AI summary by TwinMind
How Glean CEO Arvind Jain Solved the Enterprise Search Problem – and What It Means for AI at Work
Sequoia Capital
44m 48s
Company Overview and Vision
- Glean = "Google or ChatGPT inside your Enterprise"
- Employees ask questions, Glean answers using company knowledge regardless of location
- Also AI platform for building Enterprise AI applications via RAG
- Core belief: majority of knowledge work won't be done by humans in 5 years
- AI assistants will be more powerful - access to all company data, context from past conversations/meetings, don't forget anything, have reasoning capabilities
- Want Glean to be the workplace AI assistant
Background and Market Context
- Arvin Jain: Co-founder/CEO, former Google search engineer, co-founder/CTO of Rubric
- Enterprise search = "Holy Grail" for CIOs with 2 decades of failed attempts
- Even Google never cracked this category before
- Problem was startup-unfriendly: required massive R&D investment, pre-SaaS world made connections difficult (1 year to integrate Enterprise data)
Technical Architecture - Pre-AI Foundation
- Deep integrations with Enterprise systems: Salesforce, Confluence, Jira, Google Drive, SharePoint, ServiceNow
- Governance layer: understands permissions across hundreds of systems, only retrieves information user has access to
- Knowledge graph: understands people, roles, documents, relationships between individuals and documents
- Ranking system using multiple signals:
- Document popularity/engagement
- Recency (documents from last 1-2 weeks with engagement)
- Team-specific usage patterns
- Slack interactions (thumbs up responses to document links)
- Language models integrated from day one (2019) - originally BERT-based for semantic understanding
RAG Architecture and Challenges
- RAG = connecting private Enterprise data to public models (GPT, Gemini, Claude)
- Process: Question → Retrieval engine finds relevant documents → Model answers using company knowledge
- Challenges: Chaining two imperfect systems (search + LLM), both can fail
- Most Enterprise RAG uses simpler vector search without full Enterprise context
- Glean's approach: build standalone search system good enough for end users, then use for RAG
Customer Impact and Usage
- Northstar metric: 80% of user sessions successful (measured implicitly via click-through and dwell time)
- Engineering teams: troubleshooting tool, debugging errors using past solutions
- Support teams: fundamentally changed workflow from searching to validating AI answers
- Sales teams: meeting prep with 360-degree customer view, real-time question answering during calls
- Universal use case: finding subject matter experts
- Company quadrupling year-over-year
Agentic Reasoning and Workflows
- Building workflow engine for complex questions requiring multi-step processes
- Example: "Help me write weekly status report" requires understanding team structure, work locations, pulling data from multiple systems
- Current approach: human-assisted workflows, users can build manually
- Learning from user-created workflows to train models for automatic workflow generation
- Error rates compound in multi-step processes (10 steps at 90% accuracy each)
- Human assistance critical for complex workflows
Platform Strategy (Act 1 → Act 2)
- Act 1: Enterprise search product (immediate value)
- Act 2: AI application platform
- Search product provides "moral authority" - customers already trust with all data
- APIs launched for developers to build on Glean platform
- 90% of AI app development = boring infrastructure work (ETL, data pipelines, search, governance)
- Glean handles infrastructure, developers focus on business problems
Future Vision (5 years)
- Majority of knowledge work done by AI assistants
- Proactive assistance (like executive assistants) vs reactive Q&A
- AI assistants will manage daily work, guide effectiveness
- Everyone gets luxury of AI-powered personal assistant
- Glean aims to be "world's most successful AI assistant"
Competitive Strategy and Advice
- Market expanding from $600B software to $12-15T with AI
- Advice to founders: Don't worry about competition, focus on problem-solving
- Glean had no competition for 4 years (dead market), now everyone wants to build similar products
- Competition validates market, creates demand
- AI company = regular company using AI as tool, not fundamentally different
Model Training Strategy
- No plans for large model training (GPT-4 scale)
- Train smaller, customer-specific models on individual Enterprise corpus
- Models learn company lingo, acronyms, code names
- Partner with large model providers
- Model training core to Glean technology but focused on customization not foundation models
