Microsoft moves to multi‑model Copilot
Microsoft is embedding multi‑model review into Microsoft 365 Copilot, signalling a shift away from single‑model enterprise strategies toward orchestration across several models for different tasks. That changes procurement conversations from choosing a single 'best' model to selecting a control plane that can route tasks to the right underlying systems. (geekwire.com)
Microsoft just put rival artificial intelligence models inside the same Microsoft 365 Copilot workflow, so one model can draft an answer and another can challenge it before the user sees the result. Microsoft announced the new setup for its Researcher tool on March 30, 2026, with two features called Critique and Council. (techcommunity.microsoft.com) Critique works like a writer and an editor sharing one document. Microsoft says OpenAI’s GPT model produces the first draft, and Anthropic’s Claude model reviews that draft for gaps, weak logic, and missing evidence. (techcommunity.microsoft.com, geekwire.com) Council goes one step further and asks several models to tackle the same research problem in parallel. Microsoft says the system then compares those answers and synthesizes a final response with higher confidence than a single pass from one model. (techcommunity.microsoft.com) This is a sharp break from how most companies bought generative artificial intelligence in 2023 and 2024. Back then, the usual question was whether to standardize on OpenAI, Anthropic, Google, or Meta, as if one foundation model would become the company-wide default. (geekwire.com) Microsoft has been laying the groundwork for this shift for months. In September 2025, the company said it was “expanding model choice” in Microsoft 365 Copilot by adding Anthropic models alongside the OpenAI systems that had powered most Copilot experiences. (microsoft.com) The important product is no longer just the model. The important product is the layer that knows when to call a model, when to pull in company data from Microsoft Graph, and when to hand the job to an agent or another model. (github.com, learn.microsoft.com) Microsoft already uses that routing idea elsewhere in its stack. Copilot Studio lets builders choose a primary model for an agent, and Azure AI Foundry says its agent service works with a large catalog of models plus connectors to outside tools and data sources. (learn.microsoft.com, azure.microsoft.com, azure.microsoft.com) That changes the sales pitch inside big companies. A bank, hospital, or manufacturer no longer has to bet that one model will be best at writing, summarizing, coding, and checking facts at the same time; it can buy a system that routes each task to the model best suited for that step. (geekwire.com, computerworld.com) Microsoft is also making that routing more visible to customers in Europe. A new Microsoft 365 Copilot feature called flex routing is scheduled to start on April 17, 2026, for European Union and European Free Trade Association tenants, and it can send requests to nearby regions when local demand spikes. (office365itpros.com) The catch is that more models mean more governance work. Every extra model adds another vendor relationship, another set of safety rules, and another question about where data is processed and how outputs are audited. (learn.microsoft.com, azure.microsoft.com) So the new contest in enterprise artificial intelligence looks less like “which model wins” and more like “which platform can supervise all of them.” Microsoft is betting that if Copilot becomes the switchboard for OpenAI, Anthropic, Microsoft Graph, and custom agents, customers will stay with the switchboard even when the underlying models keep changing. (geekwire.com, techcommunity.microsoft.com)
