Glean bets the enterprise AI battle will be won with context, not more models

Glean has turned its annual customer conference, Glean:GO, held in San Francisco, into a statement of principles on how AI reliability should be solved in the corporate environment. Against an industry that, according to the company, has responded to model errors by stacking layers of oversight —evaluators, guardrails, human review queues— Glean argues the problem must be tackled further down: an agent that gets things wrong does not need more supervision, it needs to have been given the right information to do its job in the first place. That idea, 'context', came up as many as 61 times during a session with press and analysts, in the words of co-founder and chief executive Arvind Jain, who defined it as 'all the right information, expertise and human wisdom needed to complete a task'. According to Jain, that knowledge already exists inside companies, but it is buried, scattered and mixed in among structured documents and informal signals people leave in chats, meetings and approvals.

This insistence is no accident: it marks a narrative shift for a company long seen as an enterprise search engine. Co-founder and head of product engineering Tony Gentilcore recalled Glean's original inspiration in the old HP line about how, if a company knew everything it knows, it would be ten times more productive. That purpose, conceived so employees could find information instantly, is now being repurposed as the infrastructure that feeds context to AI agents and models. 'We found that the same thing we built for humans to do their jobs is needed for agents and models to do theirs', Gentilcore explained.

The market Glean competes in is, however, increasingly crowded. On one side are agentic knowledge-work tools such as Anthropic's Claude Cowork, Microsoft's Copilot Cowork, OpenAI's ChatGPT Work, and offerings from Atlassian, Salesforce and ServiceNow. On the other are the enterprise search vendors Glean came from, including Slack, Lucidworks, Elastic and Algolia. To differentiate itself, Glean published internal benchmarks comparing its Glean Assistant with Claude Cowork running Sonnet 5 across more than 180 knowledge-work tasks, using synthetic queries over its own production data and comparative evaluation on a five-point scale. According to those results —with automatic model routing enabled in Glean versus Sonnet 5 in high-reasoning mode in Cowork— Glean's cost per task was $0.58 versus $2.98 for Cowork, an 81% saving in token costs, with evaluators preferring Glean's answers in 78% of cases.

One of the most striking new products was Glean Tau, a desktop experience that, unlike the rest of the platform —which runs entirely in the cloud— executes on the worker's own computer. It combines local files, code and documents with Glean's context graph, allowing the AI to write and run code on the device and carry out complete workflows without relying solely on the cloud. Executive Emrecan Dogan noted that Tau has already been deployed internally and that its use has soared across different types of tasks without costs soaring in step, precisely because it combines cloud context with context that lives only on the employee's machine. It is an implicit acknowledgment that much real work —local folders, files that never sync— falls outside the reach of a purely cloud-based index, an area where Microsoft is also advancing with local AI capabilities in Windows 11 and its Copilot+ PCs. Glean has not given an official launch date for Tau beyond 'coming soon', though a spokesperson pointed to a timeframe of about three months.

Another central theme of the event was so-called 'proactive AI'. According to research by Glean's Work AI Institute, employees spend an average of 6.4 hours a week —almost a full working day— on what the company calls 'botsitting': checking that the AI has the right context, debugging errors, repeating instructions and correcting answers that are 'confident but wrong'. Dogan argues that the current AI model is reactive —the user has to remember to invoke it, spell out the steps and know which agent is for what— and that this is not sustainable because human attention is limited. Glean's answer is three features aimed at personal productivity: proactive task management, which identifies next steps and generates first drafts; email triage, which flags what is urgent and drafts context-aware replies; and a 'meeting coach' that helps prepare meetings, guides them as they unfold and outlines the conversations afterwards. It is worth noting that proactivity is not an idea exclusive to Glean: Salesforce and Asana had already floated it, though Glean insists on presenting it as digital colleagues that continuously observe how people work and start tasks on their own initiative.

Beyond personal productivity, Glean introduced its 'independent agents', comparable to Amazon Web Services' Frontier Agents: autonomous programs that persist across sessions, are shared among employees, have their own identity and permissions, and a mandate that defines the scope of their work. According to Dogan, these agents self-correct and improve with use, incorporating the feedback received in each run. Jain drew a distinction between two categories: agents that act on behalf of a worker —limited to the data that employee can access, operating within a narrow scope inside a single application— and those with their own identity, which run continuously and can initiate actions on their own, which he compared to 'employees or colleagues'. This second category, Jain acknowledged, opens up a new set of problems: role and permission design, defining the scope of responsibility, rules of behavior and collaboration, and mechanisms for accountability and oversight. The technical origin of these agents is especially revealing: Gentilcore confirmed that Glean's team went as far as forking the OpenClaw project, but abandoned that path after concluding it was not secure enough, since they were 'recreating part of the stack without the corresponding guardrails'. They then migrated everything to Glean's main agent engine, though Gentilcore himself admitted that, in essence, its independent agents are 'claw agents, without the lobster analogy'. The first three independent agents —for on-call management, voice of the customer and sales— are in beta.

The rest of the announcement included Glean Transform, a 'living map' that uses the platform's enterprise graph to observe how work flows across email, support tickets, documents and meetings, generating reports with suggested agents or automations deployable in one click and measuring their business impact; it is listed as 'coming soon'. Also unveiled was Team Chat, which lets employees, Glean's assistant and agents collaborate in the same thread or shared document —a feature already available in beta that, according to Dogan, moves AI 'from single-player to multiplayer', in line with similar moves by Slack and Microsoft. Added to that are interactive dashboards combining structured data with unstructured business context (available in beta), an expansion of its AI Gateway —launched in July— to cover more AI entry points, enforce restricted-topic policies through Glean Protect and broaden governed MCP access to organizational skills and personal memory (also in beta), and a context-based threat detection system that analyzes whether an agent's behavior is legitimate, still at the 'coming soon' stage.

On the business side, these launches come three months after Glean announced its annual recurring revenue (ARR) had passed $300 million, a figure the company says it has tripled since the start of 2025. Gentilcore said 85% of its customers deploy the platform organization-wide, and that daily active users as a share of monthly active users reaches 45% —that is, almost half of those who use Glean in a given month do so daily. Asked whether Glean could become the first AI infrastructure company to top $1 billion, Jain avoided a direct answer, saying only that there are internal plans he cannot share and that 'the market keeps surprising us'. He was more explicit about the business model: Glean is pursuing a SaaS-style revenue scheme rather than one tied to token consumption because, in his words, AI is very expensive for customers today and the company wants to position itself as the one helping businesses manage that cost, not the one profiting from runaway usage. That distinction —charging for subscription value rather than consumption— is, ultimately, consistent with the central argument of the whole event: if the right context reduces the number of failed attempts and repeated tasks, it also cuts token spending, turning Glean's context architecture itself into a commercial argument, both technical and economic, against rivals that charge for heavy model use.