Glean's AI Strategy: How the $4B Enterprise Search Leader Discovers What 700 Enterprise Customers Actually Want

TL;DR#

Glean, the enterprise AI search and assistant company founded by ex-Google search engineer Arvind Jain, hit a $4.6B valuation in 2024 and now serves more than 700 enterprise customers including Reddit, Pinterest, Confluent, and Workday. The product's core promise — answering "What does my company know?" — depends on one of the hardest customer-research problems in enterprise software: figuring out what questions employees would have asked if they thought their company's knowledge base could answer them. Page-view analytics and click-through logs can't see those questions, because the silent-failure mode of enterprise search is that the employee never asks at all. Glean's response is a layered customer-discovery program — admin interviews with knowledge owners, end-user research with daily searchers, and query-failure post-mortems that turn unanswered prompts into roadmap inputs. This post breaks down what Glean is doing, why analytics-only enterprise-search research breaks at scale, and what AI product teams should copy. Underneath all of it is a category shift: AI-first products treat the agent–customer conversation itself as the primary research instrument, not a post-hoc survey.

What is AI customer research, in the Glean context?#

AI customer research is the discipline of using conversational AI agents — not forms or click-tracking — to interview hundreds or thousands of customers in parallel about what they're trying to do, what they tried, and where the product failed them. In Glean's case, that means understanding what questions a Reddit engineer or a Workday sales rep would ask their company's brain if they trusted it to answer, and why so many of those questions never get typed.

This matters because enterprise search has a unique blind spot. Traditional search analytics — the kind that powered Google Search Quality reviews for two decades — measures clicks, dwell time, and reformulation rates. But internal search has dramatically lower query volume per user than web search, a much smaller and more idiosyncratic corpus, and a failure mode that web search rarely sees: employees giving up silently and Slacking a colleague instead. The questions a workforce would have asked their knowledge base are precisely the questions that drive the most enterprise-AI value, and they're invisible to a logging pipeline. Glean's customer-research program is engineered to surface those invisible questions.

For a primer on why AI-first companies treat the agent-customer transcript as a research artifact rather than a support log, see how AI conversations are replacing surveys and scripts in product discovery. The pattern Glean uses isn't unique to enterprise search — it's the default playbook for any AI company whose product is a conversation.

Why enterprise-search customer research breaks under analytics-only approaches#

Enterprise-search customer research breaks under analytics-only approaches because the most valuable signals — silent failures, abandoned queries, intent ambiguity — never make it into the log file. You cannot A/B-test a question that never got asked.

There are three specific failure modes that hit every internal-knowledge product, and they're worth naming because they show up in every AI assistant rollout, not just Glean's.

Silent failures. When an employee types a query into Glean (or Notion AI, or Microsoft Copilot, or Slack's search) and gets a bad answer, the most common next action isn't "click another result." It's "switch to Slack and ask a teammate." That switch is invisible to the search product. The log records a query, a result list, and a non-click — which looks indistinguishable from a query where the user did find what they needed and just didn't click because the answer was inline. Without a research overlay that asks employees about specific recent queries, the product team can't tell good outcomes from bad ones.

Query reformulation as a signal. When a knowledge worker rewrites the same question three different ways in 90 seconds — Q1 revenue, revenue Q1 2026, where is the Q1 revenue deck — that's a clear signal the first two answers missed. McKinsey's 2024 State of AI report found that employees in AI-adopting firms spend roughly 1.75 hours a day searching for and synthesizing information, and a meaningful share of that is reformulation. Most analytics dashboards bucket reformulation into "engagement" rather than flagging it as a failure pattern. Glean's research function inverts this — reformulation chains are treated as a primary input for relevance tuning.

Intent ambiguity. A query like "PTO policy" could mean ten different things depending on whether the asker is a new hire, a manager planning headcount, an HR partner answering a complaint, or a finance lead modeling accrual. Forms can't disambiguate this — they would have to ask up front, and employees won't fill them out. A conversational layer can ask one quick clarifying follow-up ("Are you asking about your own balance, or the policy in general?") and capture the intent with no friction. This is also why a 30-field survey is the wrong way to learn what employees want from an AI assistant; for more on that pattern, see why AI conversations beat surveys for real customer research.

The takeaway: enterprise search and enterprise AI assistants generate less clean behavioral signal than a consumer product, not more, because the corpus is small, the user base is small, and the failure mode is silence. Customer research has to do the work the logs can't.

Inside Glean's customer-discovery program#

Glean runs a three-layer customer-discovery program — admin interviews, end-user research, and query-failure post-mortems — because no single research method can see all the failure modes of an enterprise AI assistant. Each layer answers a different question that the others can't.

Layer 1: Admin and knowledge-owner interviews#

Admin interviews answer the question "what does the company think it knows, and where are the canonical answers supposed to live?" Glean's deployment model — connecting to Confluence, Google Drive, Slack, Notion, GitHub, ServiceNow, Salesforce, and dozens of other corpora — only works if someone inside the customer org has explained which sources are authoritative for which questions. That mapping isn't documented anywhere. It lives in the heads of IT, knowledge-management, and ops leads.

Glean's customer-success and product teams interview those admins early and often. The pattern is the same one Anthropic uses to study how enterprises actually use Claude and that Atlassian applies to Jira and Confluence: treat the admin as a power user with a roadmap of their own. The interviews surface canonical-source conflicts ("Slack thinks the policy is X, the HR wiki thinks it's Y"), permissioning edge cases, and the implicit hierarchy admins want the AI assistant to follow.

For teams trying to copy this layer, an admin / power-user interview template is the right starting point.

Layer 2: End-user research with daily searchers#

End-user research answers the question "what would you have asked, if you trusted the assistant to know?" This is the harder of the three layers because the population is large, distributed, and uninterested in talking to product researchers — most end users wouldn't show up to a Zoom interview.

This is where AI customer interviews change the math. Instead of recruiting 12 users for 30-minute Zooms and hoping the sample is representative, an AI interviewer agent can run 200 to 800 short asynchronous conversations across the workforce, each pulling at threads like "Tell me about the last question you tried to find an answer to at work — even if you didn't end up using Glean." For a step-by-step on how this scales, see how to run AI-moderated customer interviews at scale. Forrester estimates that enterprise UX teams spend 40% of research time on recruiting and scheduling alone — a number AI interviews collapse to near zero.

The end-user layer is also where Glean catches the verb-vs-noun problem in queries. An end user who asks "vacation" and an end user who asks "how do I request vacation" want fundamentally different things — one wants policy, the other wants a workflow. Log analytics can't reliably separate the two. A short, structured conversational follow-up can.

Layer 3: Query-failure post-mortems#

Query-failure post-mortems answer the question "this query got a bad answer — why?" Glean's relevance-engineering team treats failed queries like incidents: each one gets traced through the retrieval pipeline (which corpora were searched, which documents ranked, which got passed to the LLM for synthesis) and matched against an interview transcript where possible.

The post-mortem layer is where Glean's customer-research function meets its ML org. The pattern is similar to what Scale AI's forward-deployed engineers do with their enterprise AI buyers, what Harvey AI does inside BigLaw deployments, and what Mistral's FDE team does in European LLM rollouts: a tight loop between the engineer who can fix the model and the customer interview that explains why it needs fixing. The broader role pattern is covered in the State of Forward Deployed Engineering 2026 report.

The output of the three layers stacks: admin interviews tell Glean what should be findable, end-user research tells Glean what employees try to find, and query-failure post-mortems tell Glean where the gap is. The roadmap falls out of the gap.

The Glean Assistant feedback loop — agent-customer conversation as research input#

Glean Assistant — the company's chat-style AI agent layered on top of search — generates research data with every conversation, which means the product itself is now Glean's largest customer-research instrument. Each assistant turn is a question asked, a context provided, and an answer judged.

This is the structural advantage that AI-first companies have over the previous generation of enterprise software. A traditional SaaS product had to bolt on a feedback widget or quarterly NPS survey to learn anything about users. An AI agent has the user's actual question, the actual context, and (with thumbs-up/thumbs-down or follow-up signals) the actual judgment of the answer — all in one transcript. The transcript is the interview.

The same pattern shows up across the AI-native frontier: how Perplexity learns from 50 million searchers, how Cursor reads a million developers' coding sessions, how Sierra AI uses the agent transcript itself as research input, and the broader pattern in Notion's customer research function and Datadog's customer research strategy. In each case the agent is doing two jobs: serving the customer, and generating a structured record of what the customer wanted.

For Glean specifically, the assistant transcript is where the team catches:

• Out-of-corpus questions — employees asking the assistant things their company hasn't documented, which becomes a knowledge-management roadmap input.
• Workflow gaps — "Glean, can you also draft the reply?" — where the agent should plug into a downstream system.
• Persona drift — sales reps asking Glean very different questions than engineers, which feeds tuning of vertical-specific prompts and retrieval weighting.

Done well, this loop changes the cadence of product research from "quarterly studies" to "continuous learning at the speed of conversation." For a framework that any AI product team can apply to their own assistant, see the continuous discovery stack for AI-first product teams.

What this signals for enterprise AI product teams#

Enterprise AI product teams should treat conversational data — both interview transcripts and live agent traffic — as a first-class research asset, not a byproduct of support. The shift from log-based analytics to conversation-based research is the single largest tooling change happening in enterprise product orgs in 2026.

Three implications fall out of that shift, regardless of whether your product is enterprise search, an AI coding tool, an underwriting assistant, or a customer-success platform.

1. Your agent is your largest research panel. If you have 10,000 daily active users of an AI assistant, you have 10,000 daily research conversations. The team that builds the muscle to mine those conversations for roadmap signal will out-iterate the team that runs quarterly NPS. This is the explicit thesis behind the 2026 customer discovery velocity report, which found AI-first teams cut time-to-insight by 94% versus survey-led teams.

2. Static forms are the wrong intake layer. If you're collecting onboarding context, sales-qualification, or admin-setup data with a 20-field form, you're losing the highest-signal users and the highest-signal answers. Forms collapse "it depends, because we're a hybrid Salesforce + HubSpot shop with three regional knowledge bases" into a dropdown. A conversational intake captures the whole shape. See why the discovery form is the worst bug in B2B SaaS and the rise of the conversational funnel for the broader pattern.

3. The buyer for AI customer research has moved. Two years ago, "customer research tools" was a UX-research line item. Today the buyer is increasingly the product-led growth team, the CS leader, or the founder. According to Gartner's 2024 Voice of the Customer Magic Quadrant commentary, the category is consolidating around platforms that turn conversational data into action — not platforms that build prettier survey dashboards. The 2026 customer research budget report found CMOs saving over $1M annually by replacing vendor studies with AI-conducted interviews.

For Perspective AI customers building enterprise AI products, the Glean pattern is the reference implementation. Use the AI interviewer to run end-user research at scale, use a concierge intake agent to replace static forms in onboarding and qualification, and feed both transcripts plus your live agent traffic into a single research repository. Teams that adopt this stack — see the customer research stack that modern product and CX teams actually use — are running discovery cycles in days instead of quarters.

Frequently Asked Questions#

What is Glean AI and what does the product do?#

Glean is an enterprise AI assistant and search platform that connects to a company's internal applications — Slack, Confluence, Google Drive, Notion, GitHub, Salesforce, ServiceNow, and more — and lets employees ask questions in natural language. Founded in 2019 by ex-Google search engineer Arvind Jain, Glean reached a reported $4.6B valuation in 2024 and serves more than 700 enterprise customers. The product blends retrieval-augmented generation (RAG) over the company's knowledge corpus with a chat-style assistant called Glean Assistant.

How does Glean conduct customer research with enterprise buyers?#

Glean conducts customer research in three layers: admin interviews with the knowledge-management and IT leads who configure the deployment, end-user research with the employees who use Glean daily, and query-failure post-mortems that trace bad answers back through the retrieval pipeline and match them against interview transcripts. The three layers stack — admin interviews surface what should be findable, end-user research surfaces what people try to find, and post-mortems surface the gap.

Why do analytics-only approaches fail for enterprise search?#

Analytics-only approaches fail for enterprise search because the most valuable signals are silent. When an employee can't find an answer, they typically switch to Slack and ask a colleague — that switch is invisible to the search product's log file. Query reformulation looks like engagement, intent is ambiguous, and the small per-user query volume in enterprise search starves any pure-statistics approach. Customer-research interviews surface the intent and outcome data the logs miss.

What is enterprise RAG and how is it different from consumer search?#

Enterprise RAG (retrieval-augmented generation) is a pattern where an AI assistant retrieves relevant documents from a company's private corpus — wikis, files, tickets, code — and feeds them to an LLM to generate an answer grounded in that material. It differs from consumer search in three ways: the corpus is permissioned (the assistant must respect who can see what), the query volume per user is much lower, and the failure mode is silent abandonment rather than result-page bounce. Those differences make customer research more important, not less.

Who is Arvind Jain and what's his background?#

Arvind Jain is Glean's co-founder and CEO. Before founding Glean in 2019, he spent more than a decade as a distinguished engineer on Google's search team, working on ranking and infrastructure. He also co-founded Rubrik, a cloud-data-management company that went public in 2024. His core thesis at Glean is that enterprise search has been broken for two decades and that the LLM era is the first time a company can credibly answer "what does my company know?"

How can my team copy Glean's customer-research approach?#

Your team can copy Glean's customer-research approach in three steps. First, run a small batch of admin or power-user interviews to map your authoritative-source landscape — a structured customer-interview template is the right starting point. Second, deploy an AI interviewer to run end-user research at scale and capture the intent behind real workflows; see the AI-moderated customer interviews playbook. Third, treat your own product transcripts as a research corpus — review the conversations weekly with product, ML, and CS in one room. Built for product teams, Perspective AI runs all three layers in one platform.

Conclusion#

Glean's $4B run hasn't been a story about a better search algorithm. It's been a story about a company that built a customer-research function fit for the failure modes of enterprise AI — silent abandonment, query reformulation, intent ambiguity — and used that research to compound a roadmap advantage 700+ enterprise customers deep. The lesson for every AI product team is that AI customer interviews aren't a marketing layer on top of your product; in an agent-first world, they are the product's nervous system.

If you're building an enterprise AI product and you're still running discovery through static forms or quarterly NPS surveys, you're shipping with the wrong sensor stack. Start a Perspective AI study to run conversational customer interviews at Glean-scale, see how the interviewer agent works, or browse the use-case library for the workflows other enterprise AI teams are running. The AI customer interviews running on your own users today are the data your roadmap will live or die on tomorrow.