No PriorsNo Priors Ep. 115 | With Glean Founder and CEO Arvind Jain
CHAPTERS
- 0:00 – 0:35
Arvind Jain’s background and what Glean does
Elad and Sarah introduce Arvind Jain and position Glean as an AI-powered enterprise search and knowledge management platform. They frame the conversation around Arvind’s deep search background (Google) and repeat-founder experience (Rubrik).
- •Glean unifies access to internal docs, Slack, and other workplace systems
- •Productivity layer: applications/automation on top of company knowledge
- •Arvind’s prior work: early Google search + co-founding Rubrik
- •Setup for discussion on how LLMs reshape enterprise search
- 0:35 – 2:02
How LLMs shift search from keywords to understanding and answers
Arvind explains that search has long been dominated by brittle keyword matching, but LLMs enable much deeper semantic understanding. This changes the experience from returning links to directly answering questions using the underlying knowledge base.
- •LLMs improve understanding of both user intent and document meaning
- •Semantic matching reduces brittleness vs. keyword search
- •Search experience evolves from “links” to “answers” and conversation
- •Foundational shift enabling new search product capabilities
- 2:02 – 3:55
Glean’s early architecture: embeddings before “vector search” was mainstream
Elad asks about Glean’s technology transition; Arvind explains the company started in 2018–2019 already leveraging transformers. Early versions used BERT-derived embeddings customized per customer to power semantic search well before the current generative AI wave.
- •Started Glean planning in 2018; founded in early 2019
- •Transformers/embeddings were already influential inside Google search teams
- •Early product used “embedding search” (pre-vector-search terminology)
- •Customer-specific embeddings built on each company’s content
- 3:55 – 5:09
Why enterprise search requires more than embeddings: freshness, authority, and correctness
Arvind argues vector/embedding retrieval is only one component of good enterprise search. Enterprises have huge, messy, outdated corpora, so ranking must incorporate recency, correctness, and authoritative sources to return truly useful results.
- •Enterprise corpora contain obsolete and conflicting information
- •Need ranking signals beyond semantic similarity
- •Freshness and “authority” (expert-produced content) matter
- •Goal is to surface correct, current knowledge—not just relevant text
- 5:09 – 6:52
Why earlier enterprise search companies failed—and what SaaS changed
They discuss enterprise search as a historical graveyard. Arvind attributes many failures to pre-SaaS constraints: data was hard to access across on-prem systems, preventing turnkey products. SaaS APIs/connectors changed the feasibility of unified indexing.
- •Enterprise search historically had many attempts with few successes
- •Pre-SaaS/on-prem made data access and integration extremely difficult
- •SaaS systems provide standard APIs and interoperable access
- •Turnkey unified search became practical once connectors were viable
- 6:52 – 7:14
Glean’s origin story at Rubrik: ‘nothing to buy’ for internal findability
Arvind describes the practical trigger: Rubrik’s rapid growth created knowledge sprawl across hundreds of SaaS tools, and employees couldn’t find anything. When he tried to purchase a solution and found none, it motivated Glean’s creation.
- •Rubrik had information spread across ~300 SaaS systems
- •Employee surveys flagged findability as a persistent pain
- •Attempted to buy a solution; market gap was obvious
- •Glean founded to solve internal knowledge discovery at scale
- 7:14 – 8:43
Scaling challenges and the ‘low-signal’ enterprise environment
Arvind highlights the scale explosion inside companies—sometimes rivaling the early web—and notes cloud infrastructure made building scalable systems more achievable. He also contrasts web search’s abundant behavioral signals with enterprise search’s sparse usage data, increasing dependence on modeling techniques.
- •Some enterprises exceed 1B internal documents (web-scale)
- •Cloud tech reduces burden of building distributed systems from scratch
- •Enterprise lacks massive behavioral feedback loops seen on the public web
- •Transformers help compensate for limited user-signal data
- 8:43 – 10:43
Out-of-the-box models vs. curated context: why ‘just feed everything’ won’t work
Sarah asks how much bespoke IR infrastructure persists as models improve. Arvind argues that even with larger context windows, systems must still organize and curate information; dumping unstructured, out-of-order content makes reasoning difficult for humans and models alike.
- •Skepticism that infinite context windows eliminate retrieval/curation needs
- •Reasoning quality depends on how information is selected and organized
- •Chronology, structure, and relevance filtering remain essential
- •System work: prepare the right context before handing off to the model
- 10:43 – 12:14
From ‘Google at work’ to ‘ChatGPT at work’: Glean Assistant and secure RAG-style experiences
Arvind explains Glean’s product evolution toward a conversational assistant. Instead of returning links, Glean Assistant answers in a ChatGPT-like interface using both world knowledge and company knowledge, while enforcing user-specific permissions and safety constraints.
- •Product progression: search results → conversational assistant experience
- •Assistant uses internal company context plus general knowledge
- •Personalized to user identity and allowed access
- •Positioned as an always-available “sidekick” for employees
- 12:14 – 13:44
Apps/agents on top of internal knowledge: curated workflows for functions like HR
Elad asks about building applications atop the unified corpus. Arvind describes customer demand for curated, function-specific experiences (initially called “apps,” now often framed as “agents”) that don’t just answer questions but can also execute steps in enterprise systems.
- •Demand for curated, department-specific AI experiences (e.g., HR)
- •Need for “blessed” sources, controlled tone, and constrained retrieval
- •Shift from Q&A to workflow automation across business processes
- •Emergence of agents: acting in systems, not just responding
- 13:44 – 15:37
Access control and governance as the foundation for enterprise AI
Sarah probes access control in a platform aggregating unstructured data. Arvind emphasizes that most enterprise knowledge is private and permissioned; safe AI requires enforcing document-level access rules and governance end-to-end so responses never leak restricted content.
- •~90% of enterprise knowledge is private in some form
- •Cannot train/deploy a single model over all company data for all users
- •Indexing must retain permissions metadata per source system
- •AI UX must enforce identity, auth, and policy before retrieval/generation
- 15:37 – 19:13
User behavior surprises and the importance of AI education for ROI
Sarah asks about real user behaviors and superusers. Arvind notes users are trained by decades of keyword search and often don’t know how to use assistants effectively; adoption improves with guided prompts and role-relevant suggestions. He adds that beyond ROI, enterprises must invest in workforce education to become ‘AI-first.’
- •Users default to short keyword queries even when long prompts work better
- •Assistants feel unintuitive; need progressive onboarding and examples
- •Role-based prompting helps users discover valuable workflows
- •AI investment should include education/training, not just tool deployment
- 19:13 – 22:52
Hardest parts of building Glean: creating a market and fixing governance gaps
Elad asks what was unexpected versus Rubrik. Arvind explains Glean faced a ‘no budget’ category and required evangelism to create the market. Another surprise: customers feared “too-good search” because it revealed governance problems, pushing Glean to build classification and security-adjacent capabilities.
- •Unlike Rubrik, enterprise search wasn’t an established budget line item
- •Required market creation and winning innovators early
- •Good search surfaced sensitive/poorly governed content, causing fear
- •Glean expanded into governance/security features (classification, safeguards)
- 22:52 – 24:54
PLG vs enterprise sales: why Glean skewed top-down and how to run both motions
Elad asks when to do enterprise deals vs PLG. Arvind says Glean’s economics and architecture require company-wide indexing and rollout, making individual adoption hard; thus enterprise sales is structural. Ideally, he’d run PLG and enterprise motions in parallel to avoid leaving growth on the table.
- •Glean is inherently company-wide; single-user deployment doesn’t work well
- •Indexing the whole enterprise corpus makes cost structure different
- •Enterprise rollout drives cost-effectiveness and adoption
- •Best practice: start PLG + enterprise sales simultaneously when possible
- 24:54 – 27:51
Succeeding in ‘bad markets’: conviction, simplicity, and ignoring discouraging priors
Sarah asks how to think about negative market priors. Arvind argues over-indexing on past failures can kill good ideas; if a real pain persists, it signals opportunity. He describes choosing conviction over exhaustive doubt and letting details be solved through execution.
- •Priors can create analysis paralysis; everything has been tried before
- •Persistent pain implies unsolved opportunity despite past failures
- •Sometimes the right move is to commit and iterate, not over-theorize
- •Glean’s conviction came from firsthand experience across companies
- 27:51 – 31:34
What Glean wants to build next: assistants + agent platform toward ‘personal AI teams’
Elad asks about adjacent opportunities; Arvind says Glean will stay focused on its assistant and agent platform because the problem remains far from solved (retrieval correctness, missing knowledge, needle-in-haystack). He outlines a future where every worker has a personalized team of AI assistants/coaches that proactively does most work and helps upskill.
- •Strategic focus: (1) end-user assistant, (2) agent platform for processes
- •Core challenges remain: finding correct info, freshness, gaps where nothing is written
- •Hallucinations aren’t the only issue—retrieval and knowledge quality are major
- •Vision: everyone gets a proactive ‘AI team’ (assistant/coworker/coach)
