Microsoft teaches Copilot Researcher to cross-check GPT and Claude answers in a single workflow

Microsoft has started rolling out a Critique mode in Copilot Researcher, in which two models, OpenAI's GPT and Anthropic's Claude, work on a single research answer. At the same time, the company has begun deploying a new AI agent, Copilot Cowork, showing that Copilot is increasingly turning from a lone assistant into a set of coordinated helpers.

How Critique works

Critique is simple by design: one model first drafts an answer to the user's query, and a second then checks it for accuracy. In the first phase, GPT answers and Claude acts as an internal reviewer.

For a research assistant this matters more than it might seem. In search, summarization and analytics, the problem is usually not writing text quickly but catching weak spots, imprecise wording and questionable conclusions before a person sees the result.

Microsoft is already talking about the next step: it wants to make the process two-way so the models can swap roles. That means Claude would not only critique GPT's answers; GPT could also check Claude's drafts. In effect, the company is building a mini editorial desk inside a single workflow, with generation and verification kept separate.

This is a notable shift for enterprise AI: the bet is not only on the strength of an individual model, but on how models challenge and check each other to make the final answer more reliable.

Why Microsoft is doing this

For Microsoft, it is also a way to turn access to models from different vendors into a practical advantage rather than just a long compatibility list. Multi-model support often sounded like a formal perk: customers get a choice between several engines, but each task is still handled by only one of them. Here the logic is different: the models work together and cover each other's weaknesses within a single scenario.

This approach fits Microsoft 365 especially well, where users expect Copilot to deliver predictable results in real documents, spreadsheets and research rather than experiments. In practice, it offers several tangible benefits.

"Customers should get not a set of models, but the benefit of their working together," is how

Microsoft describes the idea behind the new mode.

• more accurate answers to research queries thanks to a separate verification step
• less dependence on the strengths and weaknesses of any one model
• a clearer workflow, with generation and critique split into distinct roles
• a foundation for future scenarios in which several models debate one another before delivering a result

For now, Microsoft does not promise that errors will magically disappear, and that is the right stance. One model checking another does not automatically guarantee truth: both systems can make mistakes, accept a false premise or miss problem areas. But even this setup is already better than a single answer with no internal control, especially in products whose output then flows into work correspondence, presentations or analytical notes inside companies' workflows.

Cowork launches alongside

At the same time, Microsoft has begun rolling out Copilot Cowork, another AI agent in the Copilot ecosystem. The original report gives few details about its functions, but the timing matters: the company is showing not a single local feature but a broader move toward a set of specialized agents and work modes.

Researcher handles research scenarios, Cowork extends the lineup, and combining different models within one process hints at where Copilot is heading next.

If the corporate assistant was once seen mostly as a single chatbot, its architecture is now becoming modular. One agent can gather material, another can check it, a third can help with team collaboration. For Microsoft, this is a convenient way to scale Copilot without promising a universal super-agent that is equally good at everything. Instead, the company is building a more practical system: separate roles, separate processes and more control over output quality at each stage.

What it means

Microsoft is pushing Copilot toward a multi-step AI system in which answer quality improves not only through a stronger model but through internal verification. For the market, this is an important signal: the next round of competition in enterprise AI will be fought not only between GPT, Claude and other models, but between entire workflows in which those models interact.