What changes with write actions?
Until now, Copilot connectors mainly retrieved information from other applications. With write actions, where a connector provides the right tools, Copilot can now also create, update or delete information there without users leaving the conversation. According to Microsoft, hundreds of Microsoft-, partner- and custom-built connectors are available; capabilities vary by connector and may depend on configuration and permissions.
Microsoft gives examples: a Teams meeting transcript becomes a CRM activity with a follow-up task, a draft document becomes a Canva graphic, and approved rows in an Excel workbook become records in the connected accounting system. This is generally available in Copilot Chat and Cowork. For Copilot in the Office apps, Microsoft says write actions are coming soon.
Which connectors are mentioned?
The blog post describes scenarios with six partners: Atlassian Rovo creates and updates Jira work items and creates Confluence pages. Notion pages and databases are updated, and Canva designs are created and revised. G2 plans and creates review campaigns, iManage drafts legal documents based on existing agreements, and BioRender creates scientific figures.
Which actions a connector actually offers is decided by its publisher; according to Microsoft Learn, publishers can add or change tools over time. The gallery overview on Microsoft Learn lists many more federated connectors, including ones relevant in Germany such as a third-party DATEV connector. Check whether a connector can write case by case instead of assuming it.
How does confirmation before an action work?
By default, Copilot asks for permission before it creates, updates or deletes information on the user's behalf. A confirmation card shows which tool from which connector is to be used and which details are being sent; expanded, it lets users review the parameters. According to Microsoft Learn, nothing is sent to the connected system while confirmation is pending. Declined actions are not performed.
Per tool, users choose: Allow once runs the action and asks again next time. Allow for conversation stops asking for this tool for the rest of the current conversation. Always allow stops asking until the user changes the setting or the publisher changes the tool. Cancel declines. Under Settings, Sources, permissions can be reviewed and reset; new or changed write tools return to Needs approval.
Which permissions and data does Copilot use?
Federated connectors use the Model Context Protocol and access the target system with the user's identity and permissions. Copilot can only change what the user is already allowed to change there. Data is not indexed or stored in Microsoft 365; Copilot sends the connector the information it needs for the action. Microsoft Learn notes that such actions export customer data to the connected system.
For companies in the EU, this is a data protection matter: writing content from a meeting transcript into a third-party system transfers data to another provider. Microsoft recommends reviewing a connector's capabilities, privacy terms and third-party agreements against your own security and compliance requirements before enabling it. Involve your data protection officer before connectors with write permissions are opened to everyone.
What can admins control and audit?
Admins manage federated connectors in the Microsoft 365 admin center under Copilot connectors, Your connections. They can turn connectors on or off tenant-wide, limit them to users or Entra ID groups via staged rollout and block entire publisher groups under Agents, Settings. Write, update and delete tools are part of the connector and are not enabled separately. Activity can be audited in Microsoft Purview.
According to the blog, admins can open each MCP-based connector and see which tools create, update or delete data. Microsoft Learn marks this tools view in the admin center as coming soon and not available for every connector; as a fallback, the page points to the tool list in a user's Copilot settings. The Researcher agent does not support write actions and stays read-only.
How should companies proceed?
Write actions save clicks but shift responsibility: an approved action is a real change in the target system. We recommend a controlled start.
- Inventory enabled connectors and check for each one whether it offers write tools.
- Limit write-capable connectors to a pilot group via staged rollout at first.
- Define in training and the AI policy when Always allow is acceptable, for example never for delete tools.
- Update helpdesk guidance: actions run with the user's own permissions in the third-party system.
- Plan Purview auditing for connector activity and document the data protection review per provider.
Keep it verifiable





