Generative AI promised a conversational route to workplace information, but Glean founder and CEO Arvind Jain argued that conversation was only the visible layer. For an answer to be useful inside a company, the underlying system needed strong search, knowledge of organisational context and strict enforcement of existing access permissions.
Glean was built by former Google search engineers and other industry veterans to address a growing workplace problem: employees often cannot find the information or colleagues they need. Remote work and the rapid spread of software-as-a-service applications had made that knowledge more fragmented.
Workplace Search Must Understand Each Company
Glean retrained deep-learning language models on each organisation’s own knowledge base. Jain described the resulting system as a trusted knowledge model that learned the company’s context, vocabulary, behaviour and relationships. The objective was to rank answers according to the environment in which employees actually worked.
Generic web knowledge is insufficient for questions about an internal process, project or colleague. The same term can mean different things across organisations, while useful evidence may be distributed among documents, messages and specialised applications. Enterprise search must connect those signals without treating every source as equally current or authoritative.
Personalisation also needs limits. Understanding a person’s role can improve relevance, but the system should not create hidden profiles that employees cannot inspect. Search quality depends on context, yet trust depends on explaining which context is used and keeping it tied to legitimate workplace needs.
Permissions Are Part of Answer Quality
Glean connected to enterprise knowledge while preserving the access rules already applied by source systems. Its governance engine was designed so that users saw only information they were entitled to view. That made security a foundational search feature rather than a filter added after retrieval.
This is essential when an AI system synthesises several sources. A conventional search result might expose a permission error through one visible link; a generated answer could absorb restricted information without revealing where it came from. Permission checks must happen before content reaches the model and remain effective throughout the response process.
Companies also need to test changes over time. Employees move roles, projects end and access policies evolve. A grounded assistant is dependable only if permission updates flow through quickly and if audits can show which sources supported a particular answer.
Productivity Comes From Less Time Searching
Jain said the average Glean user performed 20 searches each day and saved two to three hours per week. The company treated those time savings and productivity gains as core success measures. The figures reflected how small delays accumulate when information is spread across many tools.
Search can also affect engagement. The interview cited findings that difficulty locating data or information was a major reason employees felt disconnected. A useful knowledge system helps people complete tasks, but it can also make a distributed organisation feel more navigable by revealing expertise and relationships.
Measurement should go beyond the number of searches. A rising query count might show adoption or continuing frustration. Teams should examine whether employees reach correct answers, avoid duplicated work and act with greater confidence. Qualitative feedback can identify where the system’s ranking misses local context that usage statistics cannot explain.
Generative AI Needs a Grounded Retrieval Layer
Jain said GPT-4 and similar models were not ready for enterprise use by themselves because they needed to be grounded in the right search technology. Glean’s ambition was to provide a workplace product as useful as ChatGPT was for the web, while relying on its trusted knowledge model.
The conversational interface was only one component. Ranking remained foundational because the quality of a generated answer depends on the evidence retrieved before generation. If the system selects an outdated policy or an irrelevant discussion, fluent language can make the error harder to notice.
Grounding should therefore remain visible to users. Source links, document dates and ownership help employees judge whether an answer deserves trust. The interface can make discovery faster without removing the user’s ability to inspect the material behind a recommendation.
Glean also wanted deployment to require minimal operational overhead and no mandatory professional-services engagement. Ease of setup is valuable, but every company still needs governance ownership. Someone must define trusted sources, review permissions, handle disputes and evaluate whether the system supports important workflows safely.
The interview’s durable insight was that enterprise AI is a knowledge-management problem before it is a chat problem. Large models can summarise and converse, but organisations need accurate retrieval, local context and permission-aware evidence. Those less visible capabilities determine whether an assistant saves time or simply produces polished uncertainty.
As generative features expand, the strongest enterprise products will preserve that order of operations. They will retrieve authorised information, rank it well, show evidence and then use generation to make it easier to understand. Conversation improves access; trusted search makes the conversation worth having.
Adoption should be introduced with the same care as the technical system. Employees need to know which repositories are indexed, how permissions work and when an answer should be verified with a source owner. Teams can begin with common, low-risk questions and expand after measuring accuracy and usefulness. This gradual approach creates evidence about where search genuinely removes friction and where process or documentation must be fixed instead. An assistant should not conceal weak knowledge management; it should help the organisation identify and repair it.
