OpenAI pushes ChatGPT Work into office tasks

The company wants ChatGPT to handle multi-step work across email, Slack and calendars, but questions remain over access, setup and trust.

OpenAI is trying to move ChatGPT beyond answering questions and into everyday office work. Its new ChatGPT Work product is designed to act more like an AI agent, taking on multi-step tasks across email, Slack, calendars, documents and other workplace tools. For businesses and office workers, that promise is easy to understand. It could cut down on the constant switching between apps — but only if users are willing to give the system wide access to personal and company data.

OpenAI wants ChatGPT to do the work, not just talk about it

OpenAI’s internal goal is to make AI useful for white-collar work typically handled by accountants, investors, doctors and other office employees. ChatGPT Work is a modified version of OpenAI’s Codex coding tool, adapted for people outside engineering.

The company’s pitch is simple: instead of only answering prompts, the system should carry out longer tasks on its own. OpenAI says that fits its broader mission of helping people turn ideas into real work.

There is also a clear business case behind that push. Agent-style products tend to use more tokens over longer sessions, which can make them more profitable per user. Reaching professions beyond software development matters not just for OpenAI, but for the wider AI industry, which needs broader adoption to justify the huge cost of training and running these systems.

ChatGPT Work depends on access to workplace tools

To function as an agent, ChatGPT Work connects to the digital services people already use. In practice, that can include inboxes, Slack, phone data, Notion, Figma, Google Calendar and other tools. OpenAI says the model is meant to pull information from those places and act on it.

That design is also what makes the product sensitive. Andrew Ambrosino, the lead engineer for OpenAI’s desktop app, said the app has access to his inbox, Slack account, phone and apps such as Notion and Figma. He acknowledged that, in some cases, the system could pull information from a private message if asked to write a document, though he said he has not had a serious problem with it.

The trade-off is hard to miss. The more access the model gets, the more useful it can become. But users still have to decide how much control they are willing to hand over to software that can read, organize and potentially act on their behalf.

OpenAI is trying to move beyond engineers

OpenAI’s first successful agentic tools were built for software developers, where the payoff is easier to measure. Codex and similar tools can help write code, and engineers are already used to command-line interfaces and technical workflows.

That model does not carry over neatly to other jobs. OpenAI’s own staff found that early versions of Codex were too technical for non-engineers, with interfaces that assumed users already understood code. The company says it has been making the product more general-purpose since February.

This is where ChatGPT Work fits in. OpenAI wants it to be intuitive enough for people who do not work in software, even if that means adding more guidance, buttons and setup steps than some engineers inside the company would prefer. According to IT-PUB News, there has also been internal debate over how much interface help users need when they could simply ask the model directly.

OpenAI argues that discoverability still matters. Put simply, people need clear entry points before they are ready to trust a model with broader tasks.

OpenAI's internal use is far ahead of outside adoption

The company is betting that people will adopt these tools once they see the value. So far, though, the usage gap suggests that is not automatic.

An OpenAI-backed study found that in June, 98% of OpenAI employees were using Codex, while only 17% of organizational subscribers and less than 1% of individual subscribers were using the agentic coding tool. That gap shows the challenge clearly: a tool may work well inside the company that built it and still struggle to win over outside users.

OpenAI has not said how many people use ChatGPT Work compared with Codex, but it did say the joint app is used by 20 million people. That remains far below the more than a billion users the company says are prompting ChatGPT online.

The difference matters. OpenAI needs more than curiosity from users. It needs them to come back often enough — and trust the tool enough — to make agent-based products part of their routine.

Convenience comes with setup problems and limits

The product can be useful when it works. OpenAI describes it as best suited for routine, data-heavy coordination tasks such as weekly metrics reports, spreadsheets used for planning and assembling information from multiple sources.

The source text gives examples including creating a calendar entry from an email, building dashboards from financial data and turning information from Slack into charts. It also points to rough edges. Setting permissions for cloud-drive access was confusing, and some settings were available only in the web app. In one case, the system could create events in Google Calendar but not new calendars.

Joe Gershenson, OpenAI’s engineering lead for the harness, said the effort settings are not intuitive for new users yet. He said the company is working to improve how the product helps people choose the right level of reasoning.

That matters because a tool that is too cautious can feel limited, while one that pushes too far can become unreliable. As the source puts it, if users do not push it enough, they may end up with “the worst intern” they have ever worked with.

Cost, lock-in and trust remain open questions

There is also a business issue behind the product design. Agentic systems can be expensive to run, and heavy use can quickly consume tokens. The source says that one four-day stretch of casual use on a $20-a-month subscription used more than 80 million tokens, with a model-estimated cost of $65. OpenAI says it is working on efficiency and points to a recent 80% price cut for users of its Luna model.

Another concern is lock-in. The more data a user saves in the system — and the more accounts and permissions they connect — the harder it may be to switch away later.

For now, OpenAI is presenting ChatGPT Work as a step toward a future where AI does more than answer questions. The product is meant to operate inside the messy reality of office life, where information is scattered across tools and much of the work is hard to measure. That is also why the trust question is so central: convenience may be appealing, but the product still asks users for access, patience and confidence that it will behave as expected.