claude Enterprise among them), and “other AI apps” detected through browser activity and categorized as generative AI in the Defender for Cloud Apps catalog, with ChatGPT, Google Gemini, the consumer version of Microsoft Copilot and DeepSeek as examples.

For the public chatbots, the control is endpoint DLP. Microsoft says Windows computers onboarded to Purview can be configured with endpoint DLP policies that warn or block users when they share sensitive information with third-party generative AI sites in a browser, for example blocking a credit card number pasted into ChatGPT, or showing a warning the user can override. Microsoft’s Data Security Posture Management (DSPM), and the classic DSPM for AI, act as the front door: reports on AI use, personalized recommendations, and one-click policies that use existing Purview information protection and compliance controls. Prompts and responses for supported AI apps are captured in the unified audit log.

Strengths: it uses the same sensitive information types, labels and audit trail you may already use for email and files, so a client-data rule is written once.

Limits: the documented endpoint DLP coverage for generative AI sites is for onboarded Windows computers. Licensing varies by capability; check the Purview service description against what you own before planning around it.

Chrome Enterprise Premium: control in the browser

Google describes Chrome Enterprise Premium as giving visibility into “Shadow AI” usage and applying real-time DLP to prevent sensitive data being pasted into prompts, with DLP controls for copy and paste, uploads, downloads and printing. Google’s Workspace Admin Help lists Chrome Enterprise connector policies as a prerequisite for these rules, and gives the actions as block, allow with warning, or audit only. When a user proceeds after a warning, that choice is logged. It is a paid per-user product.

Strengths: the control travels with the browser, so it works on any network. Google says the rules can monitor user actions in Chrome on Windows, Mac, Linux and ChromeOS devices. Warn-and-log is a gentle way to start: people see the rule at the moment it applies.

Limits: it covers Chrome. If staff also use Edge, Safari or a desktop AI app, those paths need their own control or a policy that Chrome is the only browser for work.

Cloudflare One: a secure web gateway with AI prompt inspection

Cloudflare One routes device traffic through Cloudflare’s gateway. Its Shadow IT discovery dashboard, built from Gateway HTTP traffic, shows SaaS application usage with application type, users, devices and data uploaded and downloaded, and lets an administrator mark each app as approved or unapproved. Cloudflare notes that the status itself does not block anything; Gateway policies do. For AI specifically, Cloudflare’s DLP documentation lists predefined “AI Prompt” profiles (AI Security, Customer, Financial Information, PII and Technical) that evaluate prompts sent to ChatGPT, Google Gemini, Perplexity and Claude for both the content of the prompt and the intent behind it. Attached to a Gateway HTTP policy, a profile can block or log matching prompts.

Strengths: one control point for every browser on an enrolled device, plus web filtering and remote access in the same service. The prompt profiles are specific to AI rather than generic pattern matching.

Limits: the dashboard and prompt profiles depend on Gateway HTTP filtering, so device traffic has to be routed through Cloudflare and inspected, and you have to decide which traffic you will not inspect. The AI prompt profiles name four AI services; other tools fall back to app-level allow or block.

Side by side

Microsoft PurviewChrome Enterprise PremiumCloudflare One
Where the control runsWindows endpoint agentManaged Chrome browserCloud gateway
Finds AI apps in useYes, through browser activity and the Defender for Cloud Apps catalogYes, Shadow AI visibilityYes, Shadow IT discovery
Stops sensitive pastes into public chatbotsWarn or block on onboarded Windows devicesBlock, warn or audit in ChromeBlock or log prompts to four named AI services
Reuses existing data rulesPurview sensitive info types and labelsChrome and Google DLP rulesCloudflare DLP profiles
Best whenYou are on Microsoft 365 with Purview licensingYou can standardize on ChromeYou also want web filtering and remote access

Choosing

  1. Start from what you own. If Purview licensing is already in place, configure it before buying anything else.
  2. Decide what you want to see first. Most organizations should run in audit or warn mode for a month. The goal is to learn which tools are used and how often, then decide.
  3. Give people an approved tool. Controls that only block push staff to personal phones. An enterprise AI plan with the right data terms, plus a clear rule, does more than any filter.
  4. Write down what you will not see. Personal devices, mobile apps and unmanaged browsers are outside all three. Your policy and training have to cover those.
  5. Tell staff what is monitored. Prompt logging captures what people type. Say so in the policy, and limit who can read the logs.

When to get help

If you have an AI policy but no way to know whether it is followed, a short review can match one of these approaches to the systems you already have and set a rule that starts in warn mode.

Want to talk it through? Book a free 30-minute call with Security Medic Consulting.

Sources