AI Enterprise - Databricks & Glean | BG2 Guest Interview

Episode Summary

Executive Summary: The conversation argues that AI is already delivering real enterprise value, but not through generic models alone. The speakers say LLMs are becoming commoditized; the winners will be companies that combine AI with proprietary data, workflow integration, governance, and change management. They debate AGI, bubbles, and AI’s economics, while emphasizing practical use cases in finance, healthcare, retail, GTM, and internal productivity.

Main Topics: AI is already useful in enterprise, but success is uneven (Priority: 5/5): The speakers reject the idea that AI is failing broadly; they argue the 95% failure rate reflects experimentation and that real value is emerging in specific workflows where AI is productionized. LLMs as a commodity (Priority: 5/5): Both speakers frame foundation models as interchangeable infrastructure, shifting value away from model ownership toward proprietary data, security, and application design. Use cases that are working today (Priority: 5/5): Examples include RBC earnings analysis, Merck drug discovery, retail marketing automation, and internal productivity/meeting automation, showing where AI is already saving time and improving outcomes. Why AI is different from RPA (Priority: 4/5): They contrast brittle, rules-based robotic process automation with learning-based AI systems that can generalize and adapt, arguing this makes AI structurally more powerful. AI economics, capex, and the bubble debate (Priority: 4/5): The speakers acknowledge a bubble in parts of the market, especially speculative superintelligence bets, but argue practical enterprise AI is not the same as bubble-like overinvestment. Value accrual across the AI stack (Priority: 4/5): The discussion examines where value will land across data, intelligence, and application layers, with a consensus that applications and data ownership will matter most. The future of work, software, and proactive AI (Priority: 4/5): They envision AI assistants that come to users, not vice versa, with speech interfaces, note-takers, and proactive agents gradually replacing keyboard-driven software interactions.

Key Arguments: High AI deployment failure rates do not necessarily signal weakness; they may indicate healthy experimentation and a still-maturing market. LLMs are commoditized infrastructure, so competitive advantage comes from unique company data, workflows, and secure access, not from the model itself. Real enterprise value comes when AI is engineered, evaluated, and productionized around a specific business process, not when it is only demoed. AI differs from RPA because it learns and generalizes, whereas RPA was brittle, static, and rule-bound. AI spending is partly a reallocation of service dollars into software/automation rather than purely new spend, which helps explain the huge market opportunity. The bubble exists mainly in frontier superintelligence capital and some zero-revenue startups, not necessarily in practical enterprise deployments. Future winners will likely be companies that solve data entry, workflow coordination, and knowledge extraction across meetings, documents, and systems of record. Glean’s strategy is to become a personal work companion that proactively helps every employee, increasing AI literacy and moving from pull to push interaction.

Data Points: AI deployment failure rate: 95% - Referenced from the MIT report; speakers used it to argue experimentation is healthy. Royal Bank of Canada earnings report turnaround: 15 minutes - AI agent reduced equity research report production from a standard 2 hours to 15 minutes. Standard earnings report turnaround: 2 hours - Baseline for the RBC equity research workflow. AI revenue needed to justify current capex: ~$1 trillion - Investor framing around NVIDIA and total AI infrastructure spending. Semiconductor spend referenced: $250 billion - Quarter-trillion spent on NVIDIA/semis used as part of the capex argument. Implied total AI capex: ~$500 billion - Assuming semis are half of capex, total spend was estimated at half a trillion. Total software industry revenue: ~$400 billion - Used as comparison against the scale of AI revenue needed. Services industry size relative to software: 25x larger - Argument that AI can capture revenue currently spent on services, not just software. Databricks revenue run rate: $200 million - Mentioned during the closing vision segment for Glean’s growth context. Large GTM organization size at Databricks: ~6,000 people - Illustrates scale of AI-driven automation across go-to-market teams. R&D organization size at Databricks: ~3,000-4,000 people - Used to describe internal automation and workflow adoption. AI startup valuations mentioned: $10B-$30B - Examples of zero-revenue or low-revenue companies valued at bubble-like levels.

Pivotal Quotes: "I think we have AGI. I think we have artificial general intelligence. We really have it." — Ali: Opening claim in the discussion about where current AI capabilities stand. "The LLM is a commodity." — Ali: Used to explain why value will shift away from model ownership toward data and applications. "That’s actually what you want." — Arvind: On the reported 95% AI project failure rate, arguing experimentation should be expected.

Implications: The message for enterprises is to stop chasing generic demos and focus on proprietary data, workflow redesign, governance, and rapid experimentation. The next wave of value likely comes from AI-native applications, proactive assistants, and speech-based interfaces rather than model ownership alone.