OpenAI’s head of product reveals the massive adoption gap holding back AI agents

OpenAI has an internal adoption problem that most companies would kill for, and an external one that keeps its leadership up at night. Nearly every employee at the company uses its Codex tool. Outside the building, almost nobody does.

Thibault Sottiaux, who leads core product work at OpenAI and reports to co-founder Greg Brockman, laid out those numbers in a recent interview with TechCrunch. About 98% of OpenAI’s own employees actively use Codex. Among organizational subscribers, just 17% engage with AI agents. For individual subscribers, the figure drops below 1%.

The gap between builders and users

Sottiaux’s explanation centers on user experience. The current generation of AI agents requires too much configuration, too many permission settings, and too little guidance on what they can actually do.

OpenAI’s proposed solution is ChatGPT Work, which launched in July 2026 at $20 per month. The product targets non-engineers specifically, designed to handle autonomous, multi-step tasks across tools like email, Slack, and calendars.

Brockman’s consolidation play

Sottiaux’s reporting line matters here. Greg Brockman, who co-founded OpenAI, has taken on broader oversight of product strategy as part of an organizational restructuring. The goal is to merge ChatGPT, Codex, and API initiatives into a single platform framework.

Instead of running separate product tracks that occasionally overlap, OpenAI is building what it calls a unified “agentic platform.” The long-term vision resembles something closer to a super app, one interface that handles everything from coding assistance to workflow automation to general knowledge queries.

Why discoverability is the real bottleneck

Sottiaux pointed to discoverability as a core challenge. Users don’t know what agents can do, so they don’t try. And when they do try, permission settings create friction. Connecting an AI agent to your email, calendar, and Slack simultaneously requires a level of trust and configuration that most people aren’t ready for on day one.

OpenAI’s product team is working on what Sottiaux described as gradual simplification of interfaces. The idea is to lower the barrier at every step: show users what’s possible before asking them to grant access, then expand capabilities as trust builds.

The 98% internal adoption rate actually supports this approach. OpenAI employees understand what the tools can do because they’re immersed in the culture and have colleagues who can show them. External users don’t have that context, so the product itself needs to provide it.

What this means for the AI landscape

For businesses evaluating AI tools, the honest admission from OpenAI’s own product lead that external adoption remains strikingly low should be both cautionary and informative. Companies that invest in onboarding and training alongside the software license are likely to see far better returns than those that simply hand employees a login and hope for the best.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.