Ninety-eight percent of OpenAI employees used Codex in June. Outside the company, fewer than 1% of individual subscribers touched it. That single statistic explains everything about where agentic AI actually stands right now, and why OpenAI is scrambling to close the gap.
According to TechCrunch, the company released ChatGPT Work last month at the $20-per-month subscription tier. It’s a modified version of Codex, OpenAI’s coding-focused agent tool, rebuilt to serve accountants, investors, communications teams, and anyone else whose job runs through a computer but doesn’t involve writing software. The pitch is straightforward: the same autonomous, multi-step task completion that developers get from tools like Claude Code or Codex, but without requiring users to touch a command line.
Andrew Ambrosino, the lead engineer on OpenAI’s desktop app, has given the product access to his inbox, Slack, Notion, Figma, and his phone. He acknowledges the risk. A private DM could surface in a document the model writes. He says it hasn’t happened, but he’s accepted that it might. That level of trust is exactly what OpenAI needs mainstream users to develop, and right now, most of them don’t have it.
The business logic is clear. Agents that run longer tasks consume more tokens, which increases revenue per user. And coding, while lucrative, is a narrow slice of professional work. If OpenAI and its competitors want to justify the billions spent on training and compute, they need to reach lawyers, analysts, salespeople, and operations teams. But that space already has focused competitors. Harvey is going after legal. Clay is chasing sales. Both take a model-agnostic approach, which means they’ll swap in whichever underlying model performs best. OpenAI doesn’t have that flexibility with its own product.
The internal challenge is just as real. When OpenAI’s non-engineering staff started using Codex, the tool asked them about code and returned outputs like empty diffs, technical readouts that meant nothing to a finance manager. The team spent months making it more general-purpose. That work is what ChatGPT Work represents.
Still, the interface question is contested even inside OpenAI. Some engineers argue that buttons are unnecessary if users can just prompt the model directly. Ambrosino pushes back. Discoverability matters at this stage of adoption, he argues, comparing the current design philosophy to skeuomorphism, the early practice of making digital tools look like physical objects to ease the transition. The calculator app that looked like a pocket calculator wasn’t bad design. It helped people cross over.
The joint ChatGPT desktop app currently has around 20 million users. ChatGPT on the web has over a billion. That gap is the entire problem OpenAI is trying to solve, and ChatGPT Work is its current best answer.
- Weekly automated metrics reports, replacing manual data pulls
- Spreadsheets converted into live planning tools
- Investment memos assembled from Slack threads and email
- Custom dashboards built from unstructured internal data
- Engineering problem analysis turned into visual charts on request
These are the use cases OpenAI is leading with. They’re real, they’re useful, and they’re exactly the kind of routine, data-heavy coordination work where agents have a clear advantage over a human doing it manually. But getting the average worker to hand over access to their inbox and Slack requires a level of trust that most people haven’t built yet. OpenAI is betting that the product experience, not just the model capability, is what gets them there.
