The new Critique feature lets Copilot use several of the really big AI language models to answer one question. One model creates the initial answer, and another one checks it for correctness and good quality. This two-step process is intended to find errors before you see them.
Multi-model ‘Critique’ feature explained
Right now, GPT will do the first draft and Anthropic’s Claude will review it, but Microsoft says they’ll soon make it go both ways – either model can create the answer or check it, depending on the job. This should make things both faster and more likely to be right.
Using two models should decrease “AI hallucinations”, which is when the AI confidently states something that’s just plain wrong. Copilot will look at the outputs from each model to lower the chances of this happening and give you more trustworthy summaries, research notes and help with writing. This change is for professionals who need information that’s accurate and can be proven.
At the same time as Critique, Microsoft is releasing Model Council. This shows you the answers from several models at the same time, next to each other. You can compare how they sound, the details they give, and what changes they suggest without having to use separate programs or ask the question again.
Model Council for side-by-side comparisons
Model Council is good for teams that need to verify the AI’s answers or to decide on the best wording for different groups of people. It also helps with checking how things are done, because you can see how each model came to its conclusion. Being able to see this is important for following rules and keeping quality standards.
Within the comparison view, you decide which models to include and how much importance to give to each one. This lets organizations have a standard way of working with AI but still have the freedom to choose the model and the company providing it.
Microsoft is also making Copilot Cowork available to more people. Copilot Cowork is an AI tool that acts on its own, and was inspired by the idea of AI “agents” that work independently. It’s going to be available to people in the Frontier program, who get to try out certain AI features early and give Microsoft their opinions.
Copilot Cowork reaches early-access Frontier members
Copilot Cowork is part of the growing trend toward AI assistants that can handle multiple steps in a task with less human help. Testing it now will show Microsoft how well the agent can manage projects, do research and do repetitive jobs before it’s widely released.
Those in the Frontier program can test Copilot Cowork in real-world situations and give Microsoft data about how it performs. This will help Microsoft improve the safety features, how the agent plans its work, and what happens when something goes wrong. Releasing the agent to a limited group first and improving it over time is how Microsoft is approaching making AI more independent within businesses.
All of this is happening as major cloud and AI companies are all releasing their own AI systems that generate things. Microsoft is presenting Copilot as an assistant that can use many models from different companies, and this could be a benefit for customers who want to use a variety of models without having to manage many different payments. and multimodal models and specialized autonomous agents from other providers.
Market context and competitive pressures
Competition includes large AI models that work with multiple kinds of data, and AI agents that are focused on doing specific jobs from other companies. Microsoft’s plan to use multiple models is meant to keep Copilot useful by combining the best qualities of each one, and by offering strong business-level controls like Model Council.
How well Copilot is accepted will depend on how it performs, how much it costs, and how much people trust it. Businesses usually need to get the same results every time, a record of what was done, and ways to prevent the AI from giving incorrect information. Microsoft is responding to these needs by focusing on the Critique feature and the comparison views.
For businesses, these improvements mean you’ll get results from research, writing and helping with decisions that are quicker and of better quality. Teams that are already using Copilot may find that the multi-model system means they don’t have to check things over as much, and they’ll have more confidence in the work the AI helps them with.
Implications for enterprise users and developers
IT and development teams should pay attention to how Copilot connects to other systems and the rules about data. Using multiple models could make it harder to follow the rules if the different companies have different policies about how they handle data. Clear ways to set things up and controls over the data will be essential for safely using Copilot.
In general, Microsoft’s changes show a move towards AI systems that work with many models together and AI assistants that can do more on their own. Testing Copilot with the Frontier program will show how well these ideas actually increase productivity for businesses in their everyday work.
