Arvind Jain spent a decade building search at Google and co-founded Rubrik before starting Glean in early 2019. Over 7.5 years, he has scaled the enterprise AI company to 1,500 employees, thousands of enterprise customers, more than $300 million in ARR, and a $7.2 billion valuation. As CEO, Jain holds the pole position in enterprise AI while Microsoft, Google, Notion, Anthropic, and OpenAI chase the same seat—with OpenAI reportedly telling its investors in 2024 not to fund Glean. The conversation covers Glean’s multi-model architecture, why AI will not shrink headcount, and the internal agents Jain uses to run the company. These signals are from an interview that Nakul Mandan did with Arvind Jain on the Knuckle Up with Nakul podcast.
The Moat
1. The Non-AI Moat. Glean started in early 2019 as an enterprise answering tool and now acts as an AI coworker that completes tasks across company systems. After 7.5 years deploying across thousands of enterprises, Jain says the hardest part of getting an AI coworker running has nothing to do with AI. It comes down to data security, governance architecture, user permissions, and regulatory compliance. His view: educating customers on that complexity—and solving it safely—is Glean’s main edge. → 2:21
2. The 10-Step Context Graph. Frontier labs rely on Model Context Protocol servers, expecting every enterprise app to build its own integration. Jain calls that approach shallow because basic system connectors never teach a model how a company actually works. Glean builds the connectors itself, then maps a context graph by observing internal experts as they execute 10-step tasks across systems. His reasoning: half of enterprise work lives in written SOPs, and the rest is tacit judgment inside employees’ heads. → 5:12
3. Token Maxing Creates Slop. With 1,500 employees, Glean does not train frontier models, partnering instead with OpenAI, Anthropic, Google, and open-source models on Nvidia hardware. Across 1,000 enterprise talks last year, Jain heard 3 urgent problems: constant model-switching between GPT, Gemini, and Claude; annual token budgets blown in 1 to 2 months; and token maxing that leaves unread AI docs and duplicate agents. Glean routes routine tasks to open-source models costing 10x less. Customers get every lab’s innovation without lock-in. → 8:13
The Architecture
4. Keep Agents Above the CRM. Enterprises are building go-to-market agents that prospect accounts, prep for meetings, and nurture deals. Jain says those custom agents hold a company’s core intellectual property, and their intelligence compounds the longer they run. Locking that compounding intelligence inside a single CRM or document tool traps the business when better software arrives. Because Glean is not a system of record, agents built on its layer keep working if a company switches from CRM 1 to CRM 2. → 11:31
5. The 5-Minute Internal App. Core systems of record like Google Workspace, Microsoft, Salesforce, Workday, and ServiceNow will survive the next few years by becoming agentic, Jain says. Narrow SaaS apps that just put a UI wrapper over database tables will disappear as employees build their own tools. Jain hated an internal asset-management app that could not sort, filter, or answer questions like Excel. He asked Glean to code a replacement over the database, and it built the app with Claude in 5 minutes. → 13:00
6. Cutting Token Spend in Half. Glean acts as an enterprise AI gateway connecting structured data in Databricks and Snowflake with unstructured docs and personal work memory. When Claude Cowork or ChatGPT tackles a job alone, Jain says, it uses brute force and burns cycles just gathering raw context. Plugging Google Vertex, AWS Bedrock, Claude, or ChatGPT into Glean’s MCP gateway feeds them pre-assembled context from the start. Customer studies show that short-circuiting context gathering through Glean cuts token consumption by 30% to 50%. → 15:09
7. Testing on Rivals First. Glean splits engineering into 2 groups: a core context-and-intelligence platform team and a first-party Glean Coworker application team. The internal Coworker team gets no special access or extra rights over external tools that plug into Glean. In fact, Glean’s platform team does not test its work on Glean Coworker at all, testing only on Claude and ChatGPT instead. That rule forces Glean’s infrastructure to make rival interfaces faster and smarter before serving its own app. → 18:53
The Workforce
8. The 5% Power Law. Even inside Glean’s most tech-forward customer accounts, AI usage follows a strict power law. While 1 in 3 Glean users run deep tasks—like support teams where Glean resolves most tickets untouched—only 5% of enterprise workers have turned their workday into reviewing AI output. The other 95% stick to basic question answering because they lack the time to prompt. Jain’s fix: a proactive companion that spots 10 tasks on a worker’s plate and finishes 5 before they ask. → 22:03
9. Why Headcount Won’t Drop. Jain does not believe companies will cut their workforce with AI and keep their market share. If 5 competitors in an industry all get the same AI tools that let half the staff do today’s work, customer expectations will simply double. Every rival will use its full team to build stronger products rather than shrinking headcount. His rule: business success scales with human capital, and no company working 2 days a week will beat a competitor working 5. → 24:19
The CEO Playbook
10. Auditing Strategy in Minutes. Earlier this year, Glean shifted strategy so 50% of its users would come through third-party tools like Claude and ChatGPT. Checking whether engineering projects matched a new strategy used to take a once-a-year trickle-down review through department heads and managers. Instead, Jain ran deep Glean analyses over the first few months to inspect every engineer and project, spotting misaligned work to drop in minutes. A separate agent scans company work to find the top 10 quiet performers. → 29:11
11. Prism Replaces Churn Dashboards. Customer success dashboards almost never hold the truth, Jain says, because reps skip manual updates and key nuance gets lost in customer calls. Glean built an internal system called Prism that pulls live customer conversations and activity into an always-current health dashboard. Prism scores every account automatically to rank Glean’s top churn risks across the company. It then gives the team 3 specific actions to fix each at-risk customer before they leave. → 31:00
12. Managers Must Build Agents. Glean’s finance team runs dozens of agents, but Jain was still unhappy with internal AI adoption last quarter. He used an internal product called Agent Finder, which reads Glean’s work graph to rank the top 20 to 30 unautomated tasks eating up human hours. Instead of asking engineers to build those agents, Jain assigned them to managers and department leaders. His rule: leaders cannot delegate the build, and they can use Glean or any other platform. → 33:48
The Culture
13. The 90% SDR Rule. Glean refuses to use fully autonomous AI sales development reps because AI still lacks human judgment. Jain says his high school kids can spot AI-written text right away, and enterprise buyers can tell just as fast. Glean lets AI handle 90% of prospecting—picking accounts, finding contacts, and drafting outreach—while a human SDR owns the last 10%. The rep reads every draft and gives a thumbs-up or talks to the agent to tweak the angle before sending. → 36:43
14. The 2-Hour Coding Test. Glean checks every job candidate’s past work in Claude, ChatGPT, and Gemini, and it rebuilt its 2-hour engineering interview. Instead of telling candidates to use AI, Glean assigns a coding problem so complex that nobody can finish it in 2 hours without AI. Jain also got tired of Greenhouse offer packets with thin praise like “oh, I love this person, he’s great, let’s hire him.” An agent now blocks any offer packet from reaching him until interview questions are verified. → 38:33
15. Deleting Code Beats Writing It. In past software eras, winning a customer brought stability, Jain says, whereas falling 20% behind in AI pace makes a product indefensible. Rapid innovation turns old code into a liability, so he tells Glean’s 1,500 employees that “if somebody deletes lines of code, that’s actually even more valuable than writing lines of code.” Even so, he says AI “is not ready to replace even a single human today,” including an EA. Executing 10 daily tasks keeps the anxiety away. → 43:02
Watch the full Video Here.