Summary
Highlights
Sumeet explains how AI can write initial boilerplate SQL queries for advanced analyses. He introduces the concept of Claude intelligently monitoring dashboards, identifying trends, and summarizing data, automating tasks that typically consume significant analyst time. Claude can also build context from various sources like Slack and internal tickets, enriching the analysis beyond just data.
Claude can assist in crafting data stories by conducting side investigations and deep dives to fill information gaps. Furthermore, it can analyze past experiments and codebase changes to size potential impacts. The ultimate vision is for Claude to handle an entire analytical loop, from initial analysis and experiment setup to evaluating success and proposing further changes, effectively acting as an end-to-end product manager, data scientist, and engineer.
Compared to traditional dashboards, AI-powered systems can provide customized insights and proactively ask questions, ensuring that data is always reviewed critically. Sumeet shares an example where Claude identified an incident in Slack impacting metrics, saving time in debugging anomalous data. This capability allows for more focused and higher-level data analysis by human analysts.
Sumeet details the process of building an MCP (Managed Code Project) using public startup funding data. The initial setup involves defining three core queries, showing Claude how to interact with the data and establishing key fields and documentation. These queries should reflect typical joins and potential analysis patterns relevant to the domain.
The process involves instructing Claude to create a startup funding MCP using the predefined queries, providing context about the data set, and explaining the purpose of the MCP (e.g., weekly reporting, deep dive analysis, trend monitoring). Sumeet emphasizes including evaluation questions for later testing.
Sumeet discusses critical issues like managing context windows to avoid overwhelming Claude with large query results. He highlights the importance of reminding Claude about query limits and providing tight semantic context to prevent confusion, especially when multiple data segmentations exist. Integrating external context from tools like Slack and Google Drive is also crucial for comprehensive analysis.
Sumeet demonstrates the MCP by asking Claude to identify and rank the most recent Series A deals based on their likelihood of securing a Series B. Claude provides a detailed report, including rationales based on factors like company industry (AI vs. healthcare) and funding amount, showcasing its analytical capabilities.
Sumeet asks Claude to determine which AI coding tool has the most momentum. Based on the data, Claude identifies Cursor as a leader, followed by Replit, and comments on market dynamics and factors contributing to their success, such as A16Z backing for Cursor.
To prevent issues like overloading databases or incurring high token costs, Sumeet explains the use of 'skills' in Claude. These skills allow data teams to design specific behaviors and limitations for AI agents, such as imposing query limits or timeouts. This ensures that even non-technical users like PMs can leverage AI for data analysis safely and efficiently.
Sumeet shares insights from Brax's data on AI tool spending. Cursor dominates the AI coding tool market for both startups and enterprises. In large language models, OpenAI leads in enterprise adoption, often due to brand awareness and existing subscriptions, while Anthropic's Claude is gaining ground among startups for agentic applications, especially for in-app integrations.
Sumeet advises data scientists and PMs to begin by selecting three core queries and writing detailed context about their specific domain. This initial setup is crucial for building effective data MCPs, whether for internal use within Claude Code or for integrating with other BI tools. It enables AI agents to understand data structure, common joins, and analysis components, ultimately streamlining the analytical workflow.