Microsoft is turning Copilot into a central work interface for chat, internal app creation, and independently running agents. The new structure affects occasional users as well as companies that use Artificial Intelligence (AI) in Teams, Outlook, and Office. The key change is financial: Microsoft no longer intends to cover every resource-intensive capability through one flat license.
What Copilot now brings together
The redesigned Copilot app is divided into Home, Code, and Autopilot. Home combines the existing chat experience with Cowork, an environment for collaborating on tasks. Word, Excel, and PowerPoint are embedded directly, allowing you to work on documents without leaving the central interface.
Microsoft also plans to add Today, a personalized dashboard for important emails, calendar events, Teams conversations, and tasks. It is scheduled to enter private preview in October. This may sound less dramatic than an autonomous agent, but it could become the part that many people encounter most often in daily work.
Code is explicitly aimed at more than professional developers. You are supposed to be able to describe a tracker, dashboard, automation, or internal application in natural language and share the result with colleagues. It uses technology related to GitHub Copilot and, according to Microsoft, runs in an isolated managed environment within the organization’s own tenant, meaning its controlled Microsoft cloud space.
Autopilot is the new name for the agent previously introduced as Scout. An AI agent is a system that pursues a task across multiple steps, uses tools, and responds to intermediate results. Autopilot receives its own cloud computer, memory, workspace, and digital identity, and it can continue working while you are offline.
One concrete example is monitoring selected Teams channels: Autopilot can follow new activity and handle the resulting follow-up work. Microsoft also cites coordinating supplier reviews and completing recurring tasks. You can address the agent through an @ mention in Teams, Outlook, or documents.
Why billing becomes a defining issue
Microsoft is drawing a clearer line between ordinary assistance and longer-running autonomous work. According to the lead report, the standard Copilot license continues to cover chat and integrations with Office applications. Autopilot, Code, and Cowork will instead use consumption-based billing, similar to ChatGPT Enterprise once an included allowance has been exhausted.
The supplied sources do not provide exact prices or consumption limits. It is therefore not yet possible to calculate what a team running 20 agents regularly would actually pay. What is clear is that the per-person license will reveal less about total spending because the frequency, duration, and scope of delegated tasks will also matter.
For Office requests, an automatic model router is intended to select a suitable AI model for each task. A router distributes requests based on factors such as accuracy, speed, and cost. You may therefore not receive the same model for every task, while information technology administrators can decide which model families are available to particular user groups.
This optimization makes economic sense, but it creates a trade-off. More efficient models can be faster and less expensive, yet the lead report says they may not match the quality of leading frontier models, meaning the most capable systems currently available. That may be sufficient for summarizing a document, while model selection can matter more for complex analysis or a business-critical automation.
A study described by Nvidia researchers illustrates how much technical control can affect the bill. According to the researchers, they reduced a coding agent’s token use by almost half while keeping performance nearly unchanged. Tokens are small units of text that many AI services use to measure processing and cost.
The researchers optimized the harness rather than the underlying model. A harness is the control layer that manages context, tools, feedback, and stopping rules. The system examined 152 approaches across 535 executable environments and completed more than 3,000 runs; this does not prove that Microsoft will achieve the same savings, but it shows how strongly agent costs can depend on orchestration.
Rights and limits become part of the product
For enterprise agents, coding quality is only one part of the decision. A review of several coding-agent contracts compares GitHub Copilot, AWS Kiro, Cursor, Devin, and Windsurf on indemnity, data storage, audit logs, and costs for 500 seats. However, the available summary does not include complete pricing figures or the detailed data-residency findings.
According to that comparison, GitHub Copilot and Kiro provide uncapped indemnity for generated code. Indemnity is a contractual commitment to cover certain claims, including some involving intellectual property (IP), under defined conditions. Cognition’s standard terms reportedly exclude outputs entirely, although businesses still need to examine the applicable contracts for themselves.
Permissions are also becoming a product capability rather than a background setting. Microsoft says Autopilot has its own identity and includes access controls, audit features, and governance, meaning organizational rules for controlled use. Those safeguards matter more when an agent monitors channels, accesses files, or continues a process without constant human supervision.
The shift toward cloud-based agents also puts Microsoft’s earlier focus on specialized AI PCs into perspective. The company is no longer prominently using the Copilot+ PC label for new Surface devices, even though they satisfy the previous requirements. The retreat from Copilot+ PC branding also follows continuing privacy concerns surrounding the Recall feature.
The sources do not provide specific details about pricing, language support, availability, or Swiss data locations. Organizations in Switzerland therefore cannot yet infer where their content will be processed or how existing privacy requirements will be reflected in contracts. For Autopilot in particular, data residency, identity management, and the limits of audit records may be just as relevant as answer quality.
Pros and Cons of the Copilot super app
Pros:
- One interface – Chat, Office documents, internal applications, and agents are brought closer together.
- Practical automation – Autopilot can continue recurring work even when you are offline.
- Controlled access – Separate identities, permissions, and audit records can support organizational oversight.
- Lower entry barrier – Code is designed to let subject-matter specialists create simple internal tools without traditional programming experience.
Cons:
- Less predictable spending – Consumption-based billing transfers part of the financial risk to the customer.
- More contract questions – Liability, output rights, prompt storage, and data locations differ between providers.
- Unclear model selection – An automatic router can reduce costs while making the underlying model quality less predictable.
- Larger impact from mistakes – A continuously running agent can repeat an incorrect step more often and for longer than a single chat session.
What this means for your team
For beginners: A sensible first step is a tightly limited task without critical write permissions. A team could test Today for reviewing emails and meetings or use Code to draft a simple internal tracker. During the trial, you can observe which data is used, who verifies the result, and how much consumption the task generates.
For advanced users: Separate roles, spending limits, and verifiable test cases can deliver more value with better control. Administrators can release model families by user group and determine when an agent needs human approval. Comparing audit logs, data residency, prompt storage, and indemnity is also more informative than looking only at the advertised license price.
OpenAI provides a related business example involving fleet-management company Proaction. According to the Proaction case study about Codex, the company uses Codex, GPT-Live-1, and GPT-6 Astra to build, operate, and sell its software faster. The title claims a 60 percent sales increase and more than 75 hours saved, but the available summary does not explain the methodology or whether the results are transferable, so these figures remain provider claims.
Microsoft’s redesign makes Copilot more versatile, but not automatically easier to budget for. A central interface, continuously running agents, and internal app creation could reduce routine work, while consumption-based pricing and broader permissions create new oversight duties. The largest open risk is how transparent usage, model selection, and data processing will be under real working conditions.
Sources
- Microsoft baut Copilot erneut um, jetzt mit Autopilot-Agent und nutzungsbasierter Abrechnung – Unknown, 2026-09-25
- Microsoft thinks its new Copilot ‘super app’ will be as influential as Office – Unknown, 2026-09-25
- Microsoft stops insisting you need a “Copilot+ PC” – Unknown, 2026-09-25
- AI Coding Agents for Enterprise: IP Indemnity, Data Residency and 500-Seat Cost Compared – Unknown, 2026-09-27
- KI-Agent optimiert Token-Kosten: Nvidia spart fast 50 Prozent bei Coding-Agenten durch bessere Steuerung – Unknown, 2026-09-26
- Proaction boosts sales 60% and saves 75+ hours with Codex – Unknown, 2026-09-25
Image: Muhammad Jawadur Rahman via Pexels