$13B into OpenAI, yet Microsoft Copilot Cowork runs on Anthropic Claude

Microsoft has spent $13 billion backing OpenAI. It built Copilot around GPT models. It made OpenAI the default brain of its entire productivity suite. And then, on March 9, 2026, it launched the most ambitious AI agent feature in Microsoft 365 history — powered by Anthropic’s Claude.

That decision alone tells you everything about where the enterprise AI market is heading. Vendor loyalty is dead. Performance is king. And Microsoft, the company most closely identified with OpenAI, just proved it.

What Microsoft Copilot Cowork Actually Does

Copilot Cowork turns Microsoft 365 Copilot from a chatbot into a genuine autonomous agent. Instead of answering one question at a time, Cowork can take a complex request — “Prepare me for the Anderson account review on Thursday” — and break it down into a multi-step workflow that runs for minutes or even hours.

That single request might trigger Cowork to pull the latest financials from a SharePoint folder, scan recent email threads in Outlook for context, draft a slide deck in PowerPoint, schedule a prep call in Teams, and send a briefing email to two colleagues. Each step is visible to the user, who can steer, pause, or redirect the workflow at any point.

The technical backbone comes from Anthropic’s Claude model for reasoning, plus the same “agentic harness” — the orchestration layer that lets an AI model use software tools within defined guardrails — that powers Anthropic’s own Claude Cowork product. But there is a critical architectural difference: Copilot Cowork runs entirely in the cloud within a customer’s Microsoft 365 tenant, not on a local device.

This cloud-first approach plugs into what Microsoft calls “Work IQ,” an intelligence layer drawn from a user’s emails, files, documents, meetings, and chats across the entire M365 ecosystem. That means Cowork has enterprise-wide context that a local agent simply cannot match.

Why Microsoft Chose Anthropic Over OpenAI

This is the question everyone is asking, and Microsoft’s CMO for AI at Work, Jared Spataro, gave the most revealing answer: “Every 60 days at least, there’s a new king of the hill.”

That single quote captures Microsoft’s strategic pivot from single-vendor dependency to a multi-model platform. Copilot now leverages leading models from both OpenAI and Anthropic, with Claude available in the full Copilot Chat experience — not just the Researcher and Excel features where it first appeared.

The timing matters. Anthropic launched Claude Cowork to consumers in mid-January 2026, and the product was an immediate hit. It showed that AI agents could go beyond conversation to actually execute real tasks — browsing, file management, app automation. Microsoft’s stock dropped 14% in the weeks that followed, as investors worried that standalone AI agents might undercut the value of traditional SaaS subscriptions.

Microsoft’s response was characteristically pragmatic: license the technology that threatened you, wrap it in enterprise-grade infrastructure, and ship it as a Copilot feature. As Forrester VP JP Gownder put it, this is “a strategic shift” — Microsoft moving Copilot away from reliance on OpenAI alone and toward a multi-model architecture that picks the best tool for each job.

Spataro acknowledged as much, describing Anthropic’s Claude Cowork as “a fantastic tool” while noting its “limitations” in corporate environments — specifically around local-only processing and security concerns at enterprise scale. Copilot Cowork is Microsoft’s answer: the same agentic intelligence, but running inside a managed, compliant cloud environment.

The Enterprise Data Advantage

The biggest difference between Copilot Cowork and every other AI agent on the market is data access. Because it operates inside a customer’s M365 tenant, Cowork can draw on the full Microsoft Graph — the unified data layer connecting Outlook, Teams, SharePoint, OneDrive, Excel, and more.

This is not a trivial advantage. A standalone AI agent like Claude Cowork can operate on files sitting on your local machine and interact with apps through screen automation. But it cannot reach into a corporate SharePoint site to pull last quarter’s sales deck, cross-reference it with a Teams conversation from two weeks ago, and then update an Excel workbook with new projections — all while respecting the organization’s access permissions and compliance policies.

Microsoft layers its existing security stack on top of Cowork: Entra for identity and access management, Defender for threat protection, Intune for device compliance, and Purview for data governance. For IT departments that have spent years building out their Microsoft security posture, this integration means they can deploy AI agents without starting a separate governance conversation.

The accompanying Agent 365 product ($15 per user per month, available May 1) serves as the control plane for managing all AI agents within an organization — not just Cowork, but any agent built on the platform. Microsoft says it already tracks over 500,000 agents registered during the preview period.

Pricing: The $30 Question

Copilot Cowork requires a Microsoft 365 Copilot license at $30 per user per month. That is the same price point Copilot has held since its launch, so for existing Copilot customers, Cowork arrives as an upgrade at no additional cost.

For organizations wanting the full stack, the new Microsoft 365 E7 bundle launches May 1 at $99 per user per month. That includes:

Microsoft 365 E5 (valued at $60)
Copilot with Cowork capabilities ($30)
Agent 365 ($15)
Entra Suite ($12)
Advanced security features via Defender, Intune, and Purview

Purchased separately, those components would run $117 per user per month, so the E7 bundle represents roughly a 15% discount. Microsoft is clearly betting that packaging AI capabilities with security and governance tools will accelerate enterprise adoption — and make it harder for standalone AI agent products to compete on the security front.

The adoption numbers suggest the bet is working. Copilot paid seats have grown 160% year-over-year. Daily active usage is up 10x. Customers deploying at 35,000+ seat scale have tripled year-over-year. And 90% of Fortune 500 companies already use some version of Copilot.

How Copilot Cowork Stacks Up Against the Competition

The enterprise AI agent market is getting crowded fast. Here is where Copilot Cowork sits relative to the main alternatives:

Anthropic Claude Cowork is the closest technological cousin — it shares the same reasoning engine and agentic harness. But Claude Cowork runs locally on a user’s desktop, interacting with apps through screen-level automation. It is better suited for individual power users and creative workflows. Copilot Cowork is built for enterprise teams that need cloud-based execution, shared data access, and centralized governance.

Google Workspace with Gemini has been integrating AI across Docs, Sheets, and Gmail, but Google has not yet shipped a comparable autonomous agent that can execute multi-step workflows across its entire suite. Google’s AI features remain largely prompt-and-response within individual apps.

Salesforce Agentforce targets a different slice of the enterprise — CRM-centric workflows like lead qualification, case resolution, and sales coaching. It is deeply integrated with Salesforce data but does not span general productivity tasks the way Copilot Cowork does.

Notion AI Agents and other standalone tools offer lightweight automation for smaller teams, but they lack the deep enterprise data integration and security infrastructure that large organizations require.

The competitive moat Microsoft is building is not about the AI model itself — it already swapped from OpenAI to Anthropic for this feature, proving models are interchangeable. The moat is the data layer: Work IQ, the Microsoft Graph, and the security stack that wraps around it all.

What Happens Next

Copilot Cowork is currently in Research Preview with a limited set of customers. Broader access through Microsoft’s Frontier program is expected in late March 2026, with general availability likely following over the subsequent months.

The bigger question is what this means for the Microsoft-OpenAI relationship. Microsoft is not abandoning OpenAI — GPT models still power most of Copilot’s features. But by building its most visible new capability on Anthropic’s technology, Microsoft has sent an unmistakable signal: no single AI provider gets a lock on the enterprise.

For enterprise buyers, that is probably good news. A multi-model Copilot that picks the best engine for each task is more resilient and more competitive than one tied to a single vendor’s roadmap. For OpenAI and Anthropic alike, it means the real battleground is not consumer mindshare — it is which model gets selected for the next Microsoft feature.

FAQ

How much does Microsoft Copilot Cowork cost?
Copilot Cowork is included with the Microsoft 365 Copilot license at $30 per user per month. It is also available as part of the new M365 E7 bundle at $99 per user per month, which adds Agent 365, Entra Suite, and advanced security tools.

When will Copilot Cowork be available?
It is currently in Research Preview for select Microsoft Frontier program customers. Broader Frontier access is expected in late March 2026, with general availability to follow.

How is Copilot Cowork different from Anthropic’s Claude Cowork?
Both use Anthropic’s Claude model and the same agentic harness. The key difference is deployment: Claude Cowork runs locally on a user’s device, while Copilot Cowork runs in the cloud inside a customer’s M365 tenant with access to enterprise data (email, calendar, files, chats) and Microsoft’s full security stack.

Does this mean Microsoft is leaving OpenAI?
No. OpenAI’s GPT models still power the majority of Copilot features. Copilot Cowork represents Microsoft’s shift to a multi-model strategy where it selects the best model for each capability, rather than relying on a single provider.

What tasks can Copilot Cowork handle?
Cowork can execute multi-step workflows across Microsoft 365 — preparing meeting materials, drafting presentations, sending scheduled emails, rescheduling meetings, coordinating projects, and more. It breaks complex requests into visible steps that users can monitor and adjust in real time.

Top AI Product