Microsoft pulls Anthropic’s Claude into Copilot as AI “agents” move from chat to doing the work

Microsoft is widening its AI bench inside Microsoft 365, bringing in Anthropic’s Claude models to power a new “agent” mode for Copilot, software that’s supposed to execute multi-step tasks, not just spit out answers.

The move is as much about product muscle as it is about optics. AI agents have rattled the software market since mid-January, raising fears that autonomous tools could replace chunks of traditional SaaS. Microsoft wants to reassure corporate customers with tighter controls, and signal to Wall Street it isn’t chained to a single AI model provider.

Anthropic’s Claude Sonnet models will now be available to some Microsoft 365 Copilot users, reducing Copilot’s perceived dependence on OpenAI and giving Microsoft more flexibility to route different jobs to different models.

Copilot’s new “Cowork” mode aims to execute, not just respond

Microsoft’s pitch for Copilot Cowork is straightforward: stop asking AI to draft a single email and start delegating an entire workflow. Think “prep me for a client meeting,” then the agent pulls relevant files, finds key numbers, assembles a slide deck, organizes documents, messages the team, and proposes a schedule.

This is designed for longer, multi-step tasks that can run in the background while you do other work. Microsoft says the agent can handle things like building a simple app, generating and cleaning up a spreadsheet, sorting a messy folder of files, and producing a usable summary.

The key word is autonomy, but not autonomy without guardrails. Microsoft says the agent will pause for approvals at critical moments to prevent the digital equivalent of a disastrous mis-click.

In demos, the agent draws on the stuff that actually defines office work: your emails, meetings, files, and Teams messages. That context is the advantage. The value isn’t only the AI model, it’s that Microsoft 365 already knows what you’re working on, who you’re working with, and how packed your calendar is.

plenty of “agents” sold over the past year have mostly been glorified text generators that still leave humans doing the real work. Microsoft’s bet is that deep integration into Microsoft 365 will make Cowork feel like it truly acts, while staying on rails. Companies will judge it less on slick demos and more on what happens after weeks of real use.

Anthropic joins Microsoft 365, and OpenAI is no longer the only engine

Until now, the public story around Copilot has largely been “OpenAI under the hood.” By adding Anthropic, Microsoft is sending a clear message: it’s not married to one model, and it will use whatever performs best for a given task.

For CIOs and IT leaders, that looks like a multi-vendor strategy, useful for risk management and procurement leverage. For investors, it’s also a hedge against over-reliance on a partner that could influence pricing, product direction, or even access to scarce computing capacity.

On the ground, customer preferences vary. Some want the best model available at any moment. Others care more about avoiding lock-in to a single AI stack. A multi-model approach lets Microsoft position Copilot as a platform: writing, analysis, classification, planning, routed to the model that fits.

There’s also a competitive reality. As frontier AI labs push deeper into enterprise workflows, Microsoft has an incentive to bring them inside the tent rather than let them siphon off daily work from Outlook, Teams, Excel, and PowerPoint. The Anthropic partnership is Microsoft planting a flag: agentic AI should happen on Microsoft’s turf.

Pricing is still fuzzy, but the entry point is already $30 a user each month

Microsoft hasn’t published detailed pricing for Copilot Cowork. What it has said: some usage will be included with Microsoft 365 Copilot, which costs businesses $30 per user per month, with additional usage available for purchase.

That “base plus overage” model matters because agents can be far more compute-hungry than a simple chat. An agent that plans, executes, checks its work, and follows up may trigger multiple model calls, data lookups, and actions across apps. At scale, thousands of employees, this becomes a budget line IT will watch closely.

Expect companies to demand guardrails: quotas, reporting, and clear usage policies. If an agent is “working” all day in the background, costs can get confusing fast. With no public rate card yet, early customers will be the ones negotiating terms and testing whether the ROI is real.

And that $30 price tag is already controversial in some organizations as a per-seat add-on. Microsoft is betting agents make the value obvious. If an agent reliably saves 20 minutes a day, that’s tangible. If it forces employees to double-check everything, it starts to look less like an assistant and more like an intern who won’t stop talking.

Security and controls: Microsoft leans into its enterprise advantage

Microsoft is emphasizing security and data controls, because that’s where many companies draw the line. An agent that can search files, manipulate documents, and propose calendar changes is touching sensitive corporate assets.

The advantage of an agent built into Microsoft 365 is that access rules already exist: file permissions, Teams membership, internal compliance policies. In theory, the agent can’t read what you can’t read.

In practice, corporate permissions are often messy, and automation can amplify bad setups. Microsoft is also promising “limited but real” supervision, checkpoints, approvals, the ability to pause and edit. The tension is obvious: too many approvals kill automation; too few raise the risk of an agent emailing the wrong distribution list or quietly changing a shared file.

The real test will be whether companies can tune delegation by department. Finance teams won’t let an agent move numbers without oversight. Marketing might happily let it draft slides and then review. Microsoft’s credibility will hinge on logs, settings, and whether it can explain why an agent did what it did.

A staged rollout, jittery markets, and the fight for the future of office work

Copilot Cowork isn’t launching to everyone at once. Microsoft is running it as a research preview with select customers, with broader access expected through its Frontier program in late March 2026, an early-access track for experimental Microsoft 365 Copilot features.

All of this is happening in a tense market. Since Anthropic’s Claude Cowork-style agents hit the scene in January, investors have worried that agents could reduce demand for traditional SaaS tools. Microsoft itself has faced turbulence: the article cites shares down about 15% over the past year and points to a moment when disappointment around Azure growth wiped roughly $357 billion off Microsoft’s market value.

Copilot Cowork also functions as a narrative reset: Microsoft is investing, shifting resources, and trying to secure the next phase of productivity software. The company has to sell a paradox, agents will transform work without cannibalizing Microsoft’s own business, instead making Microsoft 365 even stickier.

The outcome will come down to habit. If agents become the default way people work inside Outlook, Teams, Excel, and PowerPoint, Microsoft tightens its grip on the modern office. If outside agents prove faster, smarter, or more actionable, companies could route work away from Microsoft’s ecosystem. For now, Microsoft is moving carefully: limited preview, deep integration, and a second major AI partner so it isn’t betting everything on one horse.