Skip to content

Microsoft 365

Integration · Workspace and scheduling

Connect an AI agent to Microsoft 365 through Microsoft Graph.

An AI agent works with Outlook mail and calendars, OneDrive, SharePoint and Teams through Microsoft Graph. Access comes from an app registered in Microsoft Entra ID with specific permissions, which an admin approves. The key decision is whether it acts as a signed-in user or as an app.

What can an agent do in Microsoft 365?

Microsoft Graph is one API in front of most Microsoft 365 data. An agent with the right permissions can work across it:

  • Outlook mail: read and categorise messages in a shared mailbox, draft replies, file attachments. See email triage.
  • Outlook calendars: check availability, create and update meetings, including Teams meeting links where your tenant allows.
  • OneDrive and SharePoint: find, read and save documents, for example to answer staff questions from your policies. See internal knowledge assistant.
  • Excel workbooks stored in OneDrive or SharePoint: read and update table data.
  • Teams: post notifications and handoffs into a channel or chat, so a person can pick up what the agent could not finish.

For example, a hypothetical accountancy practice could have an agent read client emails arriving in a shared mailbox, save attached documents to the right client folder in SharePoint, and post a summary to the relevant Teams channel.

Delegated or application permissions?

Every Graph integration starts with an app registration in Microsoft Entra ID (formerly Azure Active Directory). That registration requests permissions of one of two types:

The two ways an agent can access Microsoft 365
TypeWhat it means in practice
DelegatedThe agent acts on behalf of a signed-in user and can never exceed what that user can do. Suits assistants used by one person.
ApplicationThe agent acts as itself, without a user, typically across the whole tenant for that permission. Suits background processes, and always needs admin consent.

Application permissions such as reading mail apply to every mailbox by default. Exchange Online lets an admin restrict an app to specific mailboxes, and you should use that so an agent built for the accounts inbox cannot read the managing director’s email.

What does setup involve?

  1. Agree which mailboxes, sites and channels the agent needs, and what it may change.
  2. Register an app in your Entra tenant, owned by your organisation, with only the permissions on that list.
  3. Have a Global Administrator or another suitably privileged admin grant consent.
  4. Restrict mailbox access to the agreed mailboxes, and site access to the agreed SharePoint sites where possible.
  5. Store the app’s secret or certificate in a secrets manager, with an expiry date someone will be reminded about.

An app used only in your own tenant does not need Microsoft’s marketplace review. Publishing an app for other organisations involves verification and, for listing, Microsoft’s own certification process.

Before you connect Microsoft 365: a checklist.

Checklist

Connecting Microsoft 365 safely

What limits should you expect?

  • Throttling: Graph slows down or rejects requests when an app sends too many, with limits that vary by service. Build in retries that respect the wait time Microsoft returns.
  • Teams messaging: posting as an app into channels and chats has more restrictions than reading. A notification via a bot or workflow is usually simpler than full two-way chat.
  • Licences and plans: some features depend on the Microsoft 365 plan, so check before you design around them.
  • Data residency: Microsoft offers EU data boundary commitments for its cloud services. That covers Microsoft’s storage, not the AI model your agent calls, which you need to choose separately.

Power Automate, Copilot or a custom agent?

If you already pay for Microsoft 365, look at Power Automate first for rule-based flows such as saving attachments or posting alerts. Microsoft’s own Copilot features may cover personal productivity inside Office apps. A custom agent is worth it when the work crosses into systems Microsoft does not reach (your CRM, accounting or WhatsApp), needs your own rules and approvals, or you want control over which AI model handles your data. See build vs buy.

What does a sensible first release look like?

Keep the first version narrow enough that your admin is comfortable approving it. For example: one shared mailbox, read and draft permissions only, restricted to that mailbox, with notifications into one Teams channel.

Run it for a few weeks with a person sending every draft. Track how many drafts were sent unchanged, lightly edited or rewritten, and why. That record tells you whether to let the agent send routine replies, add a second mailbox, or tighten its instructions first. Our guide to human-in-the-loop AI covers how to set those approval steps up.

Common questions

Can the agent send email as a shared mailbox?

Yes, with the right permission and access to that mailbox. Many teams start by having the agent save drafts in the shared mailbox for a person to send, and move to automatic sending only for routine replies once they have reviewed enough of its output to trust it.

Does it work with on-premises Exchange?

Microsoft Graph is built for Exchange Online. Older on-premises or hybrid setups are more complicated and may need a different approach. If your mail still runs on your own servers, raise it early in scoping, because it affects both the effort and the security design.

Can we limit the agent to one SharePoint site?

Graph supports permissions that grant an app access only to specific sites, rather than all of SharePoint, with an admin assigning each site. Use that model for a knowledge assistant, and keep sensitive libraries such as HR or finance out of the agent’s reach unless the use case needs them.

What happens when the app secret expires?

The integration stops working until someone adds a new secret or certificate. This is one of the most common causes of a silent failure. Record the expiry date, alert someone well before it, and prefer certificates managed in a key vault over secrets pasted into configuration files.

Bring us the task.

Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.

Start a project.

Two people on a sofa with laptops, planning a project beside a whiteboard of sticky notes

Book a call or send us a message. We read everything.

tash@neu.ie
Working with teams in Ireland and worldwide.

Get in touch

Free 30-minute AI review.