Why AI should never write to your tenant directly

Our design principle: AI proposes typed change sets, people approve, jobs apply with undo. Here's why.

Updated · 3 min read

The principle

Language models are good at intent and bad at guarantees. So the model's job ends at a proposal: a typed list of actions with predicted impact. A person with the Approver role decides. A background job applies only what was approved, honours Microsoft Graph throttling, and records how to undo every step.

What that buys you

  • Every change is reviewable before it happens
  • Partial approval — skip what you don't like
  • 30-day undo per change or per set
  • An audit trail that names requester and approver
  • The same rule for chat, Build, Tidy and MCP assistants

Keep reading

Talk to a SharePoint specialist

See IntraBuilt build, tidy and govern an intranet live — with every change reviewed before it's applied.