Microsoft rewords Copilot positioning

Microsoft has not exactly changed what Copilot is. It has changed what Microsoft is willing to guarantee about it. In the company’s public terms of use, effective October 24, 2025, Microsoft says Copilot is “for entertainment purposes only,” can make mistakes, may not work as intended, and should not be trusted for important advice. That language sat quietly in the legal text for months. It only became a story when people noticed how sharply it clashes with the way Microsoft sells Copilot as a serious tool for work. (microsoft.com) That mismatch matters because Copilot is not a toy product on the edge of Microsoft’s business. It is supposed to be one of the company’s next big revenue engines. Microsoft has spent the past year pushing Copilot into Word, Excel, PowerPoint, Teams, Windows, and the broader Microsoft 365 stack. Then, in early April, Bloomberg reported that Microsoft had shifted its sales strategy after investor pressure, focusing more aggressively on getting customers to pay for Copilot rather than giving away AI features inside larger bundles. (bloomberg.com) The awkward part is that adoption still appears thin relative to Microsoft’s reach. Reporting this week pegged Copilot at roughly 15 million subscriptions against about 450 million commercial productivity seats. Even if that figure is directionally right rather than exact, it describes the real problem. Copilot has broad distribution, but turning that footprint into paid, habitual use has been harder than the marketing suggested. (finance.yahoo.com) That helps explain why Microsoft is now changing the product, not just the pitch. In March, the company announced Wave 3 of Microsoft 365 Copilot and made “multi-model intelligence” a central feature rather than a side experiment. Researcher, one of Copilot’s higher-end agents, now uses new modes called Critique and Council. In Critique, one model drafts and another reviews. In Council, OpenAI and Anthropic models generate separate reports side by side so users can compare where they agree and where they do not. (microsoft.com) This is more than a product tweak. It is Microsoft admitting that a single-model Copilot was not enough. The company is widening the model layer across its stack. Microsoft documentation now describes support for GPT models from OpenAI and Claude models from Anthropic inside Researcher. Admin tools also let organizations enable outside model providers such as Anthropic and xAI in parts of the Copilot ecosystem. Microsoft is still deeply tied to OpenAI, but it is no longer pretending one partner can define the whole experience. (support.microsoft.com) The legal disclaimer and the multi-model push are really the same story seen from opposite ends. One is about limiting liability when AI gets things wrong. The other is about reducing the odds that it gets things wrong in the first place. Microsoft is trying to sell Copilot as infrastructure for work while quietly writing contracts as if users should never trust it too much. The company’s newest features show why. If Copilot were already reliable enough on its own, Microsoft would not need one model checking another before handing over an answer. (microsoft.com) That shift spills outward. When a platform company moves from “one assistant” to “many models plus orchestration,” customers need more setup, more policy, and more monitoring. Someone has to decide which teams can use Claude, which can use xAI, where data can travel, and which outputs need human review. Microsoft’s own documentation notes that Anthropic-based features are handled differently in the EU, UK, and EFTA, where admin opt-in and data-boundary issues still apply. The more Copilot becomes a control plane for other models, the more value moves to the firms that implement governance around it. (learn.microsoft.com) Microsoft is also tightening the paywall around the parts of Copilot that feel most useful. Beginning April 15, 2026, many commercial users without a paid Microsoft 365 Copilot license are set to lose the in-app Copilot Chat experience inside Office apps and get pushed toward the standalone app instead. That is not the move of a company whose free funnel is converting beautifully. It is the move of a company trying to force a cleaner line between casual AI access and the version it hopes businesses will actually buy. (chrismenardtraining.com)