Your clients want Claude in their SharePoint. Your security policy says not yet.
Every MSP is getting the same call right now. A client heard about AI agents, maybe from a competitor, maybe from a keynote, and now they want Claude or Copilot wired straight into their Microsoft 365 environment. For most clients, that’s a reasonable ask you can turn around in a week. For the client who handles confidential third party data and sits under an annual security audit, it’s a different conversation entirely.
That’s the exact bind one MSP described recently: a customer pushing hard for Claude connectors into SharePoint, while the MSP has spent months locking down that same SharePoint with security groups and permissioning tighter than anything else in their book of business. The AI governance document is written. The client adopted it company wide. And the connection request keeps coming back anyway.
The gap isn’t policy, it’s implementation
Most MSPs aren’t struggling to write an AI governance policy. Templates exist, frameworks exist, and a decent policy is a half day of work. The gap shows up one layer down: how do you actually let a client use Claude or Copilot against real SharePoint data without the connector becoming the new soft spot in an otherwise hardened environment.
That’s a permissioning question as much as an AI question. Security groups control who sees what today. An AI connector needs to inherit those same boundaries exactly, not a looser approximation of them, or the audit trail gets messy fast. If a security group blocks a user from a folder, an AI agent acting on that user’s behalf needs to hit the same wall, every time, with no exceptions baked in for convenience.
Audits don’t care how good your intentions are
An annual security audit isn’t asking whether you meant well. It’s asking for evidence: who can access what, who approved it, and what happens when someone leaves or changes roles. An AI connector that isn’t mapped cleanly to existing permission structures is exactly the kind of thing an auditor flags, even if nothing has gone wrong yet. The fix isn’t to slow walk the client’s request forever. It’s to treat the AI connection the same way you’d treat any new integration touching sensitive data: scoped access, logged activity, and a clear owner on both sides.
This is where agent design earns its keep
The MSPs handling this well aren’t the ones blocking AI requests, they’re the ones who can show a client exactly what an agent can and can’t touch before it goes live. When the agent’s scope is documented and testable, the audit conversation gets a lot shorter.
If you’re working through this exact tension with a client right now, happy to talk through how other MSPs are scoping AI access without reopening the security work they already did. See how MSPs are getting 20-25% time savings per ticket in our case studies, or check out mizo.tech for more.