Microsoft 365 Copilot in September 2026: What Changed for Enterprise AI

Microsoft’s enterprise AI expansion was less a single September product launch than a continuing shift in Microsoft 365 Copilot toward agentic, multi-model, metered work. The practical changes for organizations were clearer model options, tools for orchestrating multi-step agent workflows, and more explicit controls over consumption.

Copilot Cowork moved agentic work into general availability

Microsoft announced general availability for Copilot Cowork on June 16, 2026. Cowork requires a Microsoft 365 Copilot User Subscription License and is billed according to the tasks users run. Its task charges are calculated from model use, context retrieval, tool calls, and runtime, and are denominated in Copilot Credits. Customers can use pay-as-you-go billing or commit to P3 usage volume for a discount. 7

At general availability, Cowork used Anthropic’s Opus 4.8 and Sonnet 4.6 models. 7 By July, Microsoft said Claude Opus 5 was rolling out to the model selector across Word, Excel, PowerPoint, Copilot Chat, Copilot Cowork, and Copilot Studio. 28 Microsoft’s roadmap also listed GPT-5.6 as available in Microsoft 365 Copilot. 26

The implication is significant: model selection became part of the enterprise Copilot operating model. Teams could increasingly match a model to a task while keeping the work inside Microsoft 365’s productivity and administration environment.

What “Copilot Orchestrator” actually means

The evidence supports orchestration as a core platform capability, but not a separately confirmed, no-code product called Copilot Orchestrator.

Microsoft describes the Microsoft 365 Copilot orchestrator as the layer that selects and runs the appropriate skills and actions between a user request and Copilot’s response. 43 In Copilot Studio, generative orchestration uses an LLM-driven planning layer to interpret intent, break down complex requests, choose tools and knowledge sources, and execute multistep plans with safety and compliance guardrails. 33

For more complex designs, Microsoft documents an orchestrator-and-subagent pattern: a primary agent delegates work to specialist child or connected agents, then brings their results together. 45 A Microsoft demonstration of multi-agent orchestration similarly described a central agent handling intent classification, routing, and synthesis. 38

In plain terms, the platform can enable a maker to build a workflow where one agent receives a business request, routes portions to domain-specific agents, invokes appropriate tools, and returns a unified response. That is more capable than a single Q&A bot, but organizations should still distinguish documented platform features from any claimed standalone product branding.

Cost management became a first-class administrator concern

Agentic systems can vary in cost by model, context size, tools, and runtime. Microsoft’s Cowork documentation addresses that operational reality: administrators can view usage in the Microsoft 365 admin center and set limits for individual users or groups. 10

That makes adoption a governance exercise as well as a productivity decision. Before broad rollout, IT and finance teams should:

• define which workflows are appropriate for metered agentic execution;
• assign ownership for usage monitoring and spending limits;
• test task quality, latency, and cost together rather than evaluating model output alone; and
• apply permissions and approval steps to high-impact actions.

Copilot Studio’s orchestration documentation emphasizes guardrails for safety and compliance, while its agent-building guidance distinguishes between declarative agents that use Microsoft’s orchestration and models, and custom-engine agents where organizations provide their own. 3336

Copilot Studio’s broader role

Copilot Studio is the layer for organizations that need to go beyond built-in Microsoft 365 experiences. Microsoft positions it for broad or external publishing, multistep logic, approvals, branching workflows, advanced AI models, Azure AI integrations, connectors to data outside Microsoft 365, autonomous capabilities, and lifecycle management. 37

Microsoft also said that a newer orchestration layer in Copilot Studio improved agent execution in its evaluations, reporting approximately 20% better evaluation performance and 50% lower net token usage. Those are Microsoft-reported measurements, not an independent benchmark, but they show the company’s focus on making complex agent workflows more reliable and economical. 40

The enterprise takeaway

The clearest September-era message was not simply that Copilot had added another model or chatbot feature. Microsoft was building an enterprise AI stack around four connected ideas:

1. Agentic execution: Cowork and Copilot Studio support work that can involve planning, tools, organizational context, and multiple steps. 733
2. Model flexibility: GPT-5.6 and Claude Opus 5 were presented as available options within Microsoft 365 Copilot experiences. 2628
3. Orchestration: Microsoft’s architecture supports routing work among skills, tools, and specialized agents rather than handling every request in one model interaction. 4345
4. Governance and spend control: Usage visibility and user- or group-level limits are essential when AI work is billed by consumption. 10

For buyers, the important question is therefore not only which Copilot feature is available. It is whether the organization has a repeatable way to choose models, connect trusted data and actions, supervise agent behavior, and manage variable usage costs.