One Company Now Holds as Many Documents as the Entire Early Web

Sarah Guo introduces Arvind Jain as a repeat founder who co-founded Rubrik, spent years on search at Google, and now runs Glean, an AI-powered enterprise search and knowledge platform over internal documents, Slack, and the apps built on top. Elad Gil has known him even longer. Guo’s first question is the obvious one for someone who has been “a search guy” for almost 30 years: how much did LLMs actually change the field?

Jain’s answer is that the paradigm finally moved. For a long time search was static: people asked questions, the system found words in documents, and it brought those documents back. LLMs changed the core job. They let the system deeply understand what a user is asking and, separately, what a document is about, then match the two conceptually. Search, he says, is not brittle anymore. It can go far beyond surfacing a few links and actually answer the question from the knowledge it has.

Gil remembers Glean starting in a more traditional search world and then shifting as foundation models arrived. Jain says the timing was lucky rather than prophetic. They started thinking about the company in late 2018 and founded it in early 2019. Transformers already existed, even if the whole world was not talking about them. Inside Google’s search teams, people had already seen what embeddings could do. Version one of Glean used transformers for semantic matching. Nobody called it vector search. RAG and generative AI had not been coined. Internally they called it embedding search.

The models then were not today’s models. They started from BERT, which Google had released after training on internet-scale data, then built custom embeddings on each customer’s business content to power the semantic layer. Jain immediately pours cold water on the current obsession. Embeddings and vector search are only one part of a good search system. Picture a company that has been around for decades, with information spread across many systems, much of it obsolete. Semantic match is not enough if you surface a five-year-old plan.

You also have to pick information that is correct today, up to date, and written by someone with authority on the topic. Those older information-retrieval problems—freshness, correctness, expertise—still have to be solved. That is the shape they started building, in a category almost nobody wanted to fund.

Coming off Rubrik, Jain still found investors uninterested in enterprise search, or in him for that matter. The problem was not exciting. Guo recalls vendors such as FAST, the Norwegian engine from the early Google era, and calls the category a graveyard: lots of attempts, no successes. Jain splits the failure into access, scale, and semantics. Before SaaS, you could not build a turnkey product because the data lived in data centers, on unnamed servers and storage systems you had to visit. SaaS changed that: one version for every customer, open and interoperable, hittable with APIs, so you could finally pull enterprise knowledge into one place.

Gil jokes about the Google Search Appliance—the idea that you had to slurp customer data continuously into a hardware box in order to search it, which now sounds ludicrous. Glean’s origin story is more mundane. At Rubrik, Jain ran IT. The company had information across 300 SaaS systems, nobody could find anything, and the complaints showed up in pulse surveys on his desk. He tried to buy a search product and realized there was nothing to buy. Glean began as a procurement need that did not exist in the market.

Scale was the second unlock. One of Glean’s largest customers has more than a billion documents inside the company. When Jain and Gil were working on Google Search in 2004, the entire internet was about a billion documents. You could not have built that kind of system before the cloud; you would have spent all your time on a homemade distributed stack. Transformers were the third unlock, and more necessary inside a company than on the web. On the public web you can paper over weak semantics with a billion users’ behavior. Enterprises do not have that luxury, so models have to make up for thin click logs.

Guo asks whether traditional IR signals—freshness, authority, bespoke infrastructure—persist, or whether the models eventually just do it. Jain does not think we are close to infinite context windows that let you dump everything in and walk away. He asks you to imagine a highly capable person given a million documents in no order: one from today, one from four months ago, one from three years ago, another from two days ago. Even with perfect memory, unordered information is hard to reason over.

He thinks about models the same way. You still have to present information in some organized fashion if you want them to read, reason, and answer well. You can feed them more over time, but *how* you provide the right information still determines the quality of the output. That retrieval-and-organization layer does not disappear just because context windows get longer.

Glean’s original vision was “Google in your work life.” As models gained reasoning and generation, the product moved. Glean Assistant looks and feels more like ChatGPT: you converse instead of scanning links. The difference is that it uses the world’s knowledge *and* the company’s internal data, in a way that knows who you are and what you are allowed to see. Jain describes it as a personal assistant or sidekick for every employee, always available for questions and tasks, using company context to help with the work.

Businesses were even more interested in transforming specific expensive processes. Early last year, before agents were the talk of the industry, customers asked for curated applications on the data platform. HR teams loved the assistant for benefits, PTO, and vacation policy, but it sometimes cited content their people team had not blessed. They wanted people-related questions to use only curated HR content, in a particular tone. Glean built those function-by-function experiences and called them apps, not agents. Now the same surfaces are expected to do work in systems, not just answer, so people call them agents.

Permissions are not an add-on. Jain’s number is that about 90% of knowledge inside a company is private in some shape: a document only you can see, or one shared with a few people. You cannot train one internal model on the whole corpus and open it to everyone, or engineering will see compensation data that only HR should see. When Glean indexes a Google Drive file or a Slack thread, it also stores who can access it. Users must be signed in, and the system only uses information they have permission for. Without that infrastructure, he says, you cannot roll out AI safely.

Guo, who has learned odd user behavior from search people, asks what Glean’s super-users actually do. Jain assumed the product needed no teaching: one box, type a question. Then they added natural-language capabilities and the ability to paste a paragraph of instructions, and people would not do it. Twenty years of Google trained everyone to type one or two keywords. Classic search launched to immediate high usage with no confusion. The assistant left people unsure what to ask. The curious ones asked things the product could not answer, including “what should I do in my life?”

His conclusion is that AI is unintuitive for most people. You have to expose capabilities incrementally, tied to day-to-day work. For an engineer, prompt them that they can learn a new technology right now and get a two-page tutorial. Understand the core job, then give people prompts that make them try something and get excited. Enterprises are spending on AI and demanding ROI—efficiency, top-line—while overlooking education. He asks CEOs to picture the workforce they want in three years: AI-first people who know how to use a technology that is powerful, imperfect, and hallucinatory. Experts will get a lot done with it. That has to be the leadership objective, not just buying seats.

Compared with Rubrik, rolling Glean out was unexpectedly hard. Rubrik sat in an established market with budgets, replacing old technology. Search had no line item. Everyone agreed it mattered and still ranked it as a vitamin, not a painkiller, because people were already living without it. Jain’s team had to evangelize and find innovators willing to buy a product they were not used to buying. Then came a stranger problem: the product worked, and businesses got scared of good search. Governance was a mess, sensitive information was lying around, and people started discovering it—salaries at one customer, a sensitive M&A document that was not public yet at another. They could not sell, he says, because the product was so good.

That forced them past respecting each system’s permissions into asking who you are, whether you should even ask the question, and whether the answer is safe to show. Glean ended up as a security product. A lot of companies now buy it to fix governance and become AI-ready, not only for Glean Search and Assistant but for every other AI product in the enterprise. Personally, Jain was not CEO at Rubrik; he ran R&D. At Glean he is still learning the job and still thinks like an engineer. Growing out of that into running a large business is the transformation in progress.

Guo notes that Glean went top-down into big enterprises twice, against the industry’s PLG reflex. Jain is candid: his dream was PLG, a product that sells itself on the web. Structurally it was impossible. Even one person’s search needs the whole company’s index, which is expensive, so they could never sell seats to ten employees. The recipe he would choose if he had both options is to start PLG and enterprise sales at the same time. Treat PLG as lead gen, nurture and expand with enterprise sales, and do not spend three years on self-serve before adding a sales team. Timing matters; one motion leaves a lot on the table.

On whether to trust bad market priors, his engineer’s advice is almost anti-engineering. The more you study history, the more you kill your own idea; every idea has ten reasons it will fail. A simpler test: if people feel the pain, nobody is solving it, stop arguing and do it. Google’s internal joke was that they spent their lives making it easy for the world to find things, and impossibly hard to find anything inside Google. That conviction was enough. The company will stay on two products: an assistant for every employee, and an agent platform that injects AI into business processes. The pitch is almost too large—ask any question, give any task, Glean uses world knowledge plus company data safely—and he immediately undercuts it: that promise says you do not have to work anymore, and they are a long way from it. Hallucination is not even the biggest problem. Often the information was never written down, or they pick the wrong needle in the haystack. The longer vision is a personal team of assistants, coworkers, and coaches around every person, doing 90% of the work and making new grads 10x. CEOs already live in that world. Glean wants everyone else to as well.