Claude for Word (2026): The Add-In Is Real, Still in Beta, and Anthropic Will Not Say Which Plans Get It

The short answer

Claude for Word is real, it is published by Anthropic, and it is still in beta. As of September 21, 2026, Microsoft's marketplace lists "Claude by Anthropic for Word" — "Use Claude to review, redline, and draft documents" — and Anthropic's own beta features page lists Claude for Word under Integrations with the status Beta. Claude for Excel is not on that page, which is how you know it has already graduated. Anthropic publishes no plan matrix for the Word add-in beyond the words "availability varies by plan." That last part matters, because the confident "Pro and up" you will read elsewhere is an inference from how Excel rolled out, not something Anthropic has stated about Word.

What the add-in actually is

Search Microsoft's marketplace for Claude and most of what comes back is not from Anthropic. There are Excel formula helpers, email writers and meeting transcribers that call the Claude API and charge their own subscription, and several of them have Claude in the product name. The filter that matters is the publisher field.

Filtered that way, Anthropic ships three Office add-ins and an umbrella listing:

• Claude by Anthropic for Word — "Use Claude to review, redline, and draft documents."
• Claude by Anthropic for Excel — "Use Claude to analyze, edit, and create workbooks."
• Claude by Anthropic for PowerPoint — "Use Claude to build, edit, and refine presentations."
• Claude for Microsoft 365 — "Claude for Excel, PowerPoint, and Word. Edit your files with one conversation across apps."

Anthropic calls these "Office agents" in its own documentation, and they run inside an Office WebView — a browser panel docked in the app, which is also why Anthropic's telemetry docs note that gRPC is not supported for them. The umbrella listing's phrase "one conversation across apps" is the actual pitch: the same thread that read the spreadsheet can write the memo.

Note what is not in that list. There is no Outlook add-in, from Anthropic or from OpenAI, and email is reached from the other direction entirely — that is Claude for Outlook, and it is a different product decision, not an oversight.

The beta label is the story

Anthropic maintains a page listing everything currently in beta or research preview. Today that table contains, among a handful of Claude Code entries, exactly one Office integration: Claude for Word, Integrations, Beta. Excel is absent, because it went generally available on January 24, 2026 after a research preview that started in October 2025 with roughly a thousand testers.

Anthropic's own definitions, from the same page, are worth reading literally rather than skimming:

• Beta means "further along in development... stable enough for regular use," still being refined.
• "The experience may change."
• "Performance and reliability may vary." Beta features "haven't gone through the same level of polish."
• "Not every feature graduates. Some... become a permanent part of Claude. Others don't, and we sometimes remove features that don't work out."

That last sentence is the one to carry into a rollout decision. It is not a disclaimer lawyers added; it is a vendor telling you in plain English that this could go away. Excel took about three months to go from preview to GA. Word has been carrying the beta label longer than that. Whether that means more polish or less commitment, Anthropic has not said, and I am not going to read tea leaves — but it is a fact you should hold while deciding how many people to train on it.

What your admin can see

This is the part nobody writes about and every owner should know before rolling it out, and every employee should know before pasting a résumé into a Word document.

On Claude Enterprise, an administrator can point Office agents at the organization's own OpenTelemetry collector, configured under Organization settings → Office agents in the Claude.ai admin console. Anthropic's documentation describes what lands there in unusually direct language: each turn produces spans "capturing the prompt, model calls, tool executions, file uploads, and context compaction events," and the collector "receives all span attributes, including those carrying user-generated content (prompt text, tool inputs and outputs, document URLs, and filenames). No attributes are redacted or filtered."

Two balancing facts. Assistant response text is not included in the emitted spans. And when a custom endpoint is configured, the telemetry goes exclusively to the organization — Anthropic states spans "aren't dual-sent to Anthropic," and "your organization owns this data."

For a regulated business, that is genuinely good news: it is the control that makes a document-AI rollout defensible to an auditor, and it is more granular than most SaaS tools offer. For an individual using a company Claude Enterprise seat, the practical translation is simpler. Your prompts and the filenames you worked on are loggable. Behave accordingly.

Word vs Excel: which one is worth installing first

If you only do one, do Excel, and not because Word is bad. Excel is generally available rather than beta, it has been in the wild since January, and the work it does — analysing and editing workbooks in place — is the work where the time saved is easiest to see in a week. The full picture on that one is in Claude for Excel.

Word earns its slot in a specific kind of business: one where documents come back marked up. Contracts, proposals, statements of work, policies. "Redline" is Anthropic's own word for it, and it is not a generic AI-writing verb — it describes review with tracked changes, which is a different job from drafting from scratch. Agencies, law firms, contractors and anyone who sends a scope document is the audience. A business whose documents are mostly internal notes will get less from it than the marketing suggests.

What this costs you, honestly

If you already pay for Claude, the add-in itself is not a separate purchase — it is installed from Microsoft's marketplace and authenticates against the Claude account you have. The cost that is easy to miss is duplication. If your team is already on Microsoft 365 Copilot, adding Claude for Word means paying two vendors to do overlapping work in the same window, and the honest answer is usually to pick one and mean it. The same arithmetic applies on the OpenAI side, where the per-seat picture is laid out in ChatGPT Business pricing.

The suite-level view — which plan, what the connector reads, how the three add-ins fit together — is at Claude for Microsoft 365.

What to do this week

1. Check the publisher before you install. In Microsoft's marketplace, the listing must say Anthropic. Several popular results do not.
2. Install it on your own account first and see if it authenticates. Since Anthropic publishes no plan matrix, this is faster and more reliable than any article telling you which tier you need.
3. Pilot with two people and one document type. Pick the document that comes back marked up most often. A month of that tells you more than a feature comparison.
4. If you are on Enterprise, decide the telemetry question before the rollout, not after. Either configure your own collector or knowingly accept the default — and tell your team which one you chose.
5. Do not make it load-bearing. It is beta, and Anthropic says plainly that beta features sometimes get removed. Keep a process that works without it.

If the real question underneath all of this is which assistant your business should standardize on, what data never goes into any of them, and which three workflows to automate first, that is the AI Readiness Assessment: $2,500 flat, two weeks, credited in full against an install within 90 days. Installing an add-in takes two minutes. Picking the stack you will still be on in a year is the part worth thinking about.