For years, choosing an AI assistant meant choosing a side: Copilot with OpenAI's GPT, or Anthropic's Claude, or Google's Gemini. Each in its own corner. Microsoft has just broken this pattern. Since March 30, 2026, its Copilot assistant no longer relies on a single brain: it combines GPT and Claude, working togetheron the same question, for more reliable searches.

The new feature arrives in the Researcher agent of Microsoft 365 Copilot, through two modes called Critical and Advisory. The idea is simple: instead of taking a single model at its word, Copilot asks a second model to verify the first's work. For a small or medium-sized enterprise (SME) in Trois-Rivières or the Mauricie region that is starting to rely on AI for its reports, tenders, or market research, this detail makes a significant difference.

1. What exactly Microsoft announced

On March 30, 2026, Microsoft added the ability to combine multiple AI models instead of just one to its Researcher agent (Copilot's advanced search mode). This is known as a multi-model approach. Specifically, two features were introduced.

The first method, Critique, divides the research into two stages. A first model plans, gathers information, and drafts a summary. A second model then reviews this summary to verify the accuracy, completeness, and reliability of the cited sources. In practice, Microsoft often has GPT written and assigns Claude the role of reviewer, much like an employee who writes and an experienced colleague who proofreads before submission.

The second model, Council , poses the same question to GPT and Claude simultaneously. Each produces their own complete report. A third model, acting as an arbiter, reads both versions and prepares a summary: where the two AIs agree, where they disagree, and what each found unique. You thus obtain two perspectives rather than just one.

2. Why two AIs are better than one

The biggest flaw of AI assistants is the so-called "confident error": the model invents a figure, a date, or a source, and presents it with assurance. For a company, a single false statistic in a submission or management report can be very costly in terms of credibility.

Having a second model verify an answer reduces this risk. Microsoft measured the improvement using a test called DRACO (100 complex search tasks across 10 domains, originally designed by researchers at Perplexity). According to Microsoft, Critical mode scored about 7 points higher on this test, nearly 14% better than a competing in-depth research tool. The improvements affected the depth of analysis, the quality of presentation, and the accuracy of the facts. These figures come from Microsoft; you should verify them with your own usage, but the trend is clear.

For an SME, the benefit is not theoretical. Here are some concrete examples of tasks where double-checking is important:

• Write a market analysis before a product launch.
• Compare suppliers or software solutions for a project.
• Prepare a regulatory file or a document intended for an important client.
• Summarize a long contract or tender without missing any clauses.

3. What this means in concrete terms for your SME in Quebec

Good news: there's nothing to install. If your company already has a Microsoft 365 Copilot license, these features appear in the Researcher agent. In the model selector, the "Auto" option activates the verification logic, and "Model Council" initiates the comparison between the two AI models. Access, however, depends on your plan and the permissions granted by your IT administrator (the features were initially targeted at organizations in Microsoft's Frontier program before a wider rollout).

This is exactly the kind of setup where support is invaluable. Activating the right option, deciding who has access within the company, and defining the parameters of the data entrusted to AI: it all requires planning. Our team can handle this for you as part of our managed IT services, taking into account your legal obligations, including Quebec's Bill 25.

One point to note regarding costs, without going into specific prices since they fluctuate: running two AIs consumes more resources than a single one. Microsoft indicates, as a guideline, that Critical mode costs approximately 20% more than a single model, and that Council mode can cost nearly 2.5 times more. The logic is sound: double-checking is reserved for searches that truly warrant it, not for every minor question.

4. The real signal behind this announcement

Beyond its functionality, this news reveals something important about AI in 2026. The tech giants are no longer betting everything on a single model. Microsoft, long associated with OpenAI, is now integrating its rival Claude into its flagship product. The message to businesses: no AI is perfect, and reliability depends on comparing different perspectives.

For an SME, the lesson is easy to apply. Don't rely on a single tool, a single supplier, or a single reflex. The same caution applies to your backups, access control, and security. A self-checking AI is an improvement, but it doesn't replace a human reviewing documents before they go to the client. Always ensure final approval by a real person on documents that matter.

• Use double-checking (Critical or Council) for high-stakes research.
• Keep the simple questions on a single template; it's faster and cheaper.
• Always have a human proofread your work before sending it officially.
• Provide a framework for the confidential data that your teams entrust to AI.

Frequently Asked Questions

Does Copilot really use Claude and GPT at the same time?

Yes, in the Researcher agent of Microsoft 365 Copilot. The Critical mode has one model written and then the other reviewed, and the Council mode questions both in parallel to compare their responses. This happened on March 30, 2026.

Does my SME need a special license?

A Microsoft 365 Copilot license is required. The features appear in the Researcher template selector. Depending on your plan and your IT administrator's settings, access may be phased. An IT provider can verify and activate this for you.

Does this eliminate AI errors?

No, but it reduces them. Having a response checked by a second model improves accuracy, according to Microsoft tests. A human review is still recommended before any important document, especially if it's going to a client.

Get the most out of AI, without the pitfalls

Multi-model AI is a significant advancement, but it only reaches its full potential when properly configured and managed. At OKTO Solutions in Trois-Rivières, we help SMEs in the Mauricie and Quebec regions securely deploy Microsoft 365 Copilot, train their teams, and maintain control over their data. Discover our managed IT services or contact us through our contact page to discuss your needs.