Copilot Got 50% Faster as Work IQ Reshapes the Workplace AI Surface

AI 요약

• What happened: Microsoft redesigned the Microsoft 365 Copilot app and in-app entry points on May 28, 2026.
  • The announcement moves Copilot from a static prompt line toward a task-aware workspace backed by Work IQ and a single in-app Copilot entry point.
• The numbers: Microsoft says customer testing cut app load time by more than 50% and improved complex chat first-token latency by 10%.
  • It also reported usage gains of Word 27%, Excel 33%, PowerPoint 43%, and Outlook 30%, but those figures come from Microsoft-defined rollout windows.
• Builder takeaway: Workplace AI competition is shifting from model names to context, latency, permissions, and app-native execution.

Microsoft announced a new design for Microsoft 365 Copilot on May 28, 2026. The post came from Jon Friedman, Microsoft’s chief design officer. At the surface level, the update changes the Copilot app and the Copilot buttons inside Word, Excel, PowerPoint, and Outlook. The more consequential part is the role Microsoft now gives to the prompt box. Copilot is being moved from a one-line question field into a workplace AI surface that reads task context, suggests tools, and edits inside the apps where the work already lives.

Microsoft opens the announcement by saying that work rarely happens in neat lines. The redesigned Copilot app widens the prompt line and places task-aware tools and controls under it. A user can paste a long document or structured content, preserve inline formatting before sending, and receive output that is shaped around follow-up actions rather than a plain chat answer. In this design, the prompt is closer to a work canvas than to a text input. A task that starts in chat can continue into a Word paragraph, an Excel range, or a PowerPoint slide without forcing the user to copy context back and forth.

The repeated product term in the announcement is Work IQ. Microsoft describes Work IQ as an intelligence layer that draws on work signals from email, files, chats, and meetings so Copilot can handle deeper context. Microsoft also says users can see and control Work IQ when it is active. The design separates quick-response requests from requests that need deeper reasoning, and it can connect that choice to model selection. Model choice is not presented as a standalone feature. It sits behind a workplace context layer that decides what kind of work is being attempted.

Microsoft also published performance numbers. It says the Copilot app is now close to twice as fast, with load time reduced by more than 50%. The footnote matters. The figure comes from customer testing between March 10 and March 17, 2026. Microsoft compared about 11.06 million users in a treatment group with about 11.16 million users in a control group. The metric was User Perceived Load Time, defined around the maximum latency of core elements such as the left navigation pane, prompt input box, and prompt suggestions. This is not a claim about total server processing time. It measures when the important parts of the app feel ready to a user opening Copilot.

The announcement also says first-token response time for complex chat prompts improved by 10%. Microsoft defines that measurement at the 95th percentile, which means the figure is about the slowest slice of requests rather than the average request. That distinction is useful for workplace AI. Simple prompts already have a small waiting cost. Users are more likely to feel latency on long-document questions, meeting summaries, multi-file requests, and other tasks that pull in more context. Improving the tail of complex prompts can change whether a user treats Copilot as a tool to keep open or as a separate service to visit only when they are willing to wait.

Microsoft’s performance story has three parts. Copilot app load time fell by more than 50% in the March 10 to March 17 customer test. Complex chat first-token latency improved by 10% at the 95th percentile. The comparison covered roughly 22.22 million users across the treatment and control groups. Those numbers are large, but they remain Microsoft’s own measurements of a Microsoft product surface.

Usage numbers inside the apps are more specific. Microsoft says Copilot usage among commercial users increased after the new in-app experiences rolled out: Word rose 27%, Excel 33%, PowerPoint 43%, and Outlook 30%. The comparison windows differ by app. For Word, Excel, and PowerPoint, Microsoft compared activity after the May 8 to May 12, 2026 rollout with activity before the rollout from May 1 to May 5. For Outlook, it compared daily active usage after January 27 to February 24, 2026 with the previous 28-day period from December 30, 2025 to January 27, 2026. The footnote gives more signal than the headline because the windows are short and not identical.

Those figures should not be read as long-term market validation. The Word, Excel, and PowerPoint comparisons each cover five-day windows. Outlook uses a longer 28-day daily active usage comparison, but the timing and baseline differ from the other apps. Microsoft also frames the data as short-term changes that may not represent long-term usage. The cautious reading is that discoverability, latency, and in-app entry points can move invocation behavior quickly. The figures do not prove that Copilot has become a durable habit or that the quality of completed work improved.

The product shift is bigger than a faster app shell. Microsoft says Copilot now appears as a single entry point inside Microsoft 365 apps. The entry point sits over the work and understands the document context underneath it. A user can open a side pane from that entry point and treat Copilot like an editing partner tied directly to the file. Copilot can also be invoked from the canvas itself, such as a paragraph, cell range, or slide. That direction differs from the first generation of AI tools that treated chat and the working document as separate places.

The design connects to Microsoft’s April 2026 announcement that agentic capabilities in Word, Excel, and PowerPoint had reached general availability. In that earlier update, Microsoft described Copilot performing multi-step, app-native actions inside documents, worksheets, and presentations. The May 28 design announcement answers a different question: where users discover those capabilities, how they approve or shape the action, and what kind of result they receive. As agents gain more app-native power, the interface has to show more than an input box. It has to expose execution state, change scope, undo paths, sources, and policy constraints.

Microsoft’s enterprise strategy sits behind that surface. In its March 2026 Wave 3 announcement, Microsoft positioned Agent 365 and Microsoft 365 E7 around frontier transformation. Agent 365 was described as a control plane for agent observability, governance, and security. On May 5, Microsoft tied Work Trend Index messaging to Agent 365 general availability, shadow AI agent discovery, Cowork mobile, and federated Copilot connectors. The redesigned Copilot screen is the user-facing edge of that management layer. It makes the enterprise control plane visible through the app experience rather than through an admin-only product story.

For AI product teams, the useful question is not only which model Microsoft attached to Copilot. The sharper question is where the agent is allowed to run. The same model behaves differently when the user asks it to edit a selected Word paragraph compared with when the user copies that paragraph into a separate chatbot. In-app execution can know the source document, selection range, file permissions, and surrounding state. It can show before-and-after changes and attach the action to existing Microsoft 365 identity and policy models. The cost is interface complexity. Users and administrators need clearer answers about which files were read, which actions are pending, and what will change if they approve the result.

Work IQ is abstract as a brand name, but the underlying components are concrete. Using email, files, chats, and meetings requires connectors, identity, permission trimming, retrieval, summarization, and audit logging to work together. Microsoft Graph, Entra, Purview, and Fabric are the existing product layers likely to carry much of that burden. Google Workspace Gemini, OpenAI workspace agents, Glean, and Notion AI face the same product problem. The differentiator is not whether a vendor can call an LLM. It is how consistently the vendor can preserve context and policy across the tools where people already create and review work.

This announcement is not primarily about Copilot Studio or developer APIs. It still carries a practical lesson for AI application builders. Expanding a prompt box does not automatically create a task-aware workspace. The product has to preserve structured pasted content, propose tools based on context, return results that are easy to scan, and execute follow-up actions where the user can inspect the change. Microsoft explicitly names output quality, readability, usefulness, and trustworthiness as design targets because workplace AI UX is less about where an answer appears and more about whether the answer can become the next useful action.

The limits are also part of the story. Microsoft’s numbers come from its own customer testing and rollout measurements, not from an independent benchmark. Usage increases do not directly prove retention, work quality, or return on license cost. Work IQ also raises familiar enterprise data-boundary questions. Administrators will need to know which signals can be used for which requests, what controls are available to users, which connectors are included, and how much action and reasoning context is preserved in audit logs. A faster Copilot surface is easier to try, but faster access also raises the bar for reviewability.

Community reaction to the May 28 design announcement was still limited when the Korean article was researched. Searches did not surface a large Hacker News or Reddit discussion centered on this specific announcement. Microsoft 365 changes often produce louder debate only after tenant rollouts, admin center messages, and user education materials reach customers. If the Copilot entry point becomes more prominent inside Office apps, IT teams may focus first on licensing, data protection, model choice, and third-party connector policy rather than on the visual redesign itself.

The Copilot redesign is a product-level example of the shift from chatbot to agent without relying on that phrase. The prompt line becomes a workspace. Work IQ places workplace signals before the model call. A single Copilot button moves the agent closer to the Word document, Excel sheet, PowerPoint slide, or Outlook thread. Microsoft’s 50% load-time reduction and app-level usage gains are early evidence about invocation behavior, not final evidence about workplace productivity. The next test is quality: whether Copilot can make document, spreadsheet, and slide changes more accurately while leaving behind action logs that organizations can inspect.