
Lessons from Arsalan Tavakoli
Arsalan Tavakoli is a Databricks co-founder and SVP of Field Engineering. His background includes a computer science PhD at UC Berkeley and work at McKinsey. These lessons cover enterprise data, AI implementation and customer-facing technical work. — Databricks Author Biography.
Part 1: The Transition from Academia to Industry
- On Academic Roots: Doctoral training taught Tavakoli to define ambiguous problems, identify constraints and learn through failed approaches—not just pursue technically interesting questions. — NextGenLakehouse Interview.
- On McKinsey Lessons: After McKinsey, Tavakoli’s early startup responsibilities ranged from pricing and marketing to business development and contributing code. — NextGenLakehouse Interview.
- On the Early Days: Early Databricks was focused on surviving its immediate challenges; Tavakoli did not describe its eventual scale as something the founders could confidently foresee. — Summit Live 2026.
- On Taking Risks: Tavakoli accepted startup risk because he trusted the people and technology, and saw valuable learning even if the venture failed. — NextGenLakehouse Interview.
- On Building the Core Team: The early team paired technical founders with commercial work and customer-facing solution architects; specialized sales leadership came later. — NextGenLakehouse Interview.
- On Defining the Problem: Databricks initially delivered a narrower product around notebooks and data work, then expanded to meet customers’ broader needs. — Summit Live 2025.
Part 2: The Fallacy of AI Token-Maxing
- On Capital Misallocation: A large portfolio of AI experiments is not evidence of business value; connect each initiative to a meaningful operating problem. — Enterprise AI: From Pilots to Production.
- On the AI Hype Cycle: Tavakoli cautions against adopting AI from fear of missing out; begin with the business problem rather than fashionable technology. — Enterprise AI: From Pilots to Production.
- On Measuring ROI: Define AI success by improving an existing process or enabling a new capability—not by counting deployed tools or agents. — Mastercard Interview.
- On Proof of Concepts: A pilot can demonstrate a tool without delivering its full benefit; production value may require redesigning the surrounding process. — Enterprise AI: From Pilots to Production.
- On Model Obsession: Enterprise answers need company-specific definitions and relationships; access to a powerful model alone does not supply that business context. — The Rundown Interview.
- On Sustainable AI: Match model capability to the task and balance quality with cost; routine questions do not always require the strongest reasoning model. — The Rundown Interview.
Part 3: Dashboards and Conversational Analytics
- On the Limits of BI: Dashboards remain useful visual references, but their fixed views can leave users waiting for answers to follow-up questions. — Agents Are Ready; Your Architecture Probably Isn’t.
- On Democratizing Data: Conversational data tools should let business users answer company-specific questions themselves, instead of relying on an analyst for every investigation. — The Rundown Interview.
- On Text-to-SQL Limitations: In Tavakoli’s coauthored platform essay, natural-language access depends on understanding an organization’s terminology and data—not merely adding a conversational interface. — What Is a Data Intelligence Platform.
- On the Shift to Agents: Agents create transactional as well as analytical workloads; low-latency actions need supporting infrastructure, not just another dashboard. — Agents Are Ready; Your Architecture Probably Isn’t.
- On Dashboard Fatigue: Building a dashboard for each narrow question can leave a collection of little-used reports; design for follow-up exploration instead. — Agents Are Ready; Your Architecture Probably Isn’t.
- On Real-Time Needs: Data freshness must suit the decisions agents are making; slow batch cycles can limit systems expected to respond quickly. — Agents Are Ready; Your Architecture Probably Isn’t.
Part 4: Data Context as the AI Bottleneck
- On the Real AI Challenge: After choosing a model, enterprises still face security, evaluation and data-quality problems; useful answers require relevant internal information. — TBPN Interview Transcript.
- On Unstructured Data: Tavakoli’s coauthored essay traces growing AI use cases beyond structured warehouse data, making support for unstructured data an important architectural concern. — What Is a Data Intelligence Platform.
- On Generic Models: Model selection is only part of enterprise AI. Data quality, governance, production readiness and measurement also determine whether an implementation works. — Mastercard Interview.
- On Data Governance: Do not automatically give an agent the same authority as a person; actions need explicit permissions and safeguards. — Enterprise AI: From Pilots to Production.
- On the Value of Proprietary Data: Tavakoli sees enterprise data, trust and governance as sources of defensibility—not a claim that proprietary data is the only remaining moat. — Enterprise AI: From Pilots to Production.
- On Data Silos: A unified data platform can reduce the governance and reliability complications of keeping warehouse and lake workloads in separate systems. — What Is a Data Intelligence Platform.
- On Continuous Updating: Keep enterprise context current: Tavakoli warns that previously useful data may become invalid, so systems need to distinguish relevant, recent information. — TBPN Interview Transcript.
- On Context Windows: The coauthored platform essay describes retrieval-augmented generation as a way to bring relevant enterprise data into AI applications. — What Is a Data Intelligence Platform.
- On Semantic Understanding: Data intelligence should use contents, metadata and usage patterns to understand organizational terminology, metrics and relationships. — What Is a Data Intelligence Platform.
Part 5: Enterprise AI and Legacy Migrations
- On Migration Timelines: Tavakoli’s coauthored Lakebridge announcement describes automation across assessment, conversion and validation to reduce the work involved in migrations. — Introducing Lakebridge.
- On Technical Debt: Inventory legacy objects and dependencies before migrating; their complexity helps teams prioritize business impact and plan the work in phases. — Introducing Lakebridge.
- On Automated Code Translation: Configurable conversion tools can translate legacy SQL and ETL workloads, while automated query testing identifies discrepancies that still need correction. — Welcoming BladeBridge to Databricks.
- On De-risking Migrations: Conversion is not enough: reconcile migrated data and validate results so discrepancies are visible before relying on the new system. — Introducing Lakebridge.
- On Cost Savings: Tavakoli’s coauthored migration announcement presents cost and operating complexity as reasons to modernize; it does not establish savings for every migration. — Welcoming BladeBridge to Databricks.
- On Legacy Lock-in: Open formats and conversion tooling can reduce dependence on proprietary systems, but migration still requires assessment and validation. — Introducing Lakebridge.
- On Data Lakes: The coauthored Delta Lake essay describes combining reliable data-lake storage with streaming, machine-learning and warehousing workloads. — Open Sourcing All of Delta Lake.
- On Unifying Data: Unified governance should cover tables, files and AI assets together, rather than leaving them under disconnected access and audit systems. — Open Sourcing Unity Catalog.
Part 6: Navigating Hypergrowth at Databricks
- On Scaling Revenue: Tavakoli describes proactively reducing customers’ wasted spending to build trust and encourage additional workloads—not extracting the maximum bill immediately. — The Rundown Interview.
- On Maintaining Culture: Preserve culture through hiring, promotions and recognition. Tavakoli treats these choices as deliberate work, not something growth will take care of. — Summit Live 2026.
- On Product Focus: Keep conviction in the product’s direction instead of copying every competitor announcement, while remaining willing to recognize when a pivot is needed. — Enterprise AI: From Pilots to Production.
- On Hiring: Hypergrowth keeps changing the work. Tavakoli tells new colleagues not to expect the pace of learning and change to settle down. — Summit Live 2026.
- On Customer Feedback: Use a focused pilot to learn quickly and establish what works before broadening enablement across the organization. — Agents Are Ready; Your Architecture Probably Isn’t.
- On the Platform Play: Tavakoli’s platform approach connects analytics, transactional systems, agent development and monitoring with the applications people actually use. — Agents Are Ready; Your Architecture Probably Isn’t.
Part 7: The Future of Software Monopolies
- On Pricing Pressure: Tavakoli predicts that easier software creation will challenge incumbent monopolies; this is a competitive forecast, not a guaranteed deadline. — Enterprise AI: From Pilots to Production.
- On the Cost of Code: AI-assisted development can lower the effort needed to create software, opening markets to more competitors without making entry costless. — Enterprise AI: From Pilots to Production.
- On Defensibility: Systems of record are more than database wrappers: accumulated compliance and operating capabilities can make them harder to displace. — Enterprise AI: From Pilots to Production.
- On Margin Compression: Tavakoli expects lower entry barriers and more software competition to put pressure on margins; he does not present this as a measured result. — Enterprise AI: From Pilots to Production.
- On Value Creation: Tavakoli predicts substantial AI value at the application layer, where enterprises turn models and data into useful capabilities. — Mastercard Interview.
- On Custom vs. SaaS: Tavakoli prefers buying routine software when it fits; build when important needs remain unmet, and account for maintenance—not just creation. — TBPN Interview Transcript.
Part 8: Field Engineering and Customer Success
- On the Role of Field Engineering: Field engineering brings technical resources together around the customer’s journey, rather than asking customers to navigate the vendor’s internal organization. — Summit Live 2025.
- On Technical Sales: Customer-facing technical teams need hands-on prototyping ability and the communication skills to translate business problems into technical work. — TBPN Interview Transcript.
- On Customer Empathy: Look for frustrating everyday workflows, such as claims processing and onboarding, where AI can help people get useful work done. — Mastercard Interview.
- On Time-to-Value: Help customers reach useful outcomes sooner by addressing migration, people and process together—not only the technology. — Summit Live 2025.
- On Proof of Concepts: A delivered solution should have defined success criteria and components customers can control and evolve, rather than remaining a black-box dependency. — Summit Live 2026.
- On Feedback Loops: The coauthored Delta Lake account shows how solving a customer’s reliability and workload problems helped reveal needs shared by other customers. — Open Sourcing All of Delta Lake.
- On Long-Term Success: Customer success continues after launch through enablement, training and further adoption; field engineering supports that continuing journey. — Summit Live 2026.