Microsoft 365 Copilot is no longer a single-model product. Since late 2025, Microsoft has been integrating Anthropic’s Claude and xAI’s Grok alongside OpenAI’s GPT models. This post covers what you need to know: where each model is used, what the subprocessor arrangement means for compliance, and why enabling Anthropic is less dramatic than it sounds.
The Model Landscape (April 2026)
OpenAI remains the default foundation model everywhere. Anthropic is available as a Microsoft subprocessor with enterprise data protections. xAI is available in Copilot Studio and Power Platform, under separate terms.
| Surface | OpenAI (GPT) | Anthropic (Claude) | xAI (Grok) |
|---|---|---|---|
| Copilot Chat (web, desktop, mobile) | ✅ Default | ✅ System selects best model per query; UI shows when Claude is active | ❌ |
| Copilot Cowork ³ | ✅ Default | ✅ User can select model | ❌ |
| Researcher | ✅ Available | ✅ User can select model | ❌ |
| Agent Mode in Excel ² | ✅ Available | ✅ User can select model | ❌ |
| Word, Excel, PowerPoint agents ² | ✅ Default | ✅ Admin toggle (separate for EU) | ❌ |
| Excel & PowerPoint (Copilot in apps) ² | ✅ Default | ✅ Admin toggle (separate for EU) | ❌ |
| Word (Copilot in apps) | ✅ Default | ⏳ Coming summer 2026 | ❌ |
| Copilot Studio (custom agents) | ✅ Default | ✅ Creator selects at build time (requires PPAC toggle) | ✅ Preview, US only ¹ |
| Power Platform | ✅ Default | ✅ Admin toggle in PPAC | ✅ Preview, US only ¹ |
¹ xAI Grok 4.1 Fast is in preview for US-based makers only. Unlike Anthropic, xAI operates under its own enterprise terms — not as a Microsoft subprocessor. Customer data is not retained for training, but models are hosted outside Microsoft-managed environments. xAI requires its own admin opt-in in the M365 Admin Center (under „AI providers for other large language models“ — a separate page from the Anthropic subprocessor toggle), plus the PPAC toggle for Copilot Studio.
² Agent Mode in Excel is the autonomous agent experience where users pick a model. „Copilot in apps“ refers to standard Copilot features in Excel and PowerPoint, controlled by the separate admin toggle introduced April 2026. „Word, Excel, PowerPoint agents“ are the standalone agents within those apps — distinct from the embedded Copilot experience.
³ Copilot Cowork is available in Frontier (since March 2026). Claude is selectable via the model picker. Cowork runs multi-step tasks autonomously within M365’s security boundaries.
Which Claude models, where
Microsoft doesn’t expose a single model everywhere. The specific Claude version depends on the surface and how selection works:
| Surface | Available Claude models | Selection |
|---|---|---|
| Copilot Chat | System-selected (latest available) | Automatic — Microsoft routes to the optimal model per query |
| Copilot Cowork | Claude Opus 4.7 (latest) | User picks via model selector |
| Researcher | Claude (version managed by Microsoft) | User picks Claude or GPT or Auto |
| Agent Mode in Excel | Claude (version managed by Microsoft) | User picks via model selector |
| Copilot Studio | Claude Sonnet 4, Claude Sonnet 4.5, Claude Opus 4.1, Claude Opus 4.5, Claude Opus 4.7 | Creator selects at agent build time |
| Word, Excel, PowerPoint agents | System-selected | Admin toggle — no user choice |
| Copilot in Excel & PowerPoint | System-selected | Admin toggle — no user choice |
Copilot Studio offers the broadest model catalog because creators build agents for specific tasks and need to match model strengths to use cases. In other surfaces, Microsoft manages the model version — users see „Claude“ without a specific version number. Model versions rotate as Microsoft and Anthropic ship updates; the table above reflects the state as of April 2026.
Apps where Claude is not available: Outlook, Teams, OneNote, Loop, Planner, Forms, and Whiteboard use OpenAI models exclusively. There is no announced timeline for Anthropic in these apps.
What Changed: From Separate Vendor to Microsoft Subprocessor
When Anthropic models first appeared in Copilot (September 2025), using them required accepting Anthropic’s own commercial terms and data processing agreement — essentially onboarding a new vendor. Procurement flagged it. Security demanded audits. Legal wanted a DPIA.
Since January 7, 2026, that’s over. Anthropic now operates as a Microsoft subprocessor:
| Aspect | Before (Sep 2025) | Now (Jan 2026+) |
|---|---|---|
| Contractual framework | Anthropic’s own Terms + DPA | Microsoft Product Terms + Microsoft DPA |
| Data processor | Anthropic (direct relationship) | Microsoft (Anthropic as subprocessor) |
| Enterprise Data Protection | ❌ Not covered | ✅ Covered |
| Customer Copyright Commitment | ❌ Not covered | ✅ Covered |
| Training data usage | Governed by Anthropic terms | Not used for training — same as GPT |
| Admin consent | Opt-in to Anthropic’s terms | Subprocessor toggle in M365 Admin Center |
| Vendor assessment needed | Yes — separate procurement | No — existing Microsoft agreement |
The legacy Anthropic toggle to opt in to Anthropic’s separate commercial terms and data processing agreement has been deprecated and replaced by this new Anthropic as a subprocessor admin toggle. — Microsoft Learn
Why the Subprocessor Model Matters
The short version: Anthropic is covered by your existing Microsoft enterprise agreement — no separate procurement, no additional DPA, no new vendor risk assessment. Enterprise Data Protection (encryption, tenant isolation, no training on customer data, GDPR alignment, sensitivity labels, audit logging) applies identically to Anthropic and OpenAI models. And Microsoft remains contractually accountable for what Anthropic does with your data.
The EU Data Boundary: What It Actually Means
The EU Data Boundary (EUDB) is Microsoft’s commitment to process customer data within EU/EFTA borders. It covers core M365 services and OpenAI models. Anthropic models are currently excluded. When Claude processes your prompt, data may be processed in the US.
That sounds alarming — but consider what still applies outside the EUDB:
| Protection | Inside EUDB (OpenAI) | Outside EUDB (Anthropic) |
|---|---|---|
| Microsoft DPA | ✅ | ✅ |
| Enterprise Data Protection | ✅ | ✅ |
| No training on your data | ✅ | ✅ |
| Tenant isolation | ✅ | ✅ |
| Encryption (in transit + at rest) | ✅ | ✅ |
| Customer Copyright Commitment | ✅ | ✅ |
| GDPR contractual alignment | ✅ | ✅ |
| Sensitivity labels respected | ✅ | ✅ |
| Data residency guarantee (EU only) | ✅ | ❌ |
The only difference is the geographic processing guarantee. Every other protection — DPA, EDP, encryption, tenant isolation — is identical. For organizations with strict data sovereignty requirements (public sector, regulated industries), the geographic guarantee may be non-negotiable. For most commercial enterprises, the Microsoft DPA already covers GDPR requirements for international transfers, including Standard Contractual Clauses.
This is why Microsoft’s approach makes sense: disabled by default in EU/EFTA/UK, but straightforward to enable if your compliance posture allows cross-border processing under the DPA.
How this compares to using Claude directly
If the lack of EUDB coverage gives you pause, it’s worth comparing the protections you get through Microsoft 365 Copilot with those available when using Claude directly. Anthropic’s own infrastructure stores data in the US. Their API currently offers "us" or "global" as inference regions — there is no EU-only option. On consumer plans (Free, Pro, Max), there is no DPA, no tenant isolation, no sensitivity labels, and no audit logging. Training opt-out is a per-user setting. Anthropic’s Enterprise plan adds contractual protections, but EU data residency is not among them.
| Protection | Claude via M365 Copilot | Claude directly (API / claude.ai) |
|---|---|---|
| Microsoft DPA | ✅ | ❌ Anthropic’s own terms |
| Enterprise Data Protection | ✅ | ❌ |
| Customer Copyright Commitment | ✅ | ❌ |
| Tenant isolation | ✅ | ❌ (Enterprise plan only) |
| Sensitivity labels respected | ✅ | ❌ |
| Audit logging | ✅ | ❌ (Enterprise plan only) |
| No training on your data | ✅ Organizational guarantee | ⚠️ Opt-out per user (consumer); contractual for Enterprise |
| EU data processing guarantee | ❌ | ❌ US storage; no EU-only inference option |
Both paths process data outside the EU. The difference is that using Claude through Microsoft 365 Copilot adds Microsoft’s DPA, Enterprise Data Protection, tenant isolation, sensitivity labels, and audit logging on top. For organizations already accepting cross-border processing under their Microsoft agreement, routing Claude usage through Copilot provides a governed, auditable channel rather than leaving it to individual employees.
Admin Configuration: What to Toggle Where
There are three separate admin controls — and they’re not in the same place.
| Control | Location | What it governs | Default (EU/EFTA/UK) |
|---|---|---|---|
| Anthropic as subprocessor | M365 Admin Center → Copilot → Settings → View all → AI providers operating as Microsoft subprocessors | Global prerequisite for all Anthropic usage. Directly enables Chat, Cowork, Researcher, Agent Mode in Excel. Required before Studio and M365 apps toggles take effect | Off |
| Copilot in M365 apps with Anthropic | Same page, separate toggle (new April 2026) | Anthropic as default model in Excel and PowerPoint (Word coming summer 2026) | On for tenants created after March 25, 2026; check Message Center for older tenants |
| External models in Power Platform | Power Platform Admin Center → Environment → Settings → Product → Features | Anthropic + xAI availability in Copilot Studio and Power Platform | Depends on M365 toggle |
The subprocessor toggle requires the Global Administrator role. The M365 apps toggle requires the AI Administrator role (Global Admin also works). The Power Platform toggle only works after the M365 Admin Center toggle is enabled.
Coming: Per-User and Per-Group Scoping (MC1263276)
Today, all three toggles are tenant-wide — all or nothing. That changes in late April 2026. Microsoft announced via Message Center post MC1263276 (Roadmap ID 557371) that admins will be able to assign third-party model providers to specific users or Entra ID groups:
- Scoping at the provider level (Anthropic, xAI) — not per individual model
- Enforced across M365 Admin Center, PPAC, and Copilot Studio
- Up to 999 groups and users; nested groups supported
- GA worldwide, late April 2026
This lets organizations pilot Anthropic with a specific team before going org-wide — or permanently limit it to roles that benefit most.
Regional Defaults
| Region | Anthropic | xAI | Notes |
|---|---|---|---|
| US, APAC, most commercial | 🟢 On | Preview (US makers only) | No data residency restrictions |
| EU, EFTA, UK | 🔴 Off (opt-in) | ❌ Not available | Anthropic excluded from EUDB |
| GCC, GCC High, DoD | ❌ Not available | ❌ Not available | No FedRAMP certification |
| Sovereign clouds | ❌ Not available | ❌ Not available | No toggle visible |
What to Tell Your Compliance Team
If your compliance lead asks „should we enable Anthropic?“ — four steps:
- Check your DPIA. If it already accounts for Microsoft subprocessors, Anthropic is covered — just update the documentation.
- Review data residency requirements. Is EU-only processing a hard requirement? If yes, keep the toggle off. If your org already accepts cross-border processing under the Microsoft DPA, enabling Anthropic doesn’t change the risk profile.
- Check the subprocessor list. Anthropic is on Microsoft’s Service Trust Portal — the same list your procurement team already monitors.
- Communicate to users. They’ll see a model selector in Researcher and Agent Mode. They don’t need to understand „subprocessor“ — they need to know it’s approved and secure.
Timeline
| Date | Event |
|---|---|
| September 2025 | Anthropic models first available in Researcher and Copilot Studio (under Anthropic’s own terms) |
| December 8, 2025 | New subprocessor toggle in M365 Admin Center; On by default for most commercial (Off for EU/EFTA/UK) |
| January 7, 2026 | Anthropic officially becomes Microsoft subprocessor; legacy toggle deprecated |
| March 2026 | Full subprocessor availability across all eligible tenants |
| April 3, 2026 | Separate toggle for Anthropic in Excel and PowerPoint (EU/EFTA/UK focused) |
| Late April 2026 | Per-user/group scoping (MC1263276, Roadmap ID 557371) |
| Summer 2026 | Word support for Anthropic models |
References
- Anthropic as a subprocessor for Microsoft Online Services — primary reference for subprocessor model, toggles, and regional defaults
- Copilot in Microsoft 365 apps with Anthropic models — April 2026 toggle for Excel and PowerPoint
- Enterprise data protection in Microsoft 365 Copilot — EDP overview including EUDB footnote
- Expanding model choice in Microsoft 365 Copilot — original September 2025 Anthropic announcement
- xAI models now available in Microsoft Copilot Studio — Grok 4.1 Fast preview
- Microsoft Data Protection Addendum (DPA)
- Microsoft Subprocessor List (Service Trust Portal)
- EU Data Boundary overview
- Use Claude with Researcher in Microsoft 365 Copilot — end-user model selection guide
- MC1263276 — Per-user/group scoping for third-party model providers (March 27, 2026)
Based on public Microsoft documentation as of April 21, 2026. Model availability and admin controls are subject to change — check the Message Center for the latest updates. Word support for Anthropic is announced for summer 2026 and not yet available.
Schreibe einen Kommentar