How Glean Is Turning Enterprise Context Into an AI Cost Advantage With Emrecan Dogan

• Why grounding AI in enterprise context cuts token use by about 30% almost by accident, and why 'token yield' is quietly replacing 'token maxxing' as the metric that matters.

• How indexing beats federation: serving answers from one constantly updated index instead of querying hundreds of systems live, the same reason Google search feels instant.

• The case for 'compiled institutional judgment,' and a proactive future where the best interface might be almost no interface at all.

  Let's dive in. No floaties needed…

Revenge of the Nerds

Emrecan Dogan, chief product officer of Glean

Emrecan Dogan is the chief product officer at Glean, where he leads product, data, design, operations, and partnerships, essentially all of R&D outside engineering. Before Glean, he spent years at LinkedIn (which acquired his education startup, ScoreBeyond), with earlier stops at Stripe and Amazon and an MBA from Stanford. Glean, founded in 2019 by former Google search engineer Arvind Jain with T.R. Vishwanath, Tony Gentilcore, and Piyush Prahladka, began as an enterprise search platform and grew into a Work AI platform with an assistant, agents, a Glean MCP server, auto-model selection, and Skills.

In June 2025 it raised a $150M Series F led by Wellington Management at a $7.2B valuation (backers include Sequoia, Kleiner Perkins, Lightspeed, ICONIQ, and General Catalyst), and per its own May 28, 2026 announcement it "reached $300 million in annual recurring revenue (ARR), just 15 months after reaching $100 million ARR," while nearly doubling its Fortune 500 customer count year over year.

What makes Dogan worth reading is that he is selling the opposite of the prevailing AI pitch. While much of the market sells ambition and more tokens, he argues the smarter enterprise move is to spend intelligence only where it creates value, allocating models 'in proportion to surprise' and treating context, not raw model horsepower, as the moat.

Glean's own evaluations make the case: in an August 26, 2026 benchmark across 180+ enterprise tasks, graders with a preference chose Glean's answer over Claude Cowork's 78% of the time. Glean also burned 70% fewer tokens, 1.3M vs. 4.4M, reducing the cost per task from $2.98 to $0.58 (Glean ran auto-routing across models; Cowork was held to Claude Sonnet 5).

The tension is trust: to deliver any of this, Glean has to index a bank's or pharma company's most sensitive data, and Dogan's bet is that whoever safely contains AI is the one who gets to unleash it. For anyone deciding how their organization actually adopts AI, that is the argument to wrestle with.

What is Glean?

Glean helps people use AI to get real work done inside an organization. We are all familiar with AI in our consumer world, but when you put on the head of a knowledge worker, you need all the details of what is going on in the work around you, and in the larger ecosystem of departments, subsidiaries, business units, and acquired entities, and how the company operated historically.

So Glean connects to the systems where work lives and understands the relationships among people, documents, projects, and decisions- ones being made now and ones made years ago. And it preserves the permissions already in those systems. What it creates is a context layer for work, not just what the organization knows but how it operates.

On top of that context, Glean works like the AI assistants we're familiar with, ChatGPT, Claude, and Cursor. It answers questions, creates work, and executes multi-step workflows, using the right models and action tools through the MCP pathways.

It started about seven years ago with enterprise search, because finding and understanding organizational knowledge was the foundation. Even today, in the AI maximalist world, it still is, and we're now building the Work AI platform on top of it.

Why are enterprises already optimizing AI costs this early?

Let me address that head-on. Today, when we cut token spend by 30 or 40%, that isn't even the result of an intentional effort. It comes from Glean as a side benefit, not because some engineer or PM asked how to make Glean more token-efficient. If we really focused on it, and these days we are, you could drive that much further. It's a side benefit of handling enterprise context better, and I have no doubt we could get it 4x or 5x when we really focus.

Why is it becoming more important? We, as an industry, the decision makers, are beginning to industrialize cognition, and when you do that, marginal costs make a comeback. Traditional software has effectively zero marginal cost: the more you use Jira, Salesforce, or ServiceNow, the less it costs. But AI burns compute every time it reasons, searches, and acts. A single chat response may be cheap, but an agent working across six systems for thousands of employees creates a fundamentally different economic system.

They loved traditional software because it was a fixed-cost line. You knew exactly what percent of your revenue would go to a vendor, and that's how they operated for 20 to 25 years. Now they're running AI with marginal costs, so every time employees use AI, margins get downward pressure.

The first phase of AI was exploration: introduce it to the workforce, subsidize usage, burn tokens, discover what's possible. Token maxxing became a positive term. The next phase is about getting the greatest useful outcome from every unit of machine intelligence. We call it token yield. It's not about minimizing spend; it's about maximizing the value you get out of every token. And my bet is that context is the core infrastructure for that yield.

Without context, an agent repeatedly pays to rediscover what the organization already knows. It keeps searching broadly, reading irrelevant material, retrying tools, and asking a frontier model to reconstruct the relationships between people and projects. You don't need to pay the intelligence layer to do and redo those things. Better context lets the system spend intelligence on only the novel part of the problem.

What is Glean doing differently from ChatGPT & Claude?

When Glean beats these great products, it might be using the very same frontier model to answer the same query in our evaluations. It's not that Glean is building another foundation model, or post-training, or distilling. Even apples to apples, same model, you use Glean as your context provider instead of MCPs, instead of federation, and that's where the value is created.

Imagine a simple query: 'What is the latest status on the Acme renewal?' Now picture two universes. Without Glean, you connect a bunch of MCP servers to ChatGPT, Claude, and Cursor. The harness, the frontier model, asks what tools it can access, then reaches out to Drive, Jira, Salesforce, Teams, maybe hundreds of systems, issuing individual search queries into each.

Every system returns hundreds of artifacts, snippets, and documents, with a lot of noise. You pay the harness a lot of money just to sift through those results and find the right context. That hits quality, because you get noise; latency, because you wait for every system; and cost, because you hand a ton of data to the harness.

With Glean, you're not hitting any external system. Just like Google indexes the world's internet, Glean has an index constantly updated from all those systems. When you run one query, Glean serves just the right documents, snippets, people, and project data- the minimal high-signal artifacts to answer it. That's the differentiator.

Here's the non-obvious part. Many of our customers still use Claude, Cursor, and ChatGPT for some teams. But instead of connecting those products to individual MCP servers, they connect them to one, the Glean MCP server. That's where we create value even when the interface isn't Glean.

Imagine you go to Google.com, get near-instant results, and trust that Google understands the world. Now imagine that when you gave Google a query, it first hit the Wall Street Journal, then the New York Times, then Reddit, then Twitter, getting results from each site's own search technology, some great, some terrible, then tried to sift through all of them. How much worse would Google have been? That's the equivalent for AI agents, and it's what Glean differentiates on.

What is Glean's auto model selection optimizing for?

I recently wrote a bit on this. I worked at LinkedIn before, so I care deeply about talent, the role of humans, and human craft in the world economy. I think about how humans and organizations create value in terms of who brings the alpha and who brings the beta. People like you and me create value by bringing the alpha. Beta is all the information priced at consensus: everybody knows it, it's generally the correct way of doing things, but you don't create value out of it.

Models are amazing engines for the beta, which is generally known. But a knowledge worker is expected to bring what isn't properly priced, the non-consensus, sometimes contrarian decisions that turn out correct. That matters a lot for the future of human judgment and taste, and this symbiotic relationship between what AI does and what humans do.

Back to your question, the deep principle is that we should allocate intelligence in proportion to what I call surprise. Routine, well-understood work shouldn't require the largest, most expensive model or the biggest reasoning budget. The more predictable a task, the more of it can be handled by retrieval, smaller models, faster models, and deterministic machinery. Frontier intelligence should go to tasks filled with ambiguity, with an upside of novelty, and to the moments when the world violates our expectations and humans need to create alpha.

Think of Glean's auto mode as evaluating that. As Glean's CPO, I might ask Glean to clean my inbox of unnecessary emails. High-intelligence task? No. If it does it well, will I be a better CPO? Not really. But if I ask, 'Reflecting on my last two weeks of customer conversations, what did I learn that I didn't know before, so I can refine Glean's product strategy?' that's ambiguous, high-novelty, high-intelligence.

So auto mode knows the nature of the task and a lot about the user: who they are, where their alpha contribution comes from, and what they're supposed to do. Do I deploy a frontier model to increase its creativity, craft, and taste, or am I saving time for the user and budget for the organization?

The other aspect is admins, like heads of IT, who want to govern this. These models should be available to everybody, but different departments or seniority levels get different budgets, monthly or daily. So auto mode also factors in that mishmash of rules, governance, and fallback scenarios to allocate intelligence across teams and users. It brings enterprise budgeting into the same objective function. At that moment of truth, we allocate intelligence on task, user, and enterprise governance principles.

How do Skills work as compiled institutional judgment?

This ties spot on to what we just touched on: the role of humans and how they create the alpha. AI is amazing at bringing the beta, the generally accepted ways of doing things, and it can prevent many unforced errors.

I think of it like an MBA. You learn a ton of frameworks and case studies, but no employer hires you to apply the case studies the way they are. They expect that now you have your training set and can use it to make better decisions. For a product manager, a better decision is the judgment and taste to decide what will resonate with users, what will achieve product-market fit, and what will create value for the company, the users, and the buyer.

In an organization, the real flow of decision-making is very poorly captured. The moment we write a process document, it's stale the next day. Imagine we write a document on how we decide on a candidate, who makes the call, and what the warning signs are. It might be great today, but the moment you bring a new hiring manager or re-org and change the interviewers, it's out the door.

So a skill is a more continuous understanding of how Tino, Emrecan, or Employee X gets things done and makes decisions, without taking the time to capture that explicitly. With Glean as your context engine, it can extract how you tend to behave, your patterns of decision-making, even when you're not aware of them. As a product leader, I do a lot of product reviews.

Teams bring an idea and their rationale, and I engage, leave comments, sometimes challenge, sometimes support. But I never write down my exhaustive way of approving or rejecting a product idea. My approach is a series of engagement patterns over months or years, where each engagement shows only a subset of my decisions. Now with AI, you can extract all those patterns, create that institutional judgment, and compile it so software understands it.

The big thing is that for the first time, through skills, we're creating institutional judgment that's compiled and executed by software. And when skills can auto-update, self-heal, and self-improve over time, it becomes magical because you're no longer looking at one version. You're relying on a context engine like Glean to continuously refine and refactor that skill based on your patterns.

How do you get banks & pharma comfortable indexing sensitive data?

Our luck is that we weren't founded in the last couple of years. Glean is seven years old and some change. Even before the world met generative AI, that same sensitive data question existed, and we've spent seven years building our trust and security layer to convince these same organizations to let Glean build a search index and knowledge graph of their entire company. So I have a head start. I've been investing in trust and security for seven years rather than scrambling after the generative AI moment.

What I learned is, first, they don't want their data to leave their cloud. Each of these Fortune 500 companies has its own cloud, and Glean is a single-tenant virtual private cloud system. I don't operate the bank's Glean instance. It's all self-contained within the bank's own infrastructure. If the bank runs on AWS, Google Cloud, or Azure, we let them deploy Glean there. So when Glean indexes, data isn't transmitted elsewhere. It stays within the company's own security posture.

Second, when we need to send data to LLMs, a customer might say, 'I only want models served from Vertex AI or AWS Bedrock,' and we honor that. It all minimizes or eliminates data transmission outside the organization's security boundaries. These checks and balances let banks and pharma stop telling us their data is too sensitive for Glean to index.

Let me shed light on a less obvious part. We tend to discuss security only as a defensive requirement. But I think whichever technology figures out how to safely contain AI will be the one that unlocks its next-gen power.

Today, the zeitgeist is that AI escaped its sandbox and hacked this system and that one, so we talk about containment defensively, which isn't bad. But when banks and pharmaceuticals realize they can safely contain AI in their own cloud, imagine how much more they'll invest to let AI run proactively, autonomously, and make production-grade decisions.

In manufacturing, nobody lets AI fully run the making of cars today. It helps and supports, but doesn't run it. So yes, there's the defensive downside, but there's a huge upside. Today, Glean has no problem convincing large organizations to give the data. Running actions and processes autonomously is one more level that isn't happening yet in these regulated industries. When AI is safely contained, I think that will happen too.

If Glean wins, does knowledge work collapse into one or two interfaces?

I'm future-proofing Glean even for a world of multiple interfaces, where maybe one is Glean, maybe not. As I said earlier, I'm shaping Glean so that even when you use Cursor or ChatGPT or some interface we don't know the name of yet, that product is superior when it runs on Glean as the context engine versus not. So Glean is always in the critical path of helping the knowledge worker do better work with superior context.

Of course, I'd want Glean to be the winning interface if everything boils down to one. But things might boil down to one or even zero. Let me tell you what zero could mean. As an AI-pilled knowledge worker, I run into this all the time: I can't keep up with all the features I'm getting from the AI models.

When I need to present something to you, do I use ImageGen? A PowerPoint deck or an HTML artifact like the one I showed you on indexing versus federation? Which model? Do I even have time before you send your questions to remember which feature to use? We're already seeing that squeeze on users' attention and time.

That's why I use the framing of zero. I think the future of AI is also proactive. We've been talking for 30 minutes, and in my company, so much has happened in that time. Do I need to go to AI and say 'catch me up?' Or can I rely on a proactive Glean that tells me, 'While you were away, an amazing product manager candidate applied to one of your roles, and I think you'll like meeting her, so I scheduled a one-on-one tomorrow in your free slot?' I was recently off for about a week, and Glean helped me get back up to speed, but the work had stopped. It didn't respond to customer complaints or schedule meetings with high-priority candidates.

I think the future of AI is more proactive than reactive, and that's why I think of the interface as almost zero. Maybe there's a future where Glean creates a ton of value, and you hardly go to its interface at all, because it's changing your calendar, scheduling interviews, and reacting to market news for you, and when you're commuting, it tells you through voice what it did and prepares you for your next meeting. That collapse toward zero is exciting.

Quick Poll

Additional Reads

• AI-Powered Work Assistant Glean Lands $150M at $7.2B Valuation -Crunchbase News

• In AI Agent's World: Emrecan Dogan on Building Glean's Next-Gen Search and Productivity - Jotform AI Agents Podcast (Ep. 77)

• Not All Enterprise Context Is Created Equal - Glean

Rate This Edition