Microsoft just made a significant move in the enterprise AI space that deserves attention. Their Researcher agent in Microsoft 365 Copilot now features multi-model intelligence—essentially using multiple AI models to check each other’s work. It’s a fascinating architectural decision that addresses one of the persistent challenges in AI-driven research: how do you know when your AI is wrong?
In this post, we’ll break down what Microsoft announced, how Critique and Council work under the hood, and why this multi-model approach matters for enterprise AI deployments.
The Problem with Single-Model Research
Here’s the thing about AI research assistants: they’re impressive until they’re not. A single model handles everything—planning, sourcing, synthesis, and writing. That’s a lot of responsibility for one system, and it creates blind spots. The model that wrote the claim is the same model that would need to catch its own errors. We’ve all seen how that plays out.
Microsoft’s response is architectural separation. Instead of hoping a single model catches its own mistakes, they’ve split responsibilities between two AI partners. One model focuses on deep exploration and structured synthesis. The other acts as an expert reviewer, validating claims and strengthening the output before anything reaches the user.
This isn’t just a philosophical choice—it’s backed by benchmark data showing meaningful improvements across factual accuracy, analytical depth, and presentation quality.
Critique: The Expert Reviewer Architecture
Critique is Microsoft’s new default experience in Researcher (activated when “Auto” is selected in the model picker). The system uses models from both Anthropic and OpenAI, assigning them distinct roles in a generator-reviewer workflow.
The generator model leads the initial phase: planning the research task, iterating through retrieval, and producing a draft report. Then the reviewer model takes over, applying what Microsoft describes as “rubric-based evaluation”—a structured review process borrowed from academic and professional research settings.
The reviewer examines reports across several dimensions:
Source Reliability Assessment. The reviewer prioritizes reputable, authoritative, and domain-appropriate sources. This isn’t just checking if citations exist—it’s evaluating whether those sources are actually suitable for the research context.
Report Completeness. Does the final output actually answer what was asked? The reviewer evaluates whether the report satisfies the original intent comprehensively with relevant and unique insights.
Strict Evidence Grounding Enforcement. This is where things get rigorous. The reviewer enforces what Microsoft calls a “conservative grounding standard,” requiring every key claim to be anchored to reliable sources with precise citations. No hand-waving allowed.
The key insight here is that evaluation gets as much emphasis as generation. That feedback loop—where one model’s output gets systematically challenged by another—is what drives the quality improvements.
The DRACO Benchmark Results
Microsoft evaluated Critique against the DRACO benchmark (Deep Research Accuracy, Completeness, and Objectivity), which consists of 100 complex research tasks across 10 domains. This benchmark was introduced by researchers from Perplexity and academia in February 2026, with tasks originating from anonymized real-world usage patterns.
The results are notable. Researcher with Critique achieved a +7.0 point improvement in aggregated score—a 13.88% improvement over Perplexity Deep Research (using the Claude Opus 4.6 model), which was the best system reported in the original benchmark paper.
Breaking down the improvements by evaluation dimension:
- Breadth and Depth of Analysis: +3.33 points
- Presentation Quality: +3.04 points
- Factual Accuracy: +2.58 points
- Citation Quality: Modest improvement, primarily through better use of existing sources
All dimensions showed statistically significant improvements (paired t-test, p < 0.0001). The largest gains came from analytical depth and presentation—areas where the reviewer model pushes the generator to identify missing angles, close coverage gaps, and produce more organized narratives.
At the domain level, 8 of 10 domains showed statistically significant improvements. The exceptions were Academic and Needle-in-a-Haystack queries, both of which exhibited high variance—which makes sense given the specialized nature of those domains.
Council: Side-by-Side Model Comparison
While Critique is about sequential improvement (generate, then review), Council takes a different approach: parallel execution with comparative analysis.
When users select “Model Council” in the Researcher model picker, the system runs an Anthropic and OpenAI model simultaneously. Each model produces a complete, standalone report. Then a dedicated judge model evaluates both reports and creates a distilled summary.
The judge’s output highlights:
- Where the models meaningfully agree
- Where they diverge in magnitude, framing, or interpretation
- Unique contributions from each model that the other may have overlooked
This is particularly useful when you need to understand how different models weigh and interpret the same evidence. Different models weight evidence differently, frame findings differently, and may catch details others miss. Council surfaces those differences explicitly rather than hiding them behind a single output.
Why Multi-Model Matters for Enterprise
This architectural shift carries implications beyond Microsoft’s specific implementation.
Reduced Single-Point-of-Failure Risk. When your research depends on a single model, you inherit all of that model’s biases and blind spots. Multi-model architectures distribute that risk.
Explicit Uncertainty Surfacing. Council’s approach of showing where models diverge is valuable for decision-makers. If two frontier models disagree on something, that’s important context—it suggests the question might not have a clear-cut answer or that more investigation is warranted.
Quality Through Structured Review. The Critique architecture creates a quality-focused dynamic where the reviewer model is tasked with strengthening claims and improving rigor—similar to peer review in academic and professional settings.
Model Optionality. Microsoft explicitly mentions that this design “provides clear optionality across generator and reviewer roles, with the ability to support and expand these roles over time.” As new models emerge, the architecture can incorporate them without fundamental redesign.
Current Availability
Both Critique and Council are available now through Microsoft’s Frontier program. Critique is the default experience in Researcher when Auto is selected, while Council requires explicit selection in the model picker.
For organizations already invested in the Microsoft 365 ecosystem, this represents a meaningful upgrade to research capabilities. For everyone else, it’s a signal of where enterprise AI architecture is heading: away from monolithic single-model systems and toward orchestrated multi-model approaches where different AI systems check and complement each other.
Final Thoughts
Microsoft’s move here validates something many practitioners have suspected: the future of reliable AI isn’t about waiting for a single perfect model. It’s about building systems where multiple models with different strengths work together, challenge each other, and surface disagreements rather than hiding them.
The benchmark results are compelling, but the real test will be how these capabilities perform in the messy reality of enterprise knowledge work—where queries are ambiguous, sources conflict, and “good enough” research decisions can have significant downstream consequences.
If you’re evaluating AI research tools for enterprise deployment, the multi-model question is now table stakes: does your solution use architectural approaches to catch errors before they reach users, or is it hoping a single model gets everything right the first time?