Microsoft just borrowed Anthropic’s brain for its most ambitious AI bet yet.
Copilot Cowork, announced on March 9, uses Claude’s reasoning engine to handle autonomous multi-step tasks that can run for hours. Prepare a customer presentation, pull financials, email the team, schedule prep time—all from one request. This isn’t chatbot-style back-and-forth. It’s AI that actually does the work while you do something else.
The move comes after Microsoft’s stock dropped 14% following Anthropic’s January launch of Claude Cowork. Apparently, investors were worried Microsoft was getting left behind in the agentic AI race.
How It Works
Copilot Cowork breaks complex requests into discrete steps, reasons across tools and files, and executes actions across Microsoft 365 apps. Tasks aren’t confined to a single turn or single app anymore—they can coordinate across Outlook, Teams, Word, Excel, and other M365 applications.
Microsoft calls it “the execution layer for Microsoft 365.” Users can delegate tasks, coordinate workflows, and supposedly stay in control. The system promises transparency: work is observable, actions are visible, and progress can be reviewed, guided, or stopped.
The key difference from Anthropic’s Claude Cowork: Microsoft’s version runs in the cloud within a customer’s Microsoft 365 tenant, not locally on the device. Microsoft explicitly positioned this as safer for enterprise, arguing that local deployment “raises security concerns when deployed at scale.”
What It Integrates With
Cowork taps into what Microsoft calls “Work IQ”—intelligence pulled from emails, files, documents, meetings, and chats within your organization. When you ask Cowork to prepare for a customer meeting, it’s not just pulling from the internet. It’s searching through your actual work context.
Demonstrated capabilities include:
- Inbox triage: Sorting and responding to emails based on priority
- Meeting preparation: Assembling presentations and research materials
- Workflow coordination: Chaining actions across multiple M365 apps
- Team communication: Drafting and sending messages based on context
The Price Tag
Microsoft is rolling out Cowork as part of “Wave 3”—a new version of Microsoft 365 Copilot that moves beyond simple assistance into embedded agentic capabilities.
The pricing structure:
Copilot Cowork: $30/user/month (requires existing Copilot license)
Microsoft 365 E7 “Frontier Worker” Bundle: $99/user/month (launching May 1)
- Microsoft 365 E5 ($60)
- Copilot ($30)
- Agent 365 governance platform ($15)
- Entra Suite and advanced security tools
If you bought everything separately, you’d pay $117. The bundle saves $18, assuming you need all the pieces.
Agent 365 ($15/month standalone, also May 1) is a new governance layer that lets IT teams monitor and secure AI agents across the organization using existing infrastructure like Entra, Defender, Purview, and Intune.
The Enterprise Concerns
Here’s where it gets complicated.
Microsoft is betting that enterprise customers will trust cloud-based AI agents with full access to their sensitive data. The company emphasizes that everything operates “within Microsoft’s security, identity, and governance framework.”
But analysts have raised concerns about governance, security, and data control. When an AI agent autonomously accesses emails, edits documents, and sends communications, the attack surface expands significantly. Who’s responsible when Cowork sends the wrong email to the wrong client?
The transparency promises are good in theory—observable work, visible actions, stoppable processes. But in practice, most workers aren’t going to watch their AI agent’s every move. The whole point is to delegate and do something else.
Regulated industries face particular challenges. Healthcare organizations, financial services, and government contractors have strict data handling requirements. An autonomous agent reading and processing sensitive documents needs to comply with HIPAA, SOX, or FedRAMP. Microsoft says Cowork respects existing compliance configurations, but enterprise IT teams will need to validate that claim extensively before deployment.
What This Means
Microsoft’s partnership with Anthropic for Cowork is notable. They’re not using their own Azure OpenAI models for the core reasoning—they went to Anthropic for “the agentic harness” that powers Claude Cowork. That’s a tacit admission that Anthropic’s approach to agent safety and reliability is worth licensing, even as a competitor.
The $99/month E7 bundle positions Microsoft to capture high-value enterprise customers who want the full agent stack with governance included. For organizations already paying for E5 licenses, the upgrade path is clear.
But this also raises the stakes on getting AI agents right. When autonomous AI can send emails, edit documents, and coordinate workflows, mistakes aren’t hypothetical. They’re visible. They’re email threads with incorrect information. They’re presentations with wrong numbers. They’re calendar invites that shouldn’t have been sent.
Cowork is currently in limited research preview. General availability through the “Frontier” program begins in March, with the E7 bundle and Agent 365 launching May 1.
What You Should Do
If you’re considering adoption:
- Wait for the research preview results before committing
- Audit what data Cowork would access in your M365 tenant
- Review your compliance requirements against Microsoft’s documentation
- Start with low-risk use cases (meeting prep, email triage) before high-stakes tasks
- Establish clear policies on what agents can do autonomously vs. what requires human approval
If you’re on the fence: The $30/month Copilot add-on without Cowork remains available. You can observe how early adopters fare before committing to the agent tier.
If you’re in a regulated industry: Talk to Microsoft directly about compliance certifications for Cowork. General marketing materials won’t cover your specific requirements.
The AI agent era for enterprise just got real. Microsoft is betting that enterprises will trust autonomous AI with their most sensitive workflows. Whether that trust is warranted remains to be seen.