OpenAI tries to build AI agents for everyone, not just programmers

From coding to everyday office work

OpenAI is trying to get its AI agents to break out beyond their traditional niche: software development. Last month the company launched ChatGPT Work, available for $20 a month, aiming to make independently working AI helpers accessible to accountants, investors, doctors, and other office professionals.

Enthusiastic internally, still hesitant externally

Within OpenAI itself, the company says 98% of staff use Codex, the underlying agent technology, on a daily basis. Outside the company the picture looks very different: only 17% of business subscribers and less than 1% of individual users actually use Codex. That gap illustrates both the challenge and the opportunity OpenAI sees ahead of it.

Caution around access to private data

Andrew Ambrosino, head engineer of OpenAI's desktop application, points to a tricky trade-off. Give an AI agent access to, say, private messages, and you risk it accidentally pulling that sensitive information into a document where it doesn't belong. Still, he argues for broader access, despite the risks involved.

Competition from Anthropic

Engineer Joe Gershenson stresses that a stronger underlying model ultimately matters most, but acknowledges that product design matters too. Anthropic's Claude Code reportedly does better by actively presenting users with options and asking for confirmation, rather than working fully autonomously right away the way OpenAI initially did.

Price may be a barrier

The subscription model also runs into practical friction: the author of the original article used more than 80 million tokens in four days, amounting to $65 in costs — more than three times the monthly subscription fee under normal use. OpenAI product lead Thibault Sottiaux acknowledges that work is still ongoing on efficiency, including through recent price cuts.