Microsoft on Monday introduced Copilot Cowork, a new AI agent layer for Microsoft 365 that moves beyond simple chat interactions to execute multi-step tasks across email, meetings, files, and messaging apps. The product, announced by Charles Lamanna and built with Anthropic’s Claude models running as a formal subprocessor, represents a deliberate bet that enterprise AI should act on behalf of workers rather than just answer their questions. For businesses already paying for Microsoft 365 Copilot licenses, the shift raises a practical question: what changes when an AI assistant can draft plans, run them in the background, and check in with humans only at key decision points?
From Chat Assistant to Autonomous Agent
Most AI tools embedded in office software still operate as glorified search bars. A user types a prompt, gets a response, and then manually carries out whatever the answer suggests. Copilot Cowork breaks that pattern. According to Microsoft’s official announcement, the system takes actions across Microsoft 365 rather than simply generating text. It grounds those actions in user context drawn from emails, meetings, messages, files, and organizational data, then runs plans in the background with built-in checkpoints and human oversight.
That design choice matters because it shifts the trust model. Instead of a user reviewing every AI output before acting on it, Copilot Cowork operates semi-autonomously and surfaces its work at defined intervals. The human stays in the loop, but the loop is wider. Jared Spataro, who leads Microsoft’s AI-at-Work efforts, framed the transparency angle directly in comments to Reuters reporting, emphasizing that customers can see what information the system has access to and how it is used when acting on behalf of the user.
The distinction between “chat” and “agent” is not just branding. A chat tool generates suggestions. An agent executes workflows. If Copilot Cowork can reliably schedule meetings based on email threads, draft follow-up documents from meeting notes, and compile data across SharePoint files without constant prompting, it changes the daily rhythm of knowledge work. The risk, of course, is that autonomous action without tight guardrails can produce errors at scale, which is why the checkpoint model will face scrutiny from IT departments evaluating adoption. Organizations will need to decide which tasks are safe to delegate and where they still want human review of every step.
Why Anthropic, and Why as a Subprocessor
Microsoft’s decision to integrate Anthropic’s Claude models rather than relying solely on OpenAI is the most strategically interesting element of this announcement. For years, Microsoft’s AI story centered on its exclusive partnership with OpenAI. Bringing Anthropic into the 365 ecosystem signals a move toward model diversity, where different AI engines handle different tasks based on their strengths and safety profiles.
The legal and compliance architecture tells the real story. Anthropic has been onboarded as a formal subprocessor for Microsoft Online Services, a classification that brings Claude under Microsoft’s product terms and Data Protection Addendum. This replaces an earlier arrangement where customers who wanted to use Anthropic models had to opt in under Anthropic’s separate commercial terms. The shift means enterprise data protection now applies uniformly, and IT administrators do not need to manage a second set of vendor agreements to use Claude within 365.
For enterprise buyers, this is a meaningful simplification. Data governance teams typically resist adding new AI vendors because each one introduces a separate compliance surface. By absorbing Anthropic into its existing contractual framework, Microsoft removes that friction. The trade-off is that customers now depend on Microsoft’s oversight of Anthropic’s data handling rather than negotiating directly with the AI company. Whether that concentration of trust is a net positive depends on how much confidence an organization places in Microsoft’s enforcement of its own terms and its ability to audit subprocessors effectively.
There is also a competitive dimension. By making Anthropic a subprocessor, Microsoft can present Claude as a first-class option inside its cloud and productivity stack without asking customers to think of it as a separate vendor relationship. That keeps the center of gravity inside Microsoft 365 even as the underlying AI layer becomes more heterogeneous.
Claude Powers the Researcher Agent
The first visible integration point for Anthropic’s technology is the Researcher agent inside Microsoft 365 Copilot. According to Microsoft’s support guidance, users with an active Microsoft 365 Copilot license can select Claude models within Researcher through a built-in model selection flow. The feature requires admin enablement before end users can access it, and rollout timing varies by organization and region.
The admin-gating is deliberate. Rather than flipping a switch for every 365 tenant, Microsoft is letting IT teams decide when and whether to expose Claude as an option. This gives organizations time to evaluate the subprocessor terms, test the integration in controlled environments, and set internal policies before employees start routing sensitive queries through a new model. For smaller companies without dedicated AI governance staff, the default-off posture means Claude access may lag behind larger enterprises that have the resources to evaluate and approve new tools quickly.
The Researcher agent itself is designed for deep information gathering, pulling from web sources and internal organizational data to produce detailed analysis. Adding Claude as an alternative model gives users a choice in how that research gets conducted. Different models have different strengths in reasoning, summarization, and handling ambiguity, so the ability to switch between engines could produce noticeably different outputs depending on the task. Over time, usage patterns may reveal where Claude is preferred, for example, in synthesizing long-form reports, while other models remain the default for more transactional prompts.
What This Means for the Multi-Model Enterprise
The broader implication of Copilot Cowork is that Microsoft is building a platform where multiple AI providers compete for different parts of the workflow. OpenAI models still power the core Copilot experience, but Anthropic now handles specific agent functions, starting with research and potentially extending to other complex workflows. This hybrid approach has practical advantages: if one model struggles with a particular type of reasoning or generates unreliable outputs in a specific domain, the platform can route that work to a better-suited engine.
That flexibility also creates complexity. IT administrators now need to understand which models are active, what data each model can access, and how those models behave under different conditions. Documentation around subprocessors, data residency, and logging becomes central to deployment decisions. Security teams will want clear answers about whether model selection changes where data is processed, how long it is retained, and what protections apply when agents act autonomously across multiple applications.
For end users, the multi-model design may be largely invisible at first. They will see Copilot Cowork as a single assistant that can take on more of their workload, not a collection of interchangeable models. But as interfaces like the Researcher agent expose explicit model choices, workers may begin to develop preferences and informal guidelines, using one model for brainstorming, another for policy-sensitive tasks, and relying on Cowork’s automation layer to orchestrate the rest.
In the near term, the success of Copilot Cowork will depend on how well Microsoft balances autonomy with control. Enterprises will be watching for clear configuration options: the ability to limit which apps agents can touch, to define approval workflows for high-impact actions, and to monitor what the system is doing on behalf of users. The promise of AI that actually does the work, rather than just talking about it, is compelling. But in regulated industries and large organizations, that promise will only be realized if the underlying agent framework, model choices, and subprocessor arrangements can withstand legal, security, and compliance scrutiny.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
