Microsoft turned its Copilot assistant into a multi-model, agent-driven work platform in July 2026, shipping more than 40 updates across Microsoft 365. The two standout changes: users can now select Anthropic’s Claude model directly inside Copilot Chat, and Copilot Cowork—an AI agent that can independently execute complex, multi-step tasks—reached general availability.

These aren’t just feature drops. They signal that Microsoft is betting on a strategy where Copilot becomes the orchestrator for multiple AI models, not the exclusive face of one. For the average user, that means more choice. For IT administrators, it means new governance challenges. And for small businesses, it opens a more affordable path to enterprise-grade AI.

What Actually Changed

Microsoft’s July release notes read like a product overhaul. Here’s what landed:

Claude Arrives in Copilot Chat

As first announced in Microsoft’s tech community blog on June 16, Anthropic’s Claude is now a selectable model in Microsoft 365 Copilot Chat. Users can pick Claude for tasks that demand deep document analysis, multi-page synthesis, or structured content generation, then switch back to Microsoft’s default models for quick queries. Claude Sonnet 5 joined the lineup on July 2, extending the option to Copilot Cowork and PowerPoint.

The model picker appears directly in the chat interface. You aren’t locked into one AI at account creation. You choose per task, right when you need it.

Copilot Cowork Moves from Assistant to Doer

Copilot Cowork, which exited preview and became generally available on June 16, is designed for long-running, autonomous work. It can plan a multi-step project, pull documents from SharePoint, analyze meeting transcripts, invoke plugins, and produce a finished deliverable—all while you’re away from your desk. Tasks run in the cloud, so your PC can be closed.

Microsoft describes a typical flow: you state an outcome, Cowork builds a plan, retrieves permitted data through Microsoft’s Work IQ context layer, executes steps, and logs everything. You can review, intervene, or cancel at any moment. Completed files remain subject to your organization’s retention, sensitivity labels, and sharing controls.

AI Watermarks Come to Media

An administrative policy can now apply visual or audio watermarks to video and audio content generated or modified by AI tools within Microsoft 365. Available across Windows, macOS, iOS, Android, and the web, the watermark signals that media isn’t entirely human-made. Microsoft calls it a transparency feature, not a tamper-proof seal.

Agent Builder Gets a Storefront

Employees who create custom agents using Agent Builder can now submit them for publication in the “Built by your org” section of the Agent Store. An admin reviews and approves the submission before colleagues can discover it. Lifecycle policies can also automatically reassign orphaned agents when an owner leaves the company, preventing abandoned automations from lingering.

Document Navigation Gets Smarter

PDFs can open directly inside Copilot Chat instead of forcing a separate app. Word’s Copilot can now make direct edits by default (still reversible), and generated content lands in a centralized Library. The Copilot app itself has a smoother transition between Chat and Cowork, so you can start with a conversation and hand off execution without switching interfaces.

Business Bundles Simplify Licensing

On July 1, Microsoft launched permanent Microsoft 365 Business Standard with Copilot and Business Premium with Copilot subscriptions. These combine the familiar Office apps, cloud storage, and security tools with Copilot access in a single SKU. The goal is to reduce procurement friction for small and midsize companies that might otherwise turn to unmanaged consumer AI accounts.

What It Means for You

For Everyday Users

The model picker is your most immediate win. Got a 50-page contract to review? Switch to Claude. Need a quick summary of your inbox? Go with the default Copilot model. You’ll need to experiment to learn which model handles which tasks best; Microsoft doesn’t yet offer one-click “just get it done” routing. But this flexibility can genuinely improve output quality if you’re willing to invest a little time.

Cowork is a bigger leap. It can now act like a junior teammate: prepare a project timeline, compile research from scattered files, or draft a presentation based on email threads. But treat its output like a first draft. Review everything, especially data pulled from multiple sources. The agent can misread context or surface information you didn’t intend to include.

Watermarks are a welcome transparency step. If you generate a training video with an AI voiceover, it’ll be labeled. That helps colleagues and clients know the material’s origin. However, the watermark can be cropped or stripped, so don’t rely on it for security-sensitive content.

For IT Administrators

Your job just got more complex. Claude and Cowork are powerful, but they introduce new governance vectors:

  • Model access: You can now control which models appear in users’ pickers. Consider whether to allow all frontier models or restrict to Microsoft’s defaults until you’ve assessed security and compliance implications.
  • Agent spending: Cowork uses a consumption-based model billed in Copilot Credits (roughly one cent each on pay-as-you-go). An agent that runs for hours, invokes countless plugins, and pulls large documents can rack up costs fast. Set spending limits at tenant, group, and user levels, and configure alerts for thresholds.
  • Permission hygiene: Agents operate within the user’s existing Microsoft 365 permissions. That means if your SharePoint sites are overshared, an agent can find and summarize content you’d rather keep siloed. Audit permissions before deploying Cowork broadly.
  • Agent lifecycle: The new publication workflow is a help, but you’ll need a review process for custom agents—check their instructions, data sources, connectors, and owner. Use the new reassignment policies to prevent orphaned agents from running indefinitely.
  • Data governance: Microsoft says Cowork prompts, responses, and outputs stay within the Microsoft 365 trust boundary, subject to eDiscovery, retention, and sensitivity labels. Verify the exact state of each compliance feature in your tenant. Don’t assume it’s all on by default.

For Small Businesses

The new Business bundles are the most straightforward entry point. For a company with 10 employees, buying Microsoft 365 Business Standard with Copilot is much simpler than acquiring Copilot separately. Business Premium adds device management, advanced security, and information protection—handy if you handle client data.

However, don’t skip the setup. Even a small team can create permissions chaos. Before letting everyone loose with Copilot, do a quick audit: who can access what in SharePoint and Teams? Remove stale external sharing links. Decide what kind of data should never be fed into an AI prompt. Then roll out gradually, starting with a pilot group that tests real workflows and reports back.

How We Got Here

When Microsoft 365 Copilot launched in 2023, it was essentially a wrapper around OpenAI’s GPT-4, grounded in organizational data. The value proposition was integration with Word, Excel, Outlook, and Teams. It answered questions and generated text.

Over the next three years, Copilot evolved. Microsoft built Work IQ, a context layer that lets any underlying model tap into a user’s files, emails, calendar, and people graph, all governed by existing permissions. It added Agent Builder for lightweight custom assistants. And it began testing third‑party model integration with the Frontier preview program earlier in 2026.

The July 2026 wave consolidates those threads. The addition of Claude is the clearest signal yet that Microsoft views Copilot as an orchestration platform, not a single-model chatbot. This mirrors the company’s Azure strategy, where supporting Linux and competing databases only strengthened Microsoft’s cloud position. By welcoming a rival AI, Microsoft aims to make Microsoft 365 the universal work hub—even if another company built the brain that handles your request.

What to Do Now

If You’re an End User

  1. Test the model picker. Next time you need a detailed analysis, select Claude. Compare its output with the default model. Over time, you’ll develop an intuition for which assistant fits which task.
  2. Start Cowork small. Don’t kick off a 10-step agentic process on a critical project. Ask it to compile a meeting summary from Teams chat and a shared document. Verify carefully.
  3. Check for watermarks. If you generate AI media, confirm whether your organization has enabled the watermark policy. Even if not, consider manually noting AI involvement in your deliverables.

If You’re an IT Admin

  1. Review model availability. In the Microsoft 365 admin center, decide which models users can access. You might start with just the default Copilot model and Claude, then expand after evaluation.
  2. Set up Cowork budgets immediately. Even if you delay deployment, configure tenant spending limits, per-user caps, and alert rules. A single forgotten task shouldn’t blow your AI budget.
  3. Audit file permissions. Run a tool like Microsoft Purview’s Content explorer to identify broadly shared sensitive files. Tighten access before agents can automatically surface them.
  4. Establish an agent review process. Designate who can submit agents to the store and who approves them. Include checks for data sources, connector permissions, and prompt instructions.
  5. Pilot with a controlled group. Roll out Cowork to a handful of trusted users, monitor their tasks and credit consumption, then expand based on real data—not vendor benchmarks.

If You’re a Small Business Owner

  1. Pick the right bundle. If you need only productivity apps, the Business Standard with Copilot is enough. If you want security, device management, and information protection, go with Business Premium with Copilot.
  2. Do a quick permission clean-up. Remove “Anyone with the link” shares on sensitive folders. Quietly delete outdated Teams channels that might contain client data.
  3. Train your team. Show employees how to use the model picker, how to validate Cowork’s output, and why they should never paste customer PII into a chat.
  4. Monitor usage in the admin center. Copilot usage reports will show who is using Cowork and how many credits they’re burning. Keep an eye on that dashboard weekly.

What’s Next: The Outlook

Microsoft’s immediate challenge is to make model choice invisible. Most users don’t care about the name behind the AI; they want the best result. Expect automatic routing to become a headline feature within the next year—Copilot would pick Claude for document analysis, GPT for creative writing, a smaller model for quick factual lookups, all without user intervention. That would reduce cognitive load and cut costs.

Keep an eye on “Cowork 1,” a Microsoft-developed model designed for routine agent tasks. If it offers similar quality at a lower price than Claude Opus or GPT, it could become the cost-efficient backbone of enterprise automation. This would also signal Microsoft reducing its dependence on external AI labs.

Browser-based agent actions are another frontier. Cowork can already interact with web pages via Edge, subject to enterprise policies. Once that capability matures, expect agents to book travel, update CRM records, and submit forms—with all the security, legal, and spend complexities that implies.

Finally, the industry will watch for hard ROI data. Early adopters will report whether Cowork actually saves time, reduces errors, or lets teams complete work they previously couldn’t. Those numbers, not demos, will determine whether the multi-model agent platform becomes the new normal for office work.