Skip to content
Aakash GuptaAakash Gupta

I can’t believe we built an AI employee in 62 mins (Cursor, ChatGPT, Gibson)

This is another episode from our AI PM series. This time, we’re building an AI teammate that runs user research, writes product docs, and powers customer success end-to-end with GibsonAI founder, Harish Mukhami. We're building: Preview – 00:00:00 Building AI Customer Success Agent (Tool Stack) – 00:01:46 Role of GibsonAI in Building Customer Success AI Agent – 00:07:29 Using Data from O3 Mini – 00:09:20 Ad (Amplitude) – 00:10:13 Ad (Linear) – 00:10:45 Directing GibsonAI – 00:11:45 Connecting GibsonAI via MCP – 00:17:38 Role of Cursor – 00:21:10 Python Script Inserting Data – 00:26:56 Understanding Cursor Modes – 00:29:00 Ad (Maven) – 00:30:38 Our Dashboard Is Ready – 00:31:01 Building AI Agent – 00:33:44 The the Most Important Thing Our Agent Is Doing – 00:41:46 Aakash’s Reaction to Output – 00:50:51 Role of CrewAI – 00:52:01 AI Employee Use Cases for PMs – 00:54:47 Why Harish Built GibsonAI – 00:56:35 Final Thoughts – 01:00:15 Podcast transcript: https://www.news.aakashg.com/p/harish-mukhami-podcast 💼 Check out our sponsors: Amplitude: The market-leader in product analytics - https://amplitude.com/session-replay?utm_campaign=session-replay-launch-2025&utm_source=linkedin&utm_medium=organic-social&utm_content=productgrowthpodcast Linear: Plan and build products like the best - https://linear.app/partners/aakash Maven: Check out my own curation of their courses for a discount - http://maven.com/x/aakash 👀 Where to Find Harish LinkedIn:https://www.linkedin.com/in/harishmukhami GibsonAI: https://www.gibsonai.com/?utm_medium=podcast&utm_source=aakash 👨‍💻 Where to find Aakash: Twitter: https://www.twitter.com/aakashg0 LinkedIn: https://www.linkedin.com/in/aagupta/ Instagram: https://www.instagram.com/aakashg0/ 🔑 Key Takeaways 1. Production Over Prototypes - Stop building prototypes and start shipping production-ready AI employees. Gibson AI, Cursor, and CrewAI let you go from concept to production in hours. Harish's agent was backed by a scalable database handling 10,000 users day one—no rebuilding required. 2. Amplify, Don't Replace - Your next 10x gain comes from making existing teams superhuman. AI agents analyze dashboards 24/7 and draft personalized outreach, while human CS agents focus on high-touch relationships and strategic decisions. 3. Three-Tier Implementation Strategy - Follow this roadmap: dashboard → human-approved recommendations → autonomous actions. Start with AI insights humans review, then AI recommendations humans approve, finally autonomous execution for low-risk tasks. 4. Human-Loop Insurance - Human-in-the-loop is customer relationship insurance. Harish built approval workflows because random AI emails "will only make the problem worse." AI should amplify human judgment, not bypass it. 5. Proactive Beats Reactive - Proactive churn prevention beats reactive win-back by orders of magnitude. AI agents monitor engagement patterns and usage metrics to address churn risks before customers consider leaving. 6. MCP Integration Magic - MCP makes AI tools actually talk to each other. Harish could query databases, update schemas, and deploy changes directly from Cursor—seamless integration without manual tool switching. 7. Information Processing Automation - Any role that "ingests information and sends out information" is automatable. SDRs, recruiters, executive assistants—if it involves processing data and taking action, AI handles the heavy lifting. 8. Specialized Model Selection - Different models excel at different tasks. Harish used O3 Mini for planning, Claude Sonnet for coding. Match your model choice to the specific job rather than defaulting to popularity. 9. Day-One Infrastructure - Production-grade infrastructure eliminates the prototype-to-production death valley. Starting with scalable database infrastructure means your demo can actually handle real user volumes when stakeholders want to scale. 10. Always Review Code - Read AI-generated code even when moving fast. Despite impressive capabilities, human oversight remains critical: "Make sure it is the code that you want." Speed matters, but understanding what you ship is non-negotiable. #ai #aiagents #agents 🧠 About Product Growth: The world's largest podcast focused solely on product + growth, with over 170K listeners. Hosted by Aakash Gupta, who spent 16 years in PM, rising to VP of product, this 2x/ week show covers product and growth topics in depth. 🔔 Subscribe and like the video to support our content! And turn on the bell for notifications.

Harish MukhamiguestAakash Guptahost
Jun 1, 20251h 2mWatch on YouTube ↗

CHAPTERS

  1. Goal and stack: build an AI customer success agent in three phases

    Harish lays out the end-to-end goal: an AI customer success “employee” that can detect churn risk and take actions without human involvement. He previews a three-part build (dashboard → agent recommendations → autonomous execution) and the core tools used across the demo.

  2. Planning the dashboard with o3-mini: tabs, metrics, and what to measure

    They use o3-mini in reasoning mode to design a customer success dashboard for a sample B2B e-commerce company (“Bulk Trade”). The model proposes four main dashboard tabs aligned to funnel, engagement, health, and retention outcomes.

  3. Defining the data model: translating dashboard needs into sources and fields

    After getting tab ideas, they ask o3-mini to specify what data fields are needed and then reorganize by data source. This becomes a practical “data definition” spanning CRM, analytics, and app/transactional data to power the dashboard and downstream agents.

  4. Building the Gibson database in the web UI: schema, ERD, and deployment

    Harish creates a Gibson project, chooses step-by-step modeling (since planning is done), and pastes the structured field definitions. Gibson generates an ER diagram and schema code, then deploys the database with dev and prod environments and instant CRUD APIs.

  5. Connecting Gibson to Cursor via MCP: database management from the IDE

    They move into Cursor and connect Gibson using MCP (Model Context Protocol), enabling schema management, deployment, and even natural language querying without leaving the IDE. This sets up an “infra-aware” coding workflow where Cursor can directly operate on the live backend.

  6. Generating realistic test data with Cursor + Gibson APIs

    Since the database is empty, Harish has Cursor generate a repeatable Python script to insert realistic test data using Gibson’s APIs while honoring table relationships. They install dependencies (e.g., Faker) locally and run the script to populate the deployed DB.

  7. Building the Next.js dashboard in Cursor (agent mode) using live Gibson data

    With data in place, Cursor is prompted to build a modern dashboard (Next.js + shadcn/ui) that uses Gibson APIs—not mock data. They discuss Cursor’s modes (ask vs agent), model choice (Claude Sonnet 3.7), and show the completed dashboard after the longer build process.

  8. Part 1 result: dashboard walkthrough (analytics, funnel, engagement, churn risk)

    They tour the finished dashboard: customer analytics and revenue trends, funnel conversion charts, engagement/retention metrics (NPS, active users, feature adoption), and a churn-risk view. The key limitation is highlighted: humans must still interpret insights and act.

  9. Part 2 setup: add an Agent Actions table from Cursor and redeploy schema

    To support human-in-the-loop workflows, they add a new table to store agent recommendations (customer info, churn risk/reason, proposed action, message). This demonstrates evolving the database schema directly from Cursor through Gibson MCP and redeploying it.

  10. Part 2 execution: CrewAI agents analyze churn and write recommended actions

    Harish introduces CrewAI’s structure (agents + tasks) and runs a multi-agent workflow: query/analyze data, diagnose churn reasons, propose mitigations, and ingest recommendations back into the agent_actions table. The dashboard then shows ready-to-use actions and drafted messages/emails.

  11. Part 3: making it autonomous—adding email sending and Jira ticket creation

    They extend the system to take actions automatically: send customer emails via SendGrid and create Jira tickets for product-related churn drivers. Cursor generates new tools/tasks/agents while preserving existing ones, and they prepare environment variables and accounts for execution.

  12. Autonomous run results: Jira tickets created and emails delivered

    After running the autonomous workflow, they verify outcomes in real tools: multiple Jira issues are created with labels indicating churn risk, and emails appear in the inbox with customer-appropriate outreach. This demonstrates an end-to-end loop from data → insight → action.

  13. Wrap-up: other AI employees, Gibson’s vision, and cautions on agentic code

    They discuss broader “AI employee” use cases (SDRs, exec assistants, meeting prep, recruiting) and reflect on Harish’s background (Microsoft, Apple/Siri, LeafLink) and motivation for building Gibson. The conversation closes with a vision for Gibson as a scalable backend for vibe-coded products and a reminder to review AI-generated code and add evals.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome