We're trying to figure out the best way to manage AI tools in our enterprise while preventing employees from entering sensitive company information into them. So far, we've blocked access to most AI services, allowed Copilot, and published a policy prohibiting the use of unapproved tools. The problem is that Copilot doesn't meet every department's needs, and leadership now wants one team to use Claude.
We're implementing DLP controls in Microsoft Purview and reviewing Defender cloud policies. Our Microsoft 365 environment is licensed for E5, but our domain is federated through Okta for SSO, so I'm unsure which session-control features are available. What other technical or policy controls are people using to manage AI safely? In particular, how are you stopping users from uploading things like payroll spreadsheets to personal ChatGPT or Claude accounts while still giving departments a useful, approved AI option?
4 Answers
Blocking every unapproved service usually pushes people toward personal accounts and unmanaged phones, where you lose visibility completely. The practical goal is to make the sanctioned tool easier and more capable than the workaround: provide approved accounts, useful model access, training, and clear examples of prohibited data.
Combine that with DLP, identity-based access, browser and endpoint controls, logging, and periodic reviews. Assume some sensitive data may already have been entered and focus on reducing future exposure, detecting risky behavior, and giving users a safe alternative.
For Microsoft 365 content, review sensitivity labels, Azure Rights Management permissions, and extraction or copy restrictions for sensitive files. Those controls can help prevent authorized AI experiences from handing protected content to users who shouldn’t be able to extract it.
Be careful not to assume that the same controls protect every external connector. Content brought in through a Graph or other third-party connector may be governed by the source system’s ACLs and the connector’s configuration rather than by the file’s RMS permissions. Audit connector permissions and test for oversharing separately.
That distinction is important. File permissions and sensitivity labels may protect normal SharePoint or OneDrive content, but they shouldn’t be treated as a complete control for externally indexed data.
A Microsoft-focused organization can use Purview, Defender, firewall or secure web gateway rules, Intune browser policies, and identity controls together, but expect some redesign and testing. Browser restrictions may reduce access to built-in assistants such as Gemini, although they can also affect legitimate workflows.
Don’t rely on a simple domain block as the entire strategy. AI providers change domains and users can switch browsers or devices. Track sanctioned applications, control unmanaged sessions where your identity stack supports it, monitor uploads and copy activity, and establish an exception process with an owner, expiration date, and documented data-handling requirements.
Start by defining exactly what the department needs from Claude. If it’s just a chatbot, an approved Copilot tier may provide access to multiple models while keeping the enterprise security and auditing controls. If they need Claude-specific agents, coding tools, or other advanced features, you may need an enterprise Claude agreement with contractual data-protection terms and administrative controls.
Also, treat the policy as a business decision. If leadership wants an exception, have them formally approve the exception and accept the associated risk rather than quietly bypassing the existing policy.
That makes sense. The difficult part is convincing people to stop using personal AI accounts when they already find them more useful.