Why We’re Only Using 1% of AI | Glean CEO Arvind Jain

Most engineering teams assume their best developers are already maximizing AI productivity gains. Arvind Jain has data proving they're wrong. The Glean CEO's $7.2 billion AI assistant company tracks how knowledge workers actually use artificial intelligence, and the results should worry every tech leader betting on AI-driven productivity.

Senior engineers at top companies are barely touching AI tools. "Some really good engineers, they have not they haven't adopted AI that much," Jain explains. "But they're still productive because remember that AI has become, like, helpful in terms of writing new lines of code, but a lot of work that the most senior engineers do is not writing new lines of code." The most valuable engineering work involves architecture decisions, debugging complex systems, and navigating organizational knowledge. Current AI tools excel at code generation but fail at the higher-order thinking that separates senior talent from junior developers.

This adoption gap reveals a fundamental misunderstanding about where AI creates value in knowledge work. Most AI tools target the wrong bottleneck. Writing code isn't the constraint for experienced engineers. Understanding legacy systems, finding the right person who made a critical decision two years ago, or figuring out why a service behaves differently in production than in staging consumes most senior engineering time. Glean built its business around this insight, creating AI that searches across internal company knowledge rather than just generating new content.

The implications extend beyond engineering teams. If highly technical workers who understand AI capabilities aren't adopting these tools for their most important work, other knowledge workers face even steeper challenges. The promise of AI productivity gains assumes people will naturally integrate these tools into existing workflows. Jain's experience suggests that assumption is deeply flawed.

Growth at AI companies creates unique operational challenges that traditional scaling playbooks don't address. When Glean crossed 1,000 employees, Jain experienced panic rather than celebration. "I was also surprised, you know. It was not a just just to be clear, it was not a feeling of, you know, celebration. It was a feeling of panic," he admits. The number itself triggered anxiety about maintaining company culture and decision-making speed while building products in a rapidly evolving market.

This reaction reflects deeper tensions in AI company scaling. Traditional enterprise software companies can optimize for stability once they find product-market fit. AI companies must continuously reinvent core capabilities as underlying models improve and new use cases emerge. Adding people accelerates development but also increases coordination costs at exactly the moment when technical requirements are shifting most rapidly. Jain's panic signals awareness that scale can become a liability when market conditions demand constant adaptation.

Geographic distribution amplifies these coordination challenges. Despite pressure to open San Francisco offices, Glean deliberately kept engineering teams collocated. "The main reason for us not to have an office in San Francisco for a long time was because we wanted the team to be together," Jain explains. "Our R and D team, our technical leaders, managers, even the ones who are living in San Francisco didn't like the idea of that, like, people are going to be in two different offices." Technical leaders recognized that distributed teams would slow decision-making when product requirements change weekly rather than quarterly.

The broader lesson applies beyond AI companies. Knowledge work increasingly requires rapid iteration on complex problems where context and nuance matter more than raw output. Tools that optimize for content generation miss the real bottlenecks in modern work. Understanding what happened in last quarter's customer calls, finding subject matter experts for technical decisions, or accessing institutional knowledge from departed employees creates more value than writing new emails or code from scratch.

This suggests a different framework for evaluating AI productivity tools. Instead of measuring lines of code generated or documents created, leaders should track time spent searching for information, duplicated work across teams, and decisions delayed due to missing context. The companies that win with AI will be those that solve knowledge discovery problems rather than knowledge creation problems.

The path forward requires rethinking how knowledge workers interact with information systems. Current AI tools mostly operate in isolation, generating content without accessing company-specific context. The next generation must integrate deeply with existing workflows and institutional knowledge. This means AI that understands organizational structure, historical decisions, and cross-team dependencies rather than just language patterns from internet data.

Watch for AI tools that prioritize search and discovery over generation. Monitor whether your most experienced knowledge workers are actually adopting AI tools for their most important work. If adoption remains low among your best performers, the tools probably solve the wrong problems. The companies that figure out AI-powered knowledge discovery first will create sustainable competitive advantages while others chase productivity gains that never materialize.