In Conversation with Sandipan Bhaumik, Lead Data & AI Solutions Architect, Databricks

Sandipan Bhaumik (Sandi) is a Lead Data & AI Solutions Architect at Databricks with over 18 years of experience in data engineering, data governance, machine learning, and AI. Previously at AWS, he has led data and AI transformations for enterprises across financial services, healthcare, and technology sectors.

As the founder of the AgentBuild community and newsletter, Sandi focuses on practical AI implementation for enterprise teams, helping organizations move from AI demos to measurable business outcomes. 

Sandi specializes in translating complex technical concepts into actionable frameworks for executive audiences. His work focuses on organizational readiness assessment, production AI program design, and scaling AI capabilities across multiple use cases. He regularly speaks at industry conferences on topics including AI strategy, data and AI governance, and enterprise architecture patterns for AI systems.

 Ready to explore what happens when AI systems are designed for trust, conversational intelligence is grounded in real-world context, and leadership becomes an exercise in translation between technical depth and business reality?

Let’s dive into Sandi’s perspective on building AI that people actually use, and organizations can truly scale.  

1. You've spent nearly two decades helping organizations make data and AI actually work. When you look at today's AI-driven communication, especially conversational AI and agentic platforms, what do you think separates meaningful automation from mere noise?

The difference is evaluation. Most conversational AI gets deployed without anyone defining what success actually looks like. You end up with chatbots that respond quickly but don't solve problems. Meaningful automation starts with the question: "What decision am I trying to improve?" Then you build backward from there. If you can't measure whether the conversation created value-faster resolution, better understanding, actual progress-you're just automating noise. The best systems make people feel heard while moving things forward, not just generating responses.

2. You often talk about governance being "at the core". In conversational AI, where messages and agents interact directly with customers, what does governed communication look like? How do we scale trust without slowing innovation?

Governed communication means your agents can explain their decisions and you can audit them later. It's not about locking everything down-it's about knowing what your AI said, why it said it, and having the ability to intervene when it matters. Think of it like your semantic layer for conversations: clear definitions, traceable logic, and the ability to quickly adjust when business context changes. You scale trust by making the AI's reasoning transparent, not perfect. People forgive mistakes when they understand what happened.

3. You've helped build massive data systems, hundreds of terabytes to petabytes. What's one leadership lesson from those large-scale data challenges that applies equally to scaling human conversations or organizational culture?

At scale, your biggest problems aren't technical-they're communication breakdowns. When you're moving petabytes, someone always has a different definition of "customer" or "transaction," and those misalignments compound. The same thing kills organizational culture. The fix? Obsessive clarity about definitions and relentless overcommunication about changes. Whether you're building data pipelines or team dynamics, most failures trace back to people thinking they agreed when they actually didn't. Document everything, especially the boring stuff.

4. In your view, how is the transition from traditional data pipelines to agentic communication systems (like SMS automation or AI-driven chat experiences) reshaping how businesses think about customer engagement?

Traditional pipelines assumed you controlled the flow-data moved when and where you told it to. Agentic systems flip that: the AI decides when to act based on context. For customer engagement, this is massive. Instead of "send email at 10am Tuesday," you get "reach out when the customer shows buying intent." Businesses are shifting from scheduled campaigns to contextual conversations. The challenge? Your data architecture has to support real-time decisions, not just batch reporting. Most companies aren't ready for that transition.

5. Many brands and tech leaders talk about "building community", but you've actually engineered one, from mentorship workshops to global study cohorts. What have you learned about turning passive audiences into active participants? And what advice would you give to business leaders who want to cultivate communities not just for marketing, but for genuine collaboration and shared growth?

People become participants when they see others like them succeeding. In AgentBuild, the breakthrough was showing beginners building real projects in weeks, not months. That peer proof matters more than any expert endorsement. For business leaders: stop trying to "engage" your community and start celebrating their wins publicly. Create small opportunities for people to contribute, then amplify those contributions. Communities don't grow from consumption-they grow from creation. Let people build together, not just learn together.

6. At AWS, you helped teams build "cultures of experimentation". What advice do you have for marketing or brand leaders who want to foster that same spirit when experimenting with conversational AI or automation?

Start small, measure everything, and celebrate intelligent failures. At AWS, our best experiments came from teams who defined success metrics before they wrote code. For conversational AI, that means testing one use case thoroughly before scaling. Try automating appointment reminders before you rebuild customer service. The key is creating psychological safety to try things that might not work, while being ruthless about learning from what fails. Document what you learn, not just what worked.

7. AgentBuild.ai helps people learn to build AI agents. What's one emerging capability of agents that excites you most, especially in the context of communication, customer experience, or even brand storytelling?

Multi-agent orchestration for storytelling. Imagine agents that don't just respond to customers but collaborate to craft narratives-one agent understands brand voice, another knows product details, a third tracks emotional tone. Together, they create coherent, personalized experiences that feel human. We're moving from single-bot conversations to coordinated teams of specialists. For brands, this means richer, more contextual engagement without losing consistency. The magic happens when agents hand off seamlessly, like a great improv troupe.

8. If we meet again in two years, what do you think we'll be talking about in the world of conversational AI and data intelligence that we aren't talking about enough today?

We'll be talking about agent debt-the liability from deploying AI systems without proper evaluation frameworks. Just like technical debt slows development, agent debt will cripple organizations that deployed chatbots and automation without measuring business impact. Companies will realize they have hundreds of agents making decisions they can't audit or improve. We'll need "agent archeologists" to excavate what these systems actually do. The winners will be organizations that built evaluation-first from day one, not those who moved fastest. 

 Rapid fire:

1. First thing you'd automate if you could?

Meeting follow-ups. Every conversation ends with "I'll send you that link" or "Let me circle back on X." I want an agent that listens, captures commitments, and actually closes loops without me remembering three days later.

2. Most underrated skill for future data leaders?

Storytelling. You can build the perfect architecture, but if you can't explain why it matters to someone who doesn't care about technical elegance, it won't get funded or adopted. Translation beats optimization.

3. If your career were a dataset, what would the schema say?

bridge_builder | source: technical_depth | target: business_reality | transformation: translation | quality_metric: did_anyone_actually_use_it

4. Favorite analogy you've ever used to explain AI to a non-tech audience?

AI is like hiring an intern who's read everything but has no judgment. They can find patterns and generate responses instantly, but they'll confidently tell you the wrong thing if you don't check their work. You still need a manager.

5. You can invite three people to a 'Data & Drinks' night, who's coming?

Someone who's failed spectacularly at AI deployment and will tell the truth about it. A designer who thinks about human behavior, not just interfaces. And a CFO who's killed projects for good reasons - I want to understand the "no" side.

6. Most human thing about building AI?

Realizing the technology is rarely the problem. It's always the people stuff - misaligned incentives, unclear ownership, fear of being replaced, teams that don't trust each other. You're not debugging code, you're debugging organizations.

(Thank you, Sandi, for the candid and grounded conversation. Your take on evaluation-first AI and governance as a trust builder was a refreshing reminder that real progress isn’t about hype, it’s about decisions that actually move things forward. We appreciate your honesty, your practical lens, and the way you keep both people and outcomes at the center of building AI that truly works.)

Databricks at a Glance:

Databricks is the Data and AI company, helping organizations turn data into intelligence at scale. More than 20,000 companies worldwide, including adidas, AT&T, Bayer, Block, Mastercard, Rivian, Unilever, and over 60% of the Fortune 500, use Databricks to build, deploy, and govern data, analytics, and AI applications. Headquartered in San Francisco with offices across the globe, Databricks provides a unified Data Intelligence Platform, bringing together capabilities such as Agent Bricks, Lakeflow, the Lakehouse, Lakebase, and Unity Catalog to power everything from real-time analytics to production AI and intelligent agents.

‍