Microsoft is expanding its Microsoft 365 Copilot “Researcher” agent with new multi-model capabilities designed to improve the accuracy and depth of AI-generated research outputs.
The update introduces a “Critique” system that assigns separate roles for generation and evaluation, alongside a “Council” feature that compares outputs from multiple models and highlights agreement, divergence, and unique insights.
Internal testing using the DRACO benchmark showed that Researcher with Critique outperformed previously reported systems by 13.8% (7.0 points) in aggregate score.
“We see the largest improvement in Breadth and Depth of Analysis (+3.33), followed by Presentation Quality (+3.04) and Factual Accuracy (+2.58),” Microsoft said in a blog post. “All dimensions show statistically significant improvements (paired t-test, p < 0.0001).”
The Council feature runs multiple models in parallel to generate independent reports, with a judge system synthesizing key differences and insights to help IT teams compare interpretations.
“In simple terms, it’s like having a smart professional plus a strict reviewer,” said Pareekh Jain, CEO of Pareekh Consulting. “But it’s still incremental, not magic. It reduces errors but does not eliminate them.”
Others point out that model orchestration alone may not be enough to drive meaningful enterprise outcomes.
“Multi-model systems reach their full potential when integrated with internal enterprise data such as CRM and HRM systems,” said Neil Shah, VP for research at Counterpoint Research. “This ensures that AI-driven insights are contextually nuanced, reflecting the company’s unique market position, customer heuristics, and the specific requirements of the decision-maker.”
Performance and governance concerns
Microsoft’s DRACO benchmark results appear strong, but enterprises should approach them with measured caution.
“Think of it as a best-case test; it shows AI models can check each other and catch mistakes, but real company data is much messier with conflicting info and outdated docs,” Jain said. “There’s also a risk of judge bias; if both AIs are similar, the reviewer might miss the same errors. And while benchmarks measure logic, they don’t capture real business value.”
The shift to multi-model systems introduces new layers of operational complexity for enterprise IT teams. Systems are more powerful but also harder to manage.
Instead of a single input-output flow, organizations must now track a chain of interactions that includes the initial draft, critique, and final output.
“This creates a bigger audit trail that security and compliance teams must review to understand how decisions were made,” Jain added. “It also increases cost and latency, since one question can trigger many model calls. Another challenge is accountability. If something goes wrong, it’s harder to know which part failed, like the generator, the reviewer, or the system managing them.”
Analysts say this will require enterprises to rethink governance frameworks around AI deployment.
“Enterprises must prioritize governance of the model to the output selection process, and the refinement of how multiple responses are blended or selected,” Shah said. “This continuous monitoring and calibration will become a fundamental part of Process Quality Management.” Enterprises will also need structured mechanisms to evaluate outputs and their real-world impact, ensuring traceability across the decision-making process and improving how multi-model systems are managed over time, Shah added.