OpenAI pushes agents into every workflow, adoption remains limited

OpenAI launched ChatGPT Work this month, a tailored version of its Codex coding tool aimed at knowledge workers who are not engineers. The offering is already available on the $20-per-month basic subscription. ChatGPT Work lets users attach a language model to email, Slack, Notion, Figma and other productivity apps, and run multi-step tasks without constant human oversight. The company envisions a future where “artificial intelligence moves from answering questions to helping turn big ideas into reality,” according to its statements.

Thibault Sottiaux, who leads OpenAI’s core product line including Work, explained that the tool was built to bring the same autonomous capability developers already receive from code-generation agents—writing entire projects, not just point answers—to finance, legal, medical and other departments that work behind a screen. That shift, he said, requires a significant relinquishment of control: the agent must have access to private information, direct messages and internal documents.

Andrew Ambrosino, a senior engineer on OpenAI’s desktop application, gave the app full access to his email, Slack, phone and personal work tools. He warned that the model could pull private data from a direct message and share it without knowing it is prohibited, saying to TechCrunch, “I’ll do it for the work. I’ll take the personal hit here and there if needed. And I still shouldn’t have.” This approach reflects an internal belief that without full access there is no way to test the product’s true limits.

Agents that run for longer periods consume more tokens, making them more profitable per user. The challenge is that the developer community is too small to justify the massive investments required for training and compute. Domain-specific competitors such as Harvey for lawyers and Clay for sales teams are already chasing the same customers with model-agnostic approaches, connecting whichever model performs best at the moment. Christian Catalini wrote on the a16z blog that if labs do not quickly acquire the complementary assets needed to expand the market, value will flow elsewhere.

An internal study released in June highlighted the adoption gap: 98% of OpenAI employees used Codex, versus 17% of enterprise subscribers and less than 1% of individual subscribers. Ambrosino described the pre-adjustment situation as communication and finance teams encountering a tool that was “actively hostile,” asking it for code and receiving an empty diff as output. Between February and the current launch, the company made the product more user-friendly for those who do not write code for a living.