Arvind Jain on building Glean and the future of enterprise AI
Summary
Glean’s founding thesis was enterprise search, but its early transformer bet made the 2019 product unusually well positioned for generative AI. Arvind Jain began with a universal pain point—research suggested employees spend one-third of their working time finding information—and used BERT-based models to match concepts rather than keywords. As generation and reasoning improved, Glean evolved from “a Google for you in your work life” into “ChatGPT for work life.”
Glean uses frontier models where broad capabilities already exist while building enterprise-specific retrieval. It trains small models on a customer’s corpus for custom embeddings, and separately fine-tunes small open-domain models for narrow search tasks such as spell-checking, synonym handling, and acronym expansion. It uses GPT, Gemini, or Claude for synthesis and multi-step reasoning. Jain’s rule is blunt: “Do not reinvent things that have been invented already.”
Permissions and data freshness—not merely model quality—are the central enterprise-AI constraints. Glean imports governance from systems including Google Drive, Slack, and Salesforce, bakes permissions into its index, and retrieves only documents the signed-in user may access. Because embeddings themselves might leak restricted information, customer-specific models are trained only on subsets Glean judges safe.
Jain’s repeat-founder pattern is a contrarian bet on universal problems in markets others have abandoned. Lukas Biewald pressed him on whether Rubrik and Glean reflected exceptional execution rather than novel ideas; Jain agreed enterprise search had produced “only failures” and become a “dead area” where investors did not want to invest. His conviction came from two changes: SaaS made fragmented data much worse but also more accessible, while transformers made semantic understanding technically viable. His Google-derived operating model is to put innovation first, hire smart engineers, and largely let them build.
A concrete ROI example comes from reasoning across the 95% of enterprise data that is unstructured. A non-engineer in Glean’s finance team asked an agent to combine Salesforce customer lists, shared Slack sentiment, and product usage into green/yellow/red account-risk profiles after churn appeared. Jain says the result was better than a conventional dashboard because it could incorporate subjective textual evidence.
Glean evaluates the complete answer pipeline, while conceding that enterprise AI cannot eliminate errors. It derives “golden” question-answer sets from real interactions such as well-received Slack replies, tests retrieval and model changes against them, and uses LLMs as judges. For hallucinations, it checks answers “line by line” against supplied source material and may suppress unsupported claims or abstain—but Jain says stale, missing, or poorly retrieved knowledge causes more failures than fabrication alone.
Jain rejects labor reduction as the most valuable AI strategy and instead wants every employee surrounded by a scalable “dream team.” He imagines assistants, coworkers, and coaches helping each person do up to 90% of the work they need to do, while companies retain and even expand teams whose members can do “ten times more work.” The transformation may feel incremental—one task at a time—until workers discover they are fundamentally different from two years earlier.
Deep dive
1. Enterprise search became Glean’s generative-AI wedge
Jain started Glean in early 2019, before the term “generative AI” existed, to his knowledge, to solve an older problem: information scattered across enterprise systems. Research suggested people spent one-third of their working time finding things, so the initial ambition was “a Google for people in their work lives”—one destination for documents, data, answers, and relevant colleagues.
Transformers were already changing search inside Google. After BERT-based models entered the open domain in late 2018, Glean could train small models on enterprise corpora and recognize semantic proximity between documents with almost no word overlap. Jain cautiously claims Glean was “probably the first company to use transformers in the enterprise,” because the improvement was observed rather than merely predicted.
Those 2019 models could neither generate nor reason, but their embedding-based matching created the foundation. As model capabilities arrived, Glean moved from surfacing documents to reading them and producing precise answers: “What used to be a Google for you in your work life became more like ChatGPT for work life.”
2. Glean builds enterprise retrieval while using frontier intelligence
Jain’s instruction to the technology team is to “maximally use innovation that’s happening outside.” Google, OpenAI, and Anthropic already make capable models available, so recreating them would serve neither Glean nor its customers: “Do not reinvent things that have been invented already.”
Glean does train small, customer-specific models where enterprise context matters. These learn a company’s language, code words, and internal concepts, producing custom embeddings that match questions to relevant documents more accurately than generic semantic search.
Once retrieval assembles the raw material, GPT, Gemini, or Claude synthesizes the answer; the largest foundation models also handle reasoning and decomposing complex questions into multi-step agent workflows. For narrow search operations such as spell-checking and query expansion involving synonyms and acronyms, Glean fine-tunes small open-domain models because using GPT would be effective but “overkill.”
Security lives inside retrieval. Glean reads both content and governance from Google Drive, Slack, Salesforce, and other systems, then bakes permissions into its index so each user retrieves only authorized material. For custom embeddings, it trains on a heuristically selected safe subset rather than the complete corpus, acknowledging that model internals could otherwise leak restricted information in unknowable forms.
3. Failed categories can become attractive when enabling conditions change
Biewald framed Jain’s track record as unlikely to be mere luck: Rubrik’s IPO valuation was, in his explicitly hedged recollection, over $20 billion, while a $10-billion-plus outcome might be “one in a thousand” or “one in ten thousand.” His pushback—worth keeping—was that both Rubrik and enterprise search were familiar pitches where others had repeatedly failed, suggesting execution was the differentiator.
Jain’s answer remained deliberately modest, but the selection logic was clear. Rubrik, contemplated in 2013 and started in early 2014, addressed the universal need to protect data as ransomware and other attacks rose. Jain said the best technology talent and attention were then focused on consumer technology while enterprise lagged, and that existing enterprise products were often more than two decades old. Glean likewise attacked a problem almost every employee recognized, even though decades of failures—including Google’s—had convinced investors that enterprise search was either insoluble or unnecessary. He also noted that no recognizable search company had started in the preceding 10-year period.
What turned a bad category into a contrarian opportunity was a paired technology shift. SaaS made the information problem much worse, yet also made it tractable because SaaS systems allowed Glean to read enterprise data; transformers supplied conceptual understanding. Jain still found his conviction surprising because “doubts are part of my existence,” but concluded that very large companies must tackle problems “universal in nature.”
4. Enterprise AI is moving from search box to operating layer
Jain expects every worker, regardless of seniority, to gain assistants, coworkers, and coaches capable of helping them do 90% of the work they need to do while also helping them improve. Today that surrounding team is largely a CEO’s privilege—assistant, coach, chief of staff, executives—but AI can democratize it: “You’re no longer alone.”
One current Glean agent produces up-to-date technical tutorials without requiring status meetings. It reads design documents, checks resolved Jira items and committed code for later changes, and reconciles stale descriptions with current implementation. Jain calls the result “evergreen documentation”: the system generates the latest explanation rather than returning the most prominent obsolete document.
The churn example shows the operating leverage. After some churn appeared, a finance employee with no software-engineering background asked Glean to enumerate customers from Salesforce, inspect shared Slack conversations for sentiment, combine that with product usage, and classify each account green, yellow, or red.
Jain judged that report better than a dashboard because dashboards mainly consume structured measures, while the agent could interpret subjective text alongside usage. His broader claim is that “95% of your enterprise data is unstructured”—rich in insights that previously remained inaccessible because nobody had time to extract them.
The organizational model behind this vision comes partly from Jain’s Google experience: put technology and innovation first, hire smart, motivated engineers, give them resources and autonomy, and largely let them organize and build. At Glean, he says his role is often to stay out of the way while fostering an innovation-first culture.
5. Reliability depends more on evidence quality than confident generation
Glean is its own “customer zero,” but evaluation extends to customer deployments. Real conversations already contain potential golden sets: a Slack question, a human answer, and strong reactions or thumbs-ups can establish what a good response should resemble without constructing every benchmark manually.
Each search, retrieval, model, or other system change is run against these question-answer sets. Because quality is partly subjective, Glean also uses LLMs as judges to compare generated responses with golden answers and measure the end-to-end system rather than model performance in isolation.
Jain’s honest qualification is that Glean will not “always answer questions exactly the right way.” Information may be missing, stale, or written by the wrong source, while the model itself may hallucinate. Glean checks generated claims “line by line” against the input evidence, attaches citations to supporting human-written material, and suppresses or declines answers when support cannot be found.
Model hallucination is therefore only one failure mode—and, in Jain’s view, not the largest. “Most of the losses” occur earlier: finding the correct information across many systems, determining whether it is current, and identifying whether it came from a subject-matter expert. When no answer exists, Glean can instead identify the people working on the topic; Jain also sees a legitimate future in which agents proactively ask humans for missing facts.
6. Augmentation offers a larger prize than headcount reduction
Jain thinks forecasts of tiny teams or a one-person billion-dollar company are “missing the point in a big way.” A business derives its value, intellect, and capacity from people, so AI should enable that base rather than primarily replace it.
His preferred strategy is explicit: “I don’t want to actually shrink my team size. I actually want to grow it,” while enabling everyone to do “ten times more work.” Companies will choose between using AI to reduce team size or continuing to invest in people and improve the top line significantly.
The durable success factors remain “hard work, dedication, and desire to succeed,” though task mix will shift. Engineers may need to remember fewer APIs and frameworks, spend more time thinking and designing, and less time writing code. Jain expects adaptation to arrive incrementally—“every day there’s one new task”—until the accumulated change becomes transformational.