Copilot: Cowork, Counsel & Critique

Key insights

• Wave 3 / The Big 3
  Microsoft’s Wave 3 updates introduce three major Copilot features that move the assistant from reactive help to proactive, multi-model workflows.
  These updates aim to automate long tasks, improve research accuracy, and coordinate multiple AI models inside Microsoft 365.

• Copilot Cowork
  Copilot Cowork acts as an autonomous agent: users give a goal, it plans steps, runs actions across Excel, Teams, SharePoint, Word and Outlook, and shows progress for review.
  All outputs stay in Microsoft 365 so IT can audit and govern them through existing tools.

• Copilot Critique
  Copilot Critique uses one model to draft answers and another to review them, cutting hallucinations and boosting response quality.
  Microsoft reports a notable benchmark improvement — about a 13.8% gain on DRACO-style tests — versus single-model approaches.

• Model Council / Multi-model Strategy
  The Model Council runs models side-by-side or in sequence (for example, OpenAI’s GPT and Anthropic’s Claude) to synthesize agreements and resolve differences.
  This orchestration reduces errors, improves reliability, and leverages each model’s strengths for better final answers.

• Work IQ and Governance
  Work IQ supplies context by pulling signals from email, chats, files and relationships so Copilot actions stay relevant and permission-aware.
  Data remains inside the customer’s Microsoft 365 environment to support compliance, auditing and data loss prevention.

• Benefits and Availability
  Key benefits include saved time on multi-step work, higher research accuracy, and enterprise-ready security controls.
  These features are in limited research previews and will roll out gradually to enterprises as Microsoft expands the program.

In a recent YouTube video, Sean Astrakhan (Untethered 365) walks viewers through Microsoft’s latest Copilot updates, which the industry now calls the Big 3. The video focuses on three headline features in the Wave 3 rollout: Copilot Cowork, Copilot Critique, and the Model Council. This article summarizes his explanation and places those features into practical context for enterprise readers.

Overview and Context

The video frames these features as a shift from reactive assistants toward agentic systems that can plan and execute multi-step work. Sean emphasizes that Microsoft 365 Copilot intends to embed these capabilities across the experience, spanning apps such as Word, Excel, Teams, SharePoint, PowerPoint, and Outlook. Consequently, users can expect Copilot to move beyond single-response tasks toward longer-running workflows that interact across tools.

Furthermore, the video highlights Microsoft’s multi-model approach, which combines strengths from different AI partners. For example, one model focuses on content generation while another checks for accuracy and consistency. As a result, Microsoft aims to reduce common AI errors while making outputs more auditable and traceable within enterprise systems.

How the Big 3 Work Together

Sean breaks down each component and then shows how they link into a broader system. According to the video, Copilot Cowork takes user goals and decomposes them into tasks, then executes those tasks across apps with visible progress tracking and options for human interruption. This design helps workers hand off repeatable procedures while keeping control and oversight.

At the same time, Copilot Critique applies a multi-model verification pattern where one model drafts content and another model critiques it for accuracy, hallucinations, and depth. The Model Council orchestrates these models so they can agree, disagree, or synthesize different perspectives before delivering a final result. Together, these layers aim to balance speed, accuracy, and transparency in research and document workflows.

Benefits and Practical Gains

Sean highlights tangible benefits that enterprises can expect, starting with improved efficiency for routine, multi-step tasks. For example, automating data collection, initial analysis, and slide creation can free employees to focus on judgment and strategy rather than data wrangling. Moreover, the system stores outputs within the organization’s existing environment, which supports auditing and governance.

Additionally, the video notes accuracy gains from multi-model verification, which reportedly improved benchmark performance in early testing. Consequently, organizations can expect fewer hallucinations and higher confidence in AI-supplied research, particularly when the system surfaces disagreements between models as part of the output. This transparency helps subject-matter experts validate results more quickly.

Tradeoffs, Risks, and Governance

However, Sean does not gloss over tradeoffs. Increasing autonomy introduces complexity in monitoring and troubleshooting, and automated agents can amplify mistakes if their planning or permissions are not tightly controlled. Therefore, IT and governance teams must balance flexibility with strict policies to prevent unauthorized actions or data exposure.

Moreover, while multi-model collaboration improves accuracy, it can also increase latency and cost because multiple models run in concert. Organizations will need to decide when full verification is necessary and when faster, single-model responses are acceptable. Finally, integrating these features with existing compliance frameworks requires thoughtful configuration to maintain data residency and audit trails.

Adoption Challenges and Next Steps

Sean outlines practical adoption hurdles including training, trust, and change management. Users will need clear guidance on when to rely on agentic Copilots and when to intervene, and IT teams must provide guardrails and monitoring to measure real-world performance. Without that human-in-the-loop framework, organizations risk over-reliance or misuse of automated decisions.

He also recommends phased pilots that measure productivity gains, accuracy improvements, and governance effectiveness before broad rollout. As these capabilities are currently in limited research preview, early adopters should plan controlled tests that target a few high-value scenarios while preparing policies for permissions, logging, and escalation. Ultimately, the choices around scope, oversight, and verification will determine whether the promise of the Big 3 materializes into reliable, scalable value for the business.

Conclusion: Practical Implications

In summary, Sean Astrakhan’s video offers a clear view of how Microsoft’s Wave 3 aims to transform Copilot into a proactive, multi-model system. While the potential for efficiency and improved accuracy is real, his walkthrough stresses the need for careful governance and staged adoption. Therefore, organizations should pilot purposefully, align governance with business risk, and keep humans in the loop as these agentic features mature.

This article synthesizes the video’s main points to help editorial readers assess the technology’s near-term impact and prepare practical next steps. For teams evaluating Copilot advances, the critical decisions will be how to balance autonomy against control and how to measure outcomes without sacrificing compliance or trust.

Keywords

Copilot Cowork, Copilot Counsel, Copilot Critique, Microsoft Copilot features, Copilot for enterprise, Copilot legal assistant, Copilot collaboration tools, Copilot AI review