AI Enterprise - Databricks & Glean | BG2 Guest Interview

2025-12-23 - 45 min - source - Read full transcript

Apoorv Agrawal (host)Ali GhodsiArvind Jain

Key insights

A 95% AI project failure rate is a sign of healthy experimentation, not a red flag.

Both guests argue that if nearly all pilot projects were succeeding, it would mean companies weren't taking enough risk. The MIT statistic gets treated in the enterprise press as damning, but the guests reframe it as the expected cost of exploring a genuinely new technology.

ai-enterprise-adoption

Concrete, working enterprise AI use cases already exist across very different industries.

Examples given include Royal Bank of Canada compressing equity research report turnaround from 2 hours to 15 minutes, Merck's Teddy transformer model for gene-regulatory-network drug discovery, and 7-Eleven automating marketing segmentation and content generation - offered as evidence the technology delivers real economic value, not just demos.

ai-enterprise-adoption

LLMs themselves have become an interchangeable commodity.

Ali Ghodsi compares foundation models to gas stations: buyers just compare price and quality week to week, switching with no loyalty. He contrasts this with true platform lock-in (iPhone vs Android, Mac vs Windows), arguing no one has ever switched a core platform as casually as people now switch LLMs.

llm-commoditization

Durable competitive advantage comes from proprietary company data, not from the model.

Because any company can access roughly the same commodity LLMs, the argument is that what differentiates one company's AI systems from a competitor's is the unique data and business processes it feeds into those models - the 'secret sauce' that cannot be replicated by simply buying a better model.

data-as-moat

Agentic AI differs fundamentally from RPA because it learns and generalizes instead of following frozen rules.

RPA automations were rule-based and brittle: any unexpected input broke them, and fixing them meant manually rewriting rules. Agentic AI can pattern-match and improve, though the guests note today's systems are still 'frozen' after training and don't continuously learn from live use, which they flag as the next hard problem to solve (e.g. for computer-use agents).

ai-enterprise-adoption

Both CEOs claim, by decades-old definitions, that AGI already exists.

Ali Ghodsi recalls that the AGI bar used in his UC Berkeley AI lab circa 2009 has already been cleared by current LLMs, and argues the goalposts have simply moved. His conclusion: rather than waiting for AGI, the industry should focus on expanding AI's enterprise usage from roughly 5% of tasks toward 100%.

agi-debate

The AI industry splits into three camps with different theories of value.

Camp one is the superintelligence-quest labs betting everything on scaling laws and bigger GPU clusters. Camp two is scientists (e.g. Rich Sutton, Yann LeCun) who argue current architectures are fundamentally the wrong approach and true AGI is ~20 years out. Camp three - where both guests place themselves - argues today's models are already good enough to extract enormous economic value through better engineering, without needing superintelligence.

agi-debate

Justifying current AI capex requires roughly $1 trillion in new AI revenue, but the guests argue this isn't just marginal software growth.

With about $250B of the roughly $500B AI capex flowing through Nvidia, and total software industry revenue near $400B, the numbers look impossible under a pure software-expansion lens. The guests argue AI is instead going to capture spend from the services industry, which is roughly 25x larger than the software industry, making the revenue target more plausible.

ai-capex-bubble

There is a real bubble, but it's concentrated in a narrow slice of the market, not the technology itself.

Ali Ghodsi points to pre-revenue AI startups valued at $10-30 billion as clear bubble evidence, while distinguishing this from the broader claim that AI adoption or spend is unjustified. He frames AI company valuation premiums over non-AI companies as partly rational given AI companies' faster expected growth.

ai-capex-bubble

Enterprise software will not collapse into a bare database with AI-generated UIs on top.

Arvind Jain pushes back on the idea (attributed to Satya Nadella's 'crud apps' framing) that software is becoming just a database layer. He argues that most users don't actually know what UI/workflow they want, and that software companies' real value is designing how data gets presented and acted on - meaning the application layer, not just the data layer, keeps most of its value.

data-as-moat

Internal AI automation projects often fail from organizational friction, not from AI capability gaps.

Ali Ghodsi describes Databricks' early attempt to automate software engineering failing because of how the org was structured, not because the AI was bad. Arvind Jain separately describes trying to build an agent that would auto-assign and track every employee's weekly priorities - a project that stalled not on technical grounds but because coordinating real organizational alignment is inherently hard, even with full context.

ai-enterprise-adoption

The next phase of enterprise AI is proactive - AI that comes to the user rather than waiting to be asked.

Arvind Jain's vision for Glean's future is a fully personal, privileged AI companion that already knows a person's calendar, goals, and weaknesses and starts working on tasks before being asked, rather than requiring the user to go to a product to get answers. He frames closing this gap as what will move AI from 5% power users to near-universal usage.

ai-enterprise-adoption

Media referenced

MIT study on AI deployment failure rates - paper - Cited for the finding that 95% of enterprise AI deployments do not work; both guests reframe this as normal experimentation variance rather than a failure signal.

Companies

Databricks - Ali Ghodsi's company; data and AI platform for enterprises.
Glean - Arvind Jain's company; enterprise AI search/agent platform, cited as crossing $200M ARR.
Altimeter - Investment firm employing host Apoorv Agrawal; referenced for its 'AI strategy starts with your data strategy' framing.
Royal Bank of Canada - Databricks customer; built an agent that compresses equity research report turnaround from 2 hours to 15 minutes.
Merck - Databricks customer in life sciences; built a transformer model called Teddy for gene-regulatory-network drug discovery.
7-Eleven - Databricks customer; uses agents to automate audience segmentation and marketing content generation.
OpenAI - Discussed in the rapid-fire round; both guests predict its stock/revenue is up in 12 months, driven by ChatGPT growth.
Anthropic - Discussed in rapid-fire round; predicted up given expanding coding market share.
Salesforce - Used as the example of enterprise software being called a 'database' or 'crud app' by critics; guests push back on this framing.
ServiceNow - Named alongside Salesforce as legacy enterprise software facing the 'software is becoming a database' critique.
Zoom - Cited as a company well positioned to become the 'data entry' layer for enterprise AI by capturing meeting conversations.
TSMC - Analogy for how LLM providers may become valuable but interchangeable 'fab-like' commodity suppliers.
Nvidia - Referenced for the roughly $250B being spent on AI chips/capex that needs to be justified by AI revenue.
Microsoft - Satya Nadella's 'crud apps' framing of enterprise software is referenced and debated.
Grain - AI note-taking tool Arvind Jain has personally used.
Otter.ai - AI note-taking tool Arvind Jain has personally used.
Fathom - AI note-taking tool Arvind Jain has personally used.

Techniques and frameworks

RPA (Robotic Process Automation) - Used as the contrast case to agentic AI - rule-based, brittle, and non-learning, versus AI's ability to generalize and improve.
Reinforcement learning - Referenced as the foundational technique behind much of modern AI, credited to Turing Award winner Rich Sutton.
Crawl, walk, run - Framing offered for how enterprises should roll out AI budgets: short vendor contracts, easy-to-test products, incremental scaling.

Summary

Altimeter's Apoorv Agrawal sits down with Databricks CEO Ali Ghodsi and Glean CEO/founder Arvind Jain for an operator-level conversation on what is and isn't working in enterprise AI. Both guests reject the doom reading of the widely cited MIT statistic that 95% of AI deployments fail, arguing that near-universal failure in early experimentation is the expected cost of exploring genuinely new technology, not evidence the technology doesn't work. They back this with concrete customer examples spanning finance (Royal Bank of Canada cutting equity research turnaround from 2 hours to 15 minutes), healthcare (Merck's Teddy transformer model for gene-regulatory drug discovery), and retail (7-Eleven's automated marketing segmentation) - real production use cases rather than demos.

A recurring thread is that LLMs themselves have become a commodity: Ghodsi compares them to gas stations, where buyers simply compare price and quality and switch weekly with no loyalty, a level of platform indifference he says has no precedent in prior tech cycles (unlike iPhone-vs-Android or Mac-vs-Windows loyalty). Because the model layer is commoditizing, both argue the actual moat is a company's proprietary data and unique business processes - and Ghodsi even uses Glean itself as the example, noting that Glean stripped of its customers' data would have no value. This leads into a debate about where value accrues across the data, model, and application layers; Jain pushes back hard on the idea that AI reduces enterprise software to a bare database with AI-generated UI on top (a framing he attributes to Satya Nadella's "crud apps" comment), arguing most users don't actually know what interface or workflow they want, so software companies' design work remains valuable.

The conversation pivots to comparing today's AI wave with RPA, the last enterprise automation hype cycle: both agree the fundamental difference is that RPA was rule-based and brittle with zero learning, while agentic AI can generalize and improve, though they concede current systems still "freeze" after training rather than continuously learning from live use. On the AGI question, both guests claim - somewhat provocatively - that by the definitions used in AI research circles as far back as 2009, we already have AGI, and that the debate today is really about goalposts moving rather than capability gaps. They sketch three industry camps: frontier labs chasing superintelligence through scale, academic skeptics (Rich Sutton, Yann LeCun) who think today's architecture is fundamentally wrong and true AGI is ~20 years out, and the pragmatic camp both guests place themselves in, which argues current models are already sufficient to extract enormous enterprise value through better engineering.

On the capex question, they acknowledge the math looks extreme - roughly $1 trillion in new AI revenue is needed to justify current spend against a software industry that only generates about $400 billion today - but argue AI is capturing share of the far larger $10 trillion-plus services industry rather than just expanding software spend. They do concede a real bubble exists, but locate it specifically in pre-revenue startups carrying $10-30 billion valuations, distinguishing that from the underlying technology adoption, which they see as durable. Both also share personal AI usage: Ghodsi describes agents that prep customer talking points and automate go-to-market research at Databricks, while Jain describes a "daily prep agent" and a personal shift toward asking Glean before pulling a team together to answer a question. The episode closes on Jain's vision for Glean's future as a fully proactive, privileged personal AI companion that surfaces and starts work before being asked, which he frames as the shift needed to take AI from 5% power-user adoption to near-universal use.

Notable Quotes

"I think we have AGI. I think we have artificial general intelligence. We really have it." - Ali Ghodsi

"The LLM is a commodity. People are not saying that, but it is a commodity. Like you can get gas from this gas station, you can get gas from that gas station, it doesn't matter. Just compare price." - Ali Ghodsi

"There are startups with zero revenue worth you know, 10, 20, 30 billion. That's a bubble." - Ali Ghodsi

"The products that are going to change the paradigm - instead of you building a product and expecting people to come to you, if you understand your customer very deeply and actually bring the AI to them. That's the category that I'm excited about." - Arvind Jain