OpenAI’s $20 ChatGPT Work Brings Codex Agents to Office Apps, With Access Still a Barrier

ChatGPT Work asks office workers to give an AI system access to the software that holds their schedules, messages, documents, and customer records. OpenAI created the $20-per-month product by adapting Codex, its coding tool, for non-engineers who want an agent to complete multistep work across those systems.

The ambition is to move ChatGPT beyond a prompt-and-response interface. Work can connect with email, Slack, Notion, Figma, Salesforce, Google Calendar, and web browsers, giving an agent the tools and context to carry out longer tasks instead of simply suggesting how a user might do them.

The adoption gap behind the workplace pitch

OpenAI has already seen the difference between building an agent for technical users and getting one adopted more widely. An OpenAI-backed study measured Codex use at 98% among OpenAI employees in June, versus 17% among organizational subscribers and less than 1% among individual subscribers. OpenAI has not disclosed separate user counts for Work and Codex.

98%OpenAI employees using Codex

The OpenAI-backed study measured employee use in June.

17%Organizational subscribers using Codex

Organizational-subscriber use was substantially below the internal rate.

Less than 1%Individual subscribers using Codex

Individual-subscriber use was below 1% in the same study.

More access creates a harder control problem

The product depends on a harness: software around a language model that determines what information it sees, which tools it can use, and how it returns results. For agents, that layer also provides instructions and tools for longer tasks. Its design now becomes a practical constraint on Work’s usefulness.

The current limits are concrete

• Testing described confusing cloud-drive permission setup, including read-only access errors and a prompt saying complete access was required. Some settings were available only on the web app.
• In the article’s testing, Work could create Google Calendar events but could not create a new calendar.
• OpenAI’s harness engineering lead said the effort setting was not yet intuitive for new users.

Coding offered a lesson, not a complete template

OpenAI has already altered Codex after Claude Code emphasized a back-and-forth workflow: present options, take another step after a user chooses, and provide continual updates. OpenAI later added more opportunities for user interaction after concluding its earlier design had moved ahead of what the model and harness could reliably handle.

That revision may be more difficult outside programming. OpenAI uses its GDPval benchmark, spanning 44 occupations and hundreds of knowledge-work tests, alongside user feedback. But presentations, sales pitches, and business strategy are harder to evaluate and trace than code, leaving the reliability of broad workplace automation unresolved.