AI Enterprise - Databricks & Glean | BG2 Guest Interview

Discussion of the MIT report showing 95% of AI deployments fail, but this is actually a positive sign of healthy experimentation. The speakers explain that failure rates indicate companies are trying enough new approaches, and that AI is being used extensively in both personal and work contexts despite enterprise deployment challenges.

• •95% failure rate is expected and desired - indicates sufficient experimentation with new technology
• •AI adoption is already widespread in personal and work lives (ChatGPT, Claude usage)
• •Success requires treating AI implementation as an engineering art, not just quick deployment
• •Differentiation comes from proper evaluation, productionization, and great teams - not just using AI

Concrete examples of successful AI deployments across finance, healthcare, and retail. Royal Bank of Canada reduced equity research report time from 2 hours to 15 minutes. Merck created TEDDI for drug discovery using transformer models. Seven Eleven automated their entire marketing stack with fine-grained audience segmentation.

• •Finance: RBC's agent analyzes earnings reports, competitor data, and market news to produce equity research in 15 minutes vs 2 hours industry standard
• •Healthcare: Merck's TEDDI (Transformer Enabled Drug Discovery) predicts gene regulatory networks for drug discovery
• •Retail: Seven Eleven automates marketing with agents that segment audiences and create customized content at scale
• •Success pattern: Leverage proprietary company data and unique business processes, not commodity AI capabilities

Ali argues that LLMs have become completely commoditized - interchangeable like gas stations. Unlike traditional tech religious wars (iPhone vs Android, Excel vs Google Sheets), users freely switch between LLMs based on price and slight performance differences. The real value lies in proprietary company data and unique business processes.

• •LLMs are commodities: interchangeable, price-driven, users switch models weekly without loyalty
• •Durable advantage comes from proprietary data your competitors don't have
• •Building AI that understands your unique data and business processes is not commodity work
• •Demo-ware problem: Easy to create impressive demos, but most companies build non-differentiated commodity features
• •Governance and security layers become critical when agents access sensitive company data

Arvind shares Glean's engineering failures, particularly around fine-tuning models for specific use cases. They discovered it's more effective to use existing open-source models on Databricks or foundation models rather than building custom models. Internal automation projects also take much longer than expected despite AI capabilities.

• •Fine-tuning work for specific product use cases didn't pan out - better to use existing models
• •Engineering ideas can become obsolete within two weeks due to rapid AI developments
• •Internal business process automation (like priority tracking across the company) still takes significant time despite AI
• •AI is one more tool in the toolkit - doesn't magically make complex enterprise systems easy to build overnight

Ali frames the AI landscape into three camps: (1) Super intelligence quest (frontier labs spending massive capital on scaling laws), (2) Sober researchers saying it will take 20 years with different approaches, and (3) Practical builders focused on extracting economic value from existing AGI capabilities. He argues we already have AGI by historical definitions.

• •Camp 1 (Super Intelligence): Frontier labs betting on scaling laws, massive GPU clusters, pursuing godlike AI that can cure cancer and 10x GDP
• •Camp 2 (Researchers): Nobel Prize winners like Rich Sutton and Yann LeCun say autoregressive next-token prediction is wrong approach, 20 years out
• •Camp 3 (Practical Builders): Focus on making current AGI useful in enterprises without needing super intelligence
• •We already have AGI by 2009 standards - the goalposts keep moving as capabilities improve
• •The bubble exists in Camp 1's massive capital spend, not in Camp 3's practical value creation

Discussion of where value will accrue in the AI stack. Models are commoditized like TSMC fabs - valuable but interchangeable. The intelligence layer may capture half of enterprise value. Most value will accrue to applications, but like the internet era, we don't yet know which specific apps will dominate.

• •LLMs are commodities but companies can still be valuable (like TSMC in semiconductors)
• •Intelligence layer expected to be thick, potentially capturing 50% of enterprise value
• •Most value will accrue to applications, but specific winners are unknown (like internet era before Facebook, Uber emerged)
• •Data governance and security layers become critical for agent-based systems
• •Existing companies like Amazon and Google from 1998 survived - not binary disruption

Debate on whether traditional SaaS like Salesforce will become 'just databases' with AI-generated UIs. Arvind argues this oversimplifies - software companies provide full ecosystems of workflows. The real shift is from manual data entry via screens to conversational interfaces, with Zoom positioned as ideal data entry application.

• •Salesforce is not 'just a database' - it's a full ecosystem of workflows and applications
• •Software companies add value by designing how humans naturally interact with and modify data
• •Data entry paradigm shifting from typing to voice: 'Hey chat, met with Arvind, remind me in 2 days'
• •Zoom well-positioned as data entry app - captures all conversations, can extract and store information in systems of record
• •Meeting recording agents already updating Salesforce notes automatically from customer conversations

Both CEOs share how they use AI personally and drive organizational adoption. Ali uses agents for customer insights, competitive intelligence, and meeting prep. Arvind's 'daily prep agent' provides meeting context and changed his instinct from asking people to asking AI first. Heavy automation in go-to-market (6,000 people) and R&D (4,000 people) orgs.

• •Customer intelligence agents (Rafi at Databricks) provide instant use cases, competitive insights, and meeting prep
• •Daily prep agents deliver morning briefings with meeting context, reading materials, and day planning
• •Marketing stack heavily automated - content creation, audience segmentation, campaign management
• •Engineering automation required change management - initial attempts failed due to human organization issues, not AI limitations
• •Finance moved from Excel to Python-based ML forecasting after external data science team proved value
• •Recording every meeting (internal and external) creates valuable knowledge base for AI agents

Rapid-fire predictions on AI company valuations and market dynamics. Both agree there is a bubble in super intelligence quest camp and startups with zero revenue worth billions. OpenAI and Anthropic predicted to grow. Long on agents and speech interfaces, short on coding automation hype and customer service automation claims.

• •Bubble exists in super intelligence camp and zero-revenue startups worth $10-30B, not in practical value creation
• •OpenAI and Anthropic stocks/revenue predicted up in 12 months due to ChatGPT growth and coding market expansion
• •Long: Agents, speech interfaces (keyboards will disappear), proactive AI products that come to users
• •Short: Coding automation somewhat overhyped, customer service automation hasn't made as much progress as claimed
• •AI companies deserve higher multiples than non-AI companies due to superior growth potential

Arvind outlines Glean's vision to become a personal companion for every person in every company. The companion will have privileged access to all work context, know weekly goals and career ambitions, and proactively complete tasks before being asked. Goal is moving from users coming to Glean to Glean coming to users.

• •Vision: Personal AI companion with fully privileged, confidential relationship with each worker
• •Companion knows everything about work life: schedule, meetings, goals, strengths/weaknesses, career ambitions
• •Shift from reactive (user comes to Glean) to proactive (Glean comes to user and completes tasks automatically)
• •Current challenge: High variance between power users and casual users - goal is 100% adoption through proactive AI
• •Foundation already in place at $200M ARR with $10M+ deals, focus now on AI literacy and universal adoption