Agentic Business Intelligence done right

Key Concepts

• AI Agents: Software entities designed to perform specific tasks, in this context, related to data analysis and business intelligence.
• Subagents: Specialized AI agents that are part of a larger, orchestrated workflow managed by a main agent.
• Agent Development Kit: A set of tools and resources used to build and deploy AI agents.
• Metadata: Descriptive information about data, including technical specifications and business context.
• Orchestration: The automated coordination and management of multiple AI agents to achieve a complex goal.
• Data Engineering: The process of preparing and transforming data for analysis, including mapping data to relevant dimensions.
• Business Intelligence (BI): The process of analyzing data to gain insights and make informed business decisions.
• SQL (Structured Query Language): A standard language for interacting with databases.
• Vega-Lite: A declarative format for creating interactive data visualizations.
• Gemini: Google's AI model used for tasks such as chart generation and data analysis.
• Vertex AI: Google Cloud's machine learning platform used for deploying and running AI agents.

Demo Overview: Lead Conversion Trends Analysis

The demo showcases an AI agent designed to analyze lead conversion trends by country using Salesforce data. The agent interacts with subagents specialized in CRM business analysis, data engineering, and BI engineering.

1. Initial Query: The user asks the agent about "lead conversion trends by country."
2. Agent Orchestration: The main agent activates its subagents, each with specific roles in the analytical workflow.
3. Subagent Tasks:
   • CRM Business Analyst: Analyzes the problem and formulates the analytical approach.
   • Data Engineer: Generates and validates SQL queries to retrieve the necessary data from BigQuery.
   • BI Engineer: Uses Gemini to generate interactive charts in Vega-Lite format.
4. Chart Generation: Gemini creates a chart that allows users to filter data by country for detailed exploration.
5. Data Analysis and Recommendations: The agent analyzes the data presented in the chart, makes recommendations, and suggests next steps.
6. Tool Validation: Ensuring the validation of the tools used by the subagents.

Case Study: Top 5 Customers by Value in the US in 2022

This example illustrates the complexity involved in mapping a seemingly simple question to the underlying data.

1. Data Mapping: The query requires mapping "customers" to customer master data and "value" to sales transactions.
2. Data Engineering Challenges: The data engineer must map "US" to a geographical dimension and "2022" to a time dimension (potentially a fiscal time dimension, depending on the user's perspective).
3. Dependency Mapping: The system needs to understand the dependencies between different data entities and dimensions.

Architecture and Components

The system is built on a robust architecture involving multiple components and services.

1. Agents:
   • Main Orchestration Agent: Manages the workflow and coordinates the subagents.
   • Subagents: CRM Business Analyst, Data Engineer, and BI Engineer, each with specialized tools.
2. Tools: SQL generation, SQL validation, chart generation (Vega-Lite), and chart validation tools.
3. Platforms:
   • Gemini and Vertex AI: Used for AI capabilities, including chart generation and agent deployment.
   • Cloud Storage: Stores artifacts like images.
   • Firestore: Stores session data.
   • BigQuery: Serves as the data warehouse for storing CRM data.
4. Agent Development Kit: Facilitates the creation and deployment of agents.

Best Practices: The Importance of Metadata

While a solid architecture is crucial, the quality of the metadata is equally important.

1. Data Description: It's essential to describe the data sources, including APIs and tables, along with their cardinality.
2. Business Lingo: The descriptions should use business terminology and synonyms that align with how users interact with the data.
3. Relationships: Metadata should capture the relationships between different data products.
4. "Secret Sauce": The combination of detailed technical information and business-friendly descriptions is the "secret sauce" for each customer.

Getting Started

1. GitHub Repository: The code and resources are available in a GitHub repository (linked below the video).
2. Experimentation: Users are encouraged to try the system themselves.
3. Feedback: The creators welcome questions, comments, and feedback.

Conclusion

The video highlights the power of AI agents in automating complex data analysis tasks. Key takeaways include the importance of agent orchestration, specialized subagents, a robust architecture, and comprehensive metadata that bridges the gap between technical data and business understanding. The system demonstrates how AI can mimic the human process of converting questions into SQL code and mapping data dependencies.