In this guide · 8 sections
- 01What happens when someone clicks Connect
- 02The real risk: files that are already over-shared
- 03Google Workspace: control which AI apps can connect
- 04Microsoft 365: user consent and admin approval for SharePoint access
- 05Logs: what you can check afterwards
- 06An approve-or-deny policy for AI connectors
- 07Do it yourself or get help?
- 08Frequently asked questions
Short answer
Only connect business AI accounts (ChatGPT Business, Claude Team or Enterprise). In Google Workspace use API controls and in Microsoft 365 use user consent settings so staff cannot approve AI apps alone. Fix over-shared files first, because connectors can reach everything the user can.
What happens when someone clicks Connect
Connecting ChatGPT or Claude to Google Drive or Microsoft 365 uses OAuth: the user signs in to Google or Microsoft and approves a list of permissions (scopes) for the AI app. From then on, the app can act within those permissions on that user's behalf, until the access is revoked.
Both vendors say the connector mirrors the user's existing access. Anthropic's Google Workspace connector page states that Claude can only access data for the connected account and cannot reach anything the user could not already open. OpenAI's connected apps article says app permissions do not grant new access; what the app can see is set by the access granted at connection and by workspace controls.
| ChatGPT (apps) | Claude (connectors) | |
|---|---|---|
| Who can turn it on | Business: many apps on by default; admins can disable in the Admin Console | Team and Enterprise: an Owner must enable connectors before users can authenticate |
| Approval before actions | Options from Always ask to Allow all actions | Asks before sending email or sharing files by default |
| Training on connector data | Not by default on Business, Enterprise, Edu | Not trained on Gmail, Drive or Calendar connector data |
| Where retrieved data is kept | Under your ChatGPT data and retention settings | Stored with the chat; deleting the chat deletes it |
Personal plans are different. OpenAI says that on Free, Go, Plus and Pro, information accessed from apps may be used for training if the user's Improve the model setting is on.
Google Workspace: control which AI apps can connect
Google's admin article on controlling third-party app access describes the settings under Security, Access and data control, API controls.
Review apps that already have access
In API controls, open Manage Third-Party App Access and view Accessed apps. You see each app, how many users use it and which Google services (scopes) it requested.
Decide on each AI app
Mark the approved ChatGPT or Claude app as Trusted, mark others as Limited (unrestricted services only) or Blocked. Google-owned apps are trusted automatically.
Restrict unconfigured apps
Under Settings, Unconfigured third-party apps, allow only apps that ask for Google sign-in information. New apps that want Drive or Gmail access then need admin approval.
Restrict sensitive Google services
Mark Drive and Gmail as restricted services so only Trusted apps can reach them.
Tell staff how to request an app
Explain that new AI tools need approval and who to ask. Otherwise people switch to personal accounts.
Google notes that details of newly authorised apps can take 24 to 48 hours to appear in the list, so review it again a few days after any rollout.
Logs: what you can check afterwards
Knowing who connected what is half the control.
Connector audit checklist
OpenAI notes that compliance event coverage depends on the app and workspace configuration, and only eligible Enterprise and Edu workspaces have the Compliance Platform. On smaller plans, the Google and Microsoft logs are your main record.
An approve-or-deny policy for AI connectors
Write down a short rule set and share it with staff. It can sit inside your wider ChatGPT at work policy.
Approve when
- The AI account is a company Business, Team or Enterprise plan
- The user's file access has been reviewed recently
- The use case is clear, such as searching project files or drafting from templates
- Write actions (sending email, sharing files) still ask for approval
Deny when
- The AI account is personal (Free, Plus, Pro or Max)
- The user can reach HR, payroll or customer personal data they do not need
- The app is unverified or from an unknown publisher
- Someone wants "allow all actions" without a reason
If you are deciding where team knowledge should live in the AI tools themselves, see ChatGPT Projects vs custom GPTs vs Claude Projects.
Do it yourself or get help?
A Google Workspace or Microsoft 365 admin can change the app consent settings in under an hour. The time goes into the sharing review: finding and fixing folders, links and leftover access before connectors go live.
Our Microsoft 365 / Google Security Lockdown (S$590) restricts third-party app consent, tightens sharing defaults and documents an approval process. For larger file estates, the Access Review and Leaver Tool (from S$1,200) lists who can open what across your drives so you can clean up before connecting AI.
Frequently asked questions
It can reach the files the connected user can open, not the whole company's Drive. That is why over-shared folders matter: if a folder is shared with everyone in the domain, any user's connector can search it. On ChatGPT Business, OpenAI says connector data is not used for training by default.
In the Google Admin console, go to Security, Access and data control, API controls. Set unconfigured third-party apps to allow only Google sign-in information, mark approved AI apps as Trusted and others as Limited or Blocked, and restrict sensitive services such as Drive and Gmail. Staff then need admin approval for new apps.
In the Microsoft Entra admin centre, open Enterprise apps, Consent and permissions. Turn off user consent or limit it to verified publishers with low-impact permissions, and enable the admin consent workflow so users can request an app and named reviewers approve it. Turning the workflow on needs a Global Administrator.
Anthropic says data retrieved through its Google connectors is stored on its servers with the associated chat, encrypted at rest and in transit, and can be deleted by deleting the chat. It states it does not train its models on Gmail, Drive or Calendar connector data. On Team and Enterprise plans, an Owner must enable connectors first.
It can be if personal data is reachable and the tool is used in ways your notices and policies do not cover. Use business plans, limit what each user can open, keep customer and employee personal data out of AI workflows unless there is a clear purpose, and record the decision. This is general information, not legal advice.