The short answer
ChatGPT does not see your SharePoint by default. To read SharePoint documents it needs a connector: software that signs in to Microsoft 365, calls Microsoft’s APIs, and passes results back to the model. Depending on your ChatGPT plan and your provider’s current options, that connector might be one OpenAI offers, a custom connector you build, or a gateway that sits between ChatGPT and your systems. Check your provider’s current settings, because what is available varies by plan and changes over time.
Whichever route you take, the technical connection is the easy part. The organisational questions take longer.
How the connection usually works
- Sign-in. The connector uses Microsoft’s standard OAuth sign-in through Microsoft Entra ID. Someone consents to a set of permissions (scopes), such as reading files or sites.
- Search and retrieval. When a person asks a question, the connector searches SharePoint and returns matching documents or extracts.
- Answering. ChatGPT uses the retrieved text to answer. The document content is sent to the AI provider as part of that conversation.
That third point matters. Anything the connector retrieves becomes part of what the AI provider processes, so your provider’s data-handling terms apply to it.
Delegated versus application access
There are two broad ways a connector can reach SharePoint:
- Delegated access acts as the signed-in person. It can only see what that person can see in SharePoint. This is usually the safer default.
- Application access acts as the connector itself, often with tenant-wide reach. It is convenient for background jobs but can expose far more than any single user should see.
If a connector asks for broad application permissions across all sites, treat that as a significant decision for your Microsoft 365 administrator, not a routine click.
Five things to check before you connect
- SharePoint permissions are accurate. AI search is very good at finding documents people forgot were shared too widely. Review “everyone” and “company-wide” sharing on sensitive sites first.
- Scope is limited. Decide which sites and libraries are relevant. HR, payroll and board folders are often excluded at the start.
- Client separation holds. In a professional firm, one person’s access may span many clients. Consider whether the AI should work on one client at a time.
- Provider terms suit the data. Confirm how your AI provider retains and uses business data under your plan.
- There is a record. You should be able to answer “which documents did the AI retrieve, for whom, and when?”
Reading is different from writing
Reading documents is one level of risk. Uploading, editing, moving or deleting files is another. Many organisations allow AI to search and summarise SharePoint content but not to change it, at least initially. Make that distinction explicit rather than inheriting whatever the connector happens to allow.
How Kroy approaches it
Kroy holds the Microsoft 365 connection centrally, so ChatGPT never holds a SharePoint token. ChatGPT connects to Kroy; Kroy checks the person, the agent, what the person has delegated to it and the organisation’s policies, then retrieves only the documents in scope. The same connection serves Claude, Copilot or any other AI you approve, and every retrieval is recorded. Libraries you exclude stay excluded, whichever AI is asking.