The end of the single‑model era
Microsoft 365 Copilot appears to be running more like an ensemble than a single model, with a Researcher agent that has GPT and Claude check each other's work — a signal that enterprise AI is shifting toward composed workflows. That design mirrors distributed systems thinking: redundancy, verification, and specialized components rather than one monolithic decision point. The implication is that AI platform architecture will increasingly centre on orchestration, routing and cost‑aware verification rather than a pure model selection debate. (geekwire.com) (x.com)
For years, the sales pitch in artificial intelligence was simple: pick the smartest model and use it everywhere. Microsoft is now shipping something closer to a committee, where one model writes and another model checks the work inside Microsoft 365 Copilot. (microsoft.com) (geekwire.com) The product at the center of this shift is Researcher, Microsoft 365 Copilot’s deep-research agent for work. Microsoft says Researcher can pull from the web and from work data like files, emails, meetings, and chats that a user already has permission to access. (support.microsoft.com) On March 30, 2026, Microsoft added two new Researcher features called Critique and Council. Microsoft says those features are built to improve “accuracy, depth, and confidence” by using more than one model on the same task. (microsoft.com) Critique works like an editor standing over a writer’s shoulder. Microsoft says one model produces a draft and another model reviews it, challenges weak claims, and pushes the system to revise the answer before the user sees it. (microsoft.com) (support.microsoft.com) Council works like asking several specialists for separate opinions before you make a decision. Microsoft says it can route parts of a research job across multiple models and then combine the results into one report. (microsoft.com) The notable part is the cast. Microsoft support pages now say Researcher supports both Generative Pre-trained Transformer models from OpenAI and Claude models from Anthropic, and users can choose which family helps generate or refine a report. (support.microsoft.com 1) (support.microsoft.com 2) That is a sharp break from the old idea that one company’s model stack would stay sealed inside one company’s product. Microsoft invested heavily in OpenAI, but its Copilot products now also expose Anthropic models in Researcher and in Copilot Studio, where enterprises build custom agents and workflows. (microsoft.com 1) (microsoft.com 2) This looks less like a chatbot and more like a distributed system, which is the kind of software architecture companies use when one machine is not trusted to do everything alone. Distributed systems spread work across components, add redundancy, and check results at multiple points so one failure does not ruin the whole job. (geekwire.com) (microsoft.com) That design fits enterprise software better than the single-model race did. A law firm, bank, or pharmaceutical company usually cares less about which model tops a benchmark and more about whether the system can trace sources, recover from mistakes, and spend extra compute only when a task is risky enough to justify it. (support.microsoft.com) (geekwire.com) Microsoft is also putting these experiments behind its Frontier program, which is its preview channel for early Copilot features. Microsoft said in an earlier Frontier announcement that Researcher and Analyst would begin rolling out in phases in May, before wider availability. (microsoft.com 1) (microsoft.com 2) The fight in enterprise artificial intelligence is starting to move up a layer. The question is no longer just whether OpenAI, Anthropic, or another lab has the best model on a given day, but which platform can route work between models, verify the answers, and keep the bill low enough for thousands of employees to use it every day. (geekwire.com) (microsoft.com)
