Copilot connector write actions: what should your team learn before Copilot starts changing data?

Short answer: Microsoft has made connector write actions generally available in Copilot Chat and Cowork, so Copilot can now create, update and delete information in connected apps such as Jira, Notion and Canva once a user approves the action. That shifts Copilot from answering questions to doing work in other systems, so staff need to learn how to check what they are approving, not just how to write a good prompt.

What Microsoft has announced

On 9 November 2026, Microsoft published a post on connector write capabilities on its Microsoft Copilot Blog. It says write actions are now generally available across Copilot Chat and Cowork, with hundreds of connectors built by Microsoft, partners and organisations themselves. Users can ask in plain language for Copilot to create, update or delete information without leaving the conversation. Microsoft says the feature is coming soon to Copilot in Office apps too.

Microsoft's examples include:

• turning a Teams meeting transcript into a CRM activity and a follow-up task;
• turning Outlook escalations into labelled product issues;
• creating Jira work items from a product review transcript, through Atlassian Rovo;
• updating a Notion project page from the week's status emails;
• applying team feedback to a Canva design.

Windows Report covered the launch the same day and noted that how widely organisations use these actions may depend on their security policies and approval rules.

How the approval step works

According to Microsoft, Copilot asks permission by default before it creates, updates or deletes anything. A confirmation card shows which tool Copilot wants to use, from which connector, and the details being sent.

Microsoft's documentation on federated connectors adds more detail:

• Actions run in the third-party service using the user's own permissions, so Copilot can only change what that person could already change.
• Nothing is sent to the connected app until the user responds, and declined actions are not carried out.
• The options are Allow once, Allow for conversation, Always allow and Cancel. The broader options apply to one tool, not the whole connector.
• If a publisher adds or changes a write tool, it goes back to needing approval.
• The Researcher agent stays read-only.
• Admins can audit activity in Microsoft Purview.

The same page tells organisations to update user and help desk guidance so people understand that actions are carried out with their own permissions in the third-party service. That is Microsoft telling customers, in its own documentation, that this needs explaining to staff.

Why this matters for AI training

Most Copilot training so far has been about getting good answers: writing clear prompts, checking outputs, and knowing when not to trust a summary. Write actions add a new skill. When Copilot proposes to create ten Jira tickets or update a client record, the person clicking Allow is responsible for that change.

Three things stand out.

Approving is a skill. A confirmation card is only useful if the person reads it. Staff need to know how to expand the card, check the details, and spot when Copilot has misread the request, such as the wrong project, the wrong record or too many items.

"Always allow" is a real decision. It saves clicks, but it also removes the check for that tool until the user changes the setting or the publisher changes the tool. People should know where to review and reset these choices, which Microsoft says is under Settings, then Sources, in Copilot Chat.

Clear requests matter even more. A vague prompt that produces a weak summary costs a few minutes. A vague prompt that updates the wrong records costs more. Good prompting habits, such as naming the system, the project and exactly what should change, now protect your data as well as saving time.

Practical takeaways by role

L&D leaders: Add a short module on approving Copilot actions to your Copilot training. Use realistic examples from each team, such as sales updating a CRM or project managers creating tickets, and practise reading confirmation cards before anyone works with live data. Our Microsoft Copilot training is built around practical, role-based tasks like these.

HR leaders: Check that your employee AI policy covers AI tools that take actions, not just tools that write text. Make it clear that staff remain responsible for changes they approve, and say who to tell if something goes wrong.

IT leaders: Microsoft says write tools come with the connector and are not enabled separately, so review each connector's tools before you turn it on or widen access. Use staged rollout to start with a small group, and brief the help desk, as Microsoft recommends. We focus on readiness checks and training, so connector configuration stays with your IT team.

Transformation leaders: Pick one or two workflows where moving information between systems wastes real time, such as meeting notes into a project tracker. Pilot them with a trained group, and set a baseline first so you can show whether it helped. Our Copilot rollout training plan and our guide to measuring AI training ROI can help with both.

A short checklist

1. Find out which connectors are on. Ask IT which connected apps your teams can use with Copilot, and which can change data.
2. Teach people to read the card. Make checking the confirmation card a normal habit, not an extra step.
3. Agree a rule on "Always allow". Decide whether staff may use it, and for which kinds of tools.
4. Update the policy. Cover AI actions, responsibility for approvals and how to report mistakes.
5. Start small. Pilot with one team and one workflow before rolling out more widely.

What this means for your team

Copilot is starting to do work in other systems, not just talk about it. That can save real time, especially on the dull job of copying information from meetings and emails into other tools. But it only works well when people know what they are approving and feel confident saying no. Teams that get a little training on this now will be far better placed as more connectors, and Office apps, gain the same abilities.

If you are comparing assistants, our guide to Copilot, Claude and ChatGPT training may help. You can also read our recent posts on Copilot Chat usage limits and the Copilot October 2026 update, or find out more about our corporate AI training and AI training in Leeds.

Talk to us

If you would like help getting your teams ready to use Copilot safely and with confidence, book a free 60-minute discovery call with Qasim.