For years the choice for AI inside Microsoft Office was effectively one product: Copilot. That changed when Anthropic shipped native add-ins that put Claude directly inside Excel, PowerPoint, Word, and Outlook. Instead of copying a spreadsheet into a chat window or wiring up a connector, your users open the app they already live in, click a side panel, and Claude reads and edits the document in place.
For an MSP, this is one of those quiet shifts that lands on your desk as a deployment ticket and a governance question at the same time. The good news is that these are standard Office add-ins, so they slot into the same admin tooling you already use. The part that needs thought is what it means to give an AI agent read-and-write access to your clients' financial models, contracts, and inboxes. This post covers both.
What the add-ins actually are
Claude's Microsoft 365 add-ins are task-pane add-ins that run inside the Office apps themselves. When a user opens the Claude panel in Excel, Claude can see the workbook, reason about the cells and formulas, and make edits directly in the sheet. The same pattern applies in Word (tracked changes, comment threads, company styles), PowerPoint (building and editing slides with native charts), and Outlook (inbox triage and drafting).
The headline feature is shared context across apps. Claude treats the four apps as one workspace: you can build a model in Excel, then ask Claude to turn it into a PowerPoint deck without re-explaining the assumptions. As Anthropic puts it, context carries from Outlook to Word to Excel to PowerPoint, and when you change one thing, every open file catches up. That cross-app continuity is the main thing that separates the add-ins from a one-off chat session.
Why in-app beats a connector
It is worth being precise about why this matters, because Claude can already reach Microsoft data other ways. There is an MCP-style Microsoft 365 connector that lets Claude.ai search and pull from your files. The connector is good for retrieval. The add-in is different: it operates inside the live document with full read/write, sees your current selection, and applies edits you can review in the app's own change-tracking. A connector answers questions about your files; the add-in does the work in them. For an analyst rebuilding a model or a consultant cleaning up a deck, that distinction is the whole point.
Which apps, which plans, which platforms
| App | Status | Plans |
|---|---|---|
| Excel | Generally available (May 7, 2026) | Pro, Max, Team, Enterprise |
| PowerPoint | Generally available (May 7, 2026) | Pro, Max, Team, Enterprise |
| Word | Generally available (May 7, 2026) | Pro, Max, Team, Enterprise |
| Outlook | Public beta | All paid plans |
All four require a paid Claude plan; the free tier does not include them. Users also need a current Microsoft 365 subscription. Platform support is broadest in Excel, which runs on the web, Windows (build 16.0.13127.20296 or later), Mac (16.46+), and iPad (2.51+). In practice, plan on Windows, Mac, and Office on the web for the core experience and treat older desktop builds as a prerequisite to patch before rollout.
There are two AppSource/Marketplace listings to be aware of: one bundle covering Excel, PowerPoint, and Word, and a separate listing for the Outlook beta.
How users install it
For a single user the flow is simple:
- Go to the Microsoft Marketplace listing for Claude for Microsoft 365 (Excel, PowerPoint, and Word) and click Get it now.
- Open Excel, Word, or PowerPoint and open the Claude panel.
- Sign in with a Claude account (or your organization's cloud-provider connection) to activate it.
Outlook is a separate install from its own listing while it remains in beta.
How IT admins deploy it across a tenant
For managed rollouts, do not rely on end users pulling from the store. Deploy centrally from the Microsoft 365 Admin Center:
- Confirm Settings → Org Settings → User owned apps and services has store access set the way your policy requires.
- Go to Settings → Integrated Apps → Add-ins → Deploy Add-in.
- Choose Choose from the Store, search for the Claude listing, and add it. Alternatively, use Upload custom apps and provide the add-in's
manifest.xml. - Scope the deployment: entire organization, specific users, or specific security groups. Groups are the cleanest option for a permanent rollout; specific users are best for a pilot.
A few field notes. Propagation can take up to 24 hours, though it is often faster. Some admins hit an error trying to push the AppSource SaaS listing through the generic Integrated apps → Get apps path; the supported route is the Office Deploy Add-in flow above, or the custom-manifest upload. For larger or compliance-bound tenants, Anthropic also provides a claude-in-office setup wizard that generates manifests, provisions cloud resources where needed, and handles tenant-wide admin consent through Microsoft Entra ID.
Data governance
This is the part to slow down on before you flip it on for a client. You are granting an AI agent live read/write access to spreadsheets, documents, and email.
- Retention. Anthropic states that inputs and outputs are deleted from its backend within 30 days except in specific circumstances, and that commercial data is not used to train models by default. Chat history lives locally in the browser (IndexedDB) and is not synced across devices.
- Keep inference in your boundary. Rather than sending prompts to Anthropic's API, you can route through Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, or an LLM gateway so prompts and responses stay inside your existing trust boundary. In that model, Anthropic collects operational telemetry (feature usage, performance) but not prompt content. Note that a few features, including working across apps, require a Claude account sign-in.
- Scope, don't blanket-deploy. Use group-based deployment so only users who need it get the panel, and run a pilot first.
- Treat it like any AI tool. The same discipline from our AI coding CLI best practices applies: review every change before it lands, keep secrets out of prompts, and document the policy. The human-in-the-loop design in Outlook (drafts only, nothing sends until you click send) is a model worth enforcing elsewhere. For a deeper look at guardrails, see our notes on Claude Code security guidance.
The bottom line
Claude's Microsoft 365 add-ins move AI from a separate chat window into the documents your clients actually work in, with shared context that follows a task from inbox to spreadsheet to slide deck. Excel, PowerPoint, and Word are generally available on paid plans; Outlook is in public beta. Deployment is standard Office add-in territory, so it fits your existing admin workflow, scope it with security groups and pilot it first. The real work for an MSP is on the governance side: decide whether inference runs through Anthropic or your own cloud boundary, confirm retention and training settings meet each client's requirements, and roll it out deliberately rather than tenant-wide on day one.