Researcher Agent in Microsoft 365 Copilot: Deep Research for Knowledge Workers

Researcher Agent in Microsoft 365 Copilot: Your Assistant for Deep Research at Work

This post introduces the Researcher agent, a new capability within Microsoft 365 Copilot designed to assist knowledge workers with deep research tasks. It leverages advanced reasoning models to distill vast amounts of data from both enterprise and web sources into well-founded conclusions, aiming to influence market-entry strategies, sales pitches, and R&D investments.

The Need for Deep Research

Modern information workers require AI models that can reason across enterprise data (emails, chats, documents, applications) and web data. Existing tools often focus on web-centric research, leaving a gap for comprehensive enterprise data analysis. The Researcher agent addresses this by providing thorough, accurate, and deeply contextualized research reports, saving employees significant time and effort.

Our Approach: Mimicking Human Research

The Researcher agent's approach is modeled after how a human researcher would tackle a complex task:

Initial Planning Phase (P0): The agent analyzes the user's request and context to create a high-level plan. It may ask clarifying questions to ensure alignment with user expectations regarding content and format. Insights from this phase are denoted as I0.
Iterative Research Phase: The agent enters a loop of Reason → Retrieve → Review for each subtask until diminishing returns are met.
- Reasoning (Rj): Deep analysis to determine the next subtask and identify missing information.
- Retrieval (Tj): Searching across various data sources (documents, emails, messages, calendars, web data) to fetch necessary details.
- Review (Vj): Evaluating the retrieved information for relevance to the original query and storing findings on a 'scratch pad'. New insights gained in iteration j are denoted as ΔIj, which are added to the cumulative insights: Ij = Ij-1 ∪ ΔIj. The agent stops when the marginal insight ΔIm falls below a threshold ε.
Synthesis Phase: The agent consolidates all findings (Im) to synthesize a coherent report, including explanations and source citations for traceability.

The Researcher Agent in Action

Example Scenario: A user asks, "How did our Product P perform in Q4 compared to industry trends?"

Planning: The agent identifies subtasks: (1) retrieve internal Q4 sales numbers for Product P; (2) find industry news or analyst reports on Q4 trends. It might ask for clarification on specific regions or competitors.
Iterative Research:
- Iteration 1: Reasons to start with internal sales data, retrieves the Q4 sales report, and reviews the contribution of feature F to sales growth.
- Iteration 2: Adapts the plan to explore feature F, retrieves related internal/external communications, and searches the web for competitor offerings. Reviews customer reception and industry news related to F.
- This process continues, gathering information until new iterations yield minimal new insights.
Synthesis: The agent generates a report comparing Product P's Q4 performance to the market, citing internal sales data and external industry analysis, and highlighting feature F as a competitive differentiator.

Technical Implementation

Core Technology: Leverages OpenAI's deep research model, trained on an upcoming OpenAI o3 model variant. Achieves high accuracy on benchmarks like Humanity's Last Exam (HLEx) and GAIA.
Reasoning over Enterprise Data: Extends the model with Copilot tools for retrieving enterprise data (meetings, documents) and third-party content via graph connectors (wikis, CRMs). These tools are managed by the Copilot Control System for security and administration.
Enterprise Contextualization: Integrates with the enterprise knowledge graph to personalize results based on user and organizational context (people, projects, products). This allows for more nuanced clarifying questions and tailored starting conditions for the research model.
Deep Retrieval: Employs broad, shallow retrieval across diverse data sources first, followed by semantic passage retrieval within documents to maximize insights per iteration. This contrasts with a purely serial approach.
Integrating Specialized Agents: Seamlessly integrates with domain-specific agents (e.g., Sales Agent) to delegate complex subtasks, enhancing the Researcher's capabilities with specialized knowledge. Agents can be chained for multi-step workflows (e.g., preparing for customer meetings by accessing calendar, communications, and CRM data).

Results and Impact

Response Quality: Evaluated using the ACRU framework (Accuracy, Completeness, Relevance, Usefulness), Researcher shows significant improvements over M365 Copilot Chat:
- +88.5% Accuracy
- +70.4% Completeness
- +25.9% Relevance
- +22.2% Utility
- Cites an average of ~10.1 sources per response, with enterprise documents, web pages, emails, and meeting transcripts frequently used.
Time Savings: Pilot users (Product Managers, Account Managers) reported saving 6-8 hours per week, with tasks that previously took days now completed in minutes. Users trust the agent's ability to search across all accessible data sources.

What's Next

Reinforcement Learning: Further improving report quality by training reasoning models on real-world work tasks using reinforcement learning, optimizing the agent's decision-making process.
User Control: Introducing 'steerability' to allow users to control the information sources used for report generation.
Agentic Orchestration: Generalizing the integration of Microsoft agents and enabling users/admins to incorporate custom agents into the Researcher workflow for specialized output formatting (e.g., legal briefs).

Conclusion

The Researcher agent has the potential to significantly transform knowledge work by providing accurate, detailed, and time-saving research capabilities. Future developments aim to enhance its quality, customization, and integration with other specialized agents, solidifying its role as an indispensable workplace tool.

For more details, refer to the blog post on reasoning agents within M365 Copilot.