Microsoft brings Grok into Copilot for Word, Excel and PowerPoint

• FACT (CONFIRMED): On 2026-09-12, Microsoft published "Expanding model choice in Copilot with Grok" on the Microsoft Copilot Blog (Tech Community), announcing that Grok models from SpaceXAI are being added to Copilot in Word, Excel and PowerPoint, delivered through the Microsoft Frontier Program as a "focused release to gather customer feedback." (Primary post fetched directly; URL: techcommunity.microsoft.com/blog/microsoft-copilot-blog/expanding-model-choice-in-copilot-with-grok/4555749. Post dated Sep 12, 2026, updated Sep 14, 2026, v2.0.)
• FACT (CONFIRMED): The post's first line: "Today we are expanding model choice in Copilot with the addition of Grok models from SpaceXAI." Microsoft frames this as "our commitment to bring the latest AI innovation in the industry to Copilot, tuned for work and tailored to business needs."
• FACT (CONFIRMED): To support this, SpaceXAI has been added to Microsoft's Online Services Subprocessor List — the legal register of third parties allowed to process Microsoft customer data. Access to SpaceXAI models is managed through a dedicated administrative setting that is disabled by default; organizations must enable the setting first.
• FACT (CONFIRMED): Grok is not available to Frontier customers in the European Union (EU), European Free Trade Association (EFTA), and the United Kingdom during the preview. Microsoft gives no reason in the post; it only says it will evaluate customer feedback, usage patterns, and product fit before broadening availability "across additional Copilot experiences and surfaces."
• FACT (CONFIRMED): CEO Satya Nadella posted on X on 2026-09-12: "More model choice coming to Copilot. Welcome Grok!" and the @Microsoft365 account posted: "Launching even more models in Copilot. Grok models from SpaceXAI are rolling out in Copilot in Word, Excel, and PowerPoint—starting with a focused release to customers in the Microsoft Frontier program." (Quoted verbatim by techAU and NDTV Profit.) Elon Musk subsequently confirmed on X: "Grok available in Microsoft Copilot." (Big Hat Group.)
• FACT (CONFIRMED context): This is a third external model family in Copilot: Microsoft had already added Anthropic Claude across Copilot experiences (Researcher agent and Copilot Studio last year; Copilot Cowork, Copilot Studio, and Copilot in Excel from early 2026), plus its own in-house MAI family. Grok makes OpenAI + Anthropic + Grok + MAI coexist. (Windows Central, thrumos, eesel, Big Hat Group; Microsoft Learn "Overview of AI Subprocessors" documents the Anthropic examples.)
• FACT (CONFIRMED context): The Grok-in-Copilot story is three separate rollouts on three different model bodies: (a) GitHub Copilot has offered Grok 4.6 in its model picker since 2026-08-14; (b) Copilot Studio has offered Grok 4.1 Fast since 2026-02-19 (Microsoft Copilot Blog, Feb 19, 2026); (c) the new Microsoft 365 Office preview is described only as "the model family from SpaceXAI" with no version number named. (eesel; Microsoft Copilot Blog Copilot Studio post; GSMDome.)
• FACT (CONFIRMED context): xAI/SpaceXAI already sold standalone Grok add-ins for Microsoft 365 (PowerPoint 2026-06-16, Word 2026-06-18, Excel 2026-07-20) and Grok Bot integrations for Outlook/Calendar/Teams/SharePoint/OneDrive after May 2026. The September announcement is different: Grok moves inside Microsoft's own Copilot experience rather than a separate Grok panel. (Sebertech, The Tech Boom, NDTV Profit.)

• Copilot is no longer effectively OpenAI-first inside Office. For the first time, a Musk-lab model becomes an in-app, licensed Copilot model option in Word, Excel and PowerPoint — the apps where enterprise knowledge work happens — rather than a separate add-in, API call, or side panel.
• The "exclusive-era" framing died in Office. Copilot now explicitly offers OpenAI, Anthropic and Grok in the same interface (drop-down model selection, per Windows Central), plus Microsoft's own MAI models in the broader catalog. "Copilot used to mean 'OpenAI, wrapped in Microsoft's apps.' Now it means a menu" (eesel).
• A new legal/architectural pattern was established: third-party frontier models as subprocessors with off-Microsoft data processing — Grok is the second major instance after Anthropic, and the pattern (subprocessor list + admin gate + non-Microsoft terms) is now the template for any future provider (Google, Meta, Mistral, etc.).
• Enterprise AI procurement shifted from "which model does Microsoft give me?" to "which model does my admin let me pick?" — model choice is now an admin-controlled governance decision inside the Microsoft 365 Admin Center, not a Microsoft product decision.
• SpaceXAI (xAI) gained enterprise legitimacy at the distribution layer: it does not need its own sales force to reach Office's install base; it needs a Microsoft admin to flip a switch.

• Before (through 2026-09-11): Copilot in Word/Excel/PowerPoint ran on Microsoft-hosted models (primarily OpenAI GPT family, with Microsoft's MAI and Anthropic Claude arriving across selected experiences such as Researcher, Cowork, and Copilot Studio). Grok reached Office only via separate marketplace add-ins (June–July 2026) or Grok Bot; Europeans and UK users of those standalone add-ins were served by SpaceXAI/xAI directly. GitHub Copilot got Grok 4.6 in its picker on Aug 14; Copilot Studio got Grok 4.1 Fast on Feb 19. No Grok inside the Microsoft 365 document/sheet/deck Copilot experience itself.
• Change (2026-09-12): Microsoft announced Grok models from SpaceXAI rolling into Copilot inside Word, Excel and PowerPoint via the pending Frontier release; SpaceXAI added to the subprocessor list; EU/EFTA/UK Frontier customers excluded; admin opt-in gate, off by default. Nadella publicly welcomed Grok; Musk confirmed.
• After (as of 2026-09-14+): Eligible Frontier tenants whose admins enable SpaceXAI see GPT / Claude / Grok as selectable models (or auto-pick) inside Office apps on web and desktop (Windows Central). The same "model choice" message now defines Copilot's enterprise positioning; SpaceXAI is a distribution-layer partner. Broader availability (Teams, Outlook, GA outside Frontier) is explicitly pending feedback and usage.

• Where it appears: Copilot in Word, Excel, PowerPoint (web and desktop) for organizations/users in the Microsoft Frontier Program — Microsoft's tenant-level early-access tier for AI features pre-GA. Eligibility requires a Microsoft Copilot license (for business) plus Frontier enrollment; individuals on Microsoft 365 Pro/Premium/Personal/Family can enable "Preview Frontier Features" via Copilot settings in supported web versions, but the Grok piece is primarily positioned as a business/enterprise feature. (Office Watch, Sebertech, The Grid Fin.)
• The admin gate (COMPANY CLAIM, corroborated by Office Watch/The Grid Fin): in the Microsoft 365 Admin Center → Copilot → Settings → "AI providers for other large language models" → SpaceXAI, the admin enables the provider by accepting SpaceXAI's Terms and Conditions. The setting is disabled by default; nothing changes until an admin deliberately opts in. Microsoft can also restrict a provider to specific users/Entra ID security groups (Microsoft Learn "Assign AI provider access to users and groups", which links a "Connect to SpaceXAI models" page).
• The user-facing picker (INDEPENDENT EVIDENCE — Windows Central, 2026-09-14): "Once an admin enables the integration, you can set Copilot in apps like Word, Excel, and PowerPoint to automatically pick the best model for a task, or manually pick between GPT, Claude, or Grok directly from a drop-down menu in the interface." So: model selection is manual or auto, and Grok is an option, not the default.
• Subprocessor mechanics and data flow (COMPANY CLAIM + INDEPENDENT EVIDENCE): SpaceXAI is on the Online Services Subprocessor List. Per Microsoft Learn and Office Watch, when an organization uses a Grok model, data "is processed outside all Microsoft managed environments and audit controls" — Microsoft's own customer agreements, Product Terms, Data Protection Addendum, data-residency commitments, audit/compliance requirements, SLAs and Customer Copyright Commitment do not apply to SpaceXAI usage; instead xAI's Enterprise Terms of Service and Data Processing Addendum govern (identical structure to the Copilot Studio Grok 4.1 Fast rollout, per Microsoft's Feb 2026 Copilot Studio post). The Copilot Studio post also states customer data is not retained or used to train xAI models in that surface.
• Guardrails: Grok runs below Microsoft's system prompts and enterprise guardrails, so output register is professional — Windows Central explicitly notes users "should not expect the witty or irreverent tone... on X."
• Versioning: the Office preview is version-agnostic publicly ("Grok model family"); in parallel, GitHub Copilot's picker names Grok 4.6 (latest frontier flagship, ~500k-token context per eesel; Foundry docs list grok-4.6/4.3/4-20 variants) and Copilot Studio runs Grok 4.1 Fast.

• Grok becomes a mainstream office-AI option. Office is the default workplace software for hundreds of millions of users; putting Grok behind the same Copilot pane as GPT and Claude converts xAI from a consumer/developer brand into an enterprise-distributed model family — precisely the distribution xAI could not build itself (The Information, Jan 2026, framed xAI's uphill enterprise selling battle).
• The OpenAI–Microsoft "exclusivity era" is officially over at the product layer. Copilot is now a model marketplace ("the model switchboard inside Word, Excel and PowerPoint" — The Tech Boom). Microsoft's leverage is its distribution, not any single model.
• It legitimizes "multi-model procurement" as the enterprise norm. CIOs must now govern which models employees use for which tasks — model selection becomes a per-task optimization and compliance decision, not a tenant-wide standard (Big Hat Group).
• It strengthens the platform-owner thesis. Forbes reads the terms (off by default, EU/UK exclusion, subprocessor status, no Microsoft DPA coverage) as Microsoft "dictating terms to model makers," positioning itself as the essential gateway; SpaceXAI accepts the terms to reach enterprise customers.
• Timing signal: the announcement lands days after Anthropic/OpenAI/xAI CEOs publicly endorsed pacing the frontier (Amodei's "We Must Pace the Frontier" essay, Sep 12) and Microsoft published its MAI Code of Conduct draft (Sep 14) — yet Microsoft simultaneously expands third-party frontier distribution, underscoring the commercial/regulatory tension in this window.

• Per-task model routing in Office: draft in Word with one model, run Excel analysis with Grok, structure slides with Claude — without leaving the app or adding licenses per provider.
• A/B and redundancy practices: teams can compare model outputs for the same document/sheet/deck task; Grok serves as redundancy against outages or quality issues in other subprocessors (Windows Central).
• Competition on office-work quality: model providers now compete for the same Word/Excel/PowerPoint workflows — quality, latency, and responsible-AI posture become visible to enterprises at the point of use.
• A blueprint for future providers: Google, Meta, Mistral, or open-weight hosts can now enter Copilot through the same subprocessor + admin-gate mechanism (the Fireworks AI-on-Foundry open-model example shows the pattern extending beyond frontier labs).
• Grok's enterprise credentials, under Microsoft guardrails: enterprises get Grok's reasoning/tool-use strengths inside a governed, professional-register wrapper without adopting X/Grok's public persona.

• Which specific Grok model(s) power the Office preview, and does it vary by task (drafting vs. formula generation vs. deck structure)?
• What are the commercial terms between Microsoft and SpaceXAI — revenue share, inference placement, exclusivity constraints (none claimed), and did EU/UK exclusion come from Microsoft or SpaceXAI?
• When and where does GA arrive — will Teams/Outlook follow, and will EU/EFTA/UK unblock change?
• Does data used for Grok in M365 Copilot train xAI models? (Copilot Studio said no; the M365 post does not say.)
• How does auto-pick decide "best model for a task," and can enterprises audit those routing decisions?
• Will the MAI Code of Conduct (scope: first-party only) ever extend to subprocessor models like Grok, or will Microsoft publish per-provider responsible-AI reports?
• Does the Excel COPILOT function retirement (September 14, 2026, per Windows Central) shift users toward the new picker and change the Grok-in-Excel calculus?
• What does the SpaceX/xAI IPO-era structure (SpaceXAI) mean for Grok's enterprise governance and for continued Microsoft partnership stability?

• Short term (weeks): the Frontier preview rolls out to eligible tenants ("coming week with admin opt-in" per the post's footnote); Microsoft aggregates feedback and usage; expect community and press teardowns of Grok's actual Office behavior and version identification.
• Medium term (months): Microsoft will likely extend Grok beyond Word/Excel/PowerPoint (Teams and Outlook named as candidate surfaces by techAU/others) and, if pilot metrics hold, move toward GA — with the EU/EFTA/UK exclusion revisited as a compliance prerequisite.
• Model-catalog cadence: expect more providers through the same subprocessor template; the picker may become the flagship demo of Copilot's "model-agnostic platform" story at Ignite-season events, alongside MAI-first-party positioning and the Code of Conduct consultation (which closes ~late October 2026).
• Watch items: (a) named Grok version in Office + safety documentation; (b) data-training/no-training commitment for M365 Copilot usage; (c) first enterprise data-incident or regulator questions around off-Microsoft processing; (d) whether OpenAI responds with new exclusivity/partnership terms; (e) SpaceXAI IPO-era enterprise wins attributable to this channel.

The headline — "Grok is now in Copilot" — is real but small in the short term (a gated, Frontier-only, EU/UK-excluded preview in three Office apps). The structural story is large: Microsoft has turned Copilot into a multi-model marketplace with a governance switch, and xAI's Grok just walked through the Anthropic-sized door. The practical lesson for enterprises is not "should we enable Grok?" but "do we have a model-governance policy at all?" — because the picker, the subprocessor list, and the off-Microsoft data path are now the default vocabulary of office AI. Microsoft's genius and its risk are the same move: it sells everyone else's models, curates them, and keeps the data — and now it also has to defend a Grok whose alignment its own evaluations once flagged. Watch the version number, the DPA, and the EU/UK unblock: those three details will tell you whether this is a genuine multi-model era or a very well-governed experiment.

Research window: 2026-09-10 → 2026-09-17 (RESEARCH_CONFIG.json). Event date 2026-09-12 verified in-window. Evidence discipline: FACT / COMPANY CLAIM / INDEPENDENT EVIDENCE / INTERPRETATION / PREDICTION kept distinct throughout; all web sources persisted in sources/S42.md with complete absolute URLs.