How much control are you willing to give an AI model over your digital life? OpenAI is testing that question with ChatGPT Work, a product that connects large language models to email, messaging apps, calendars and a fast-growing list of software services used in modern offices. The promise is simple: instead of answering questions, an AI agent can take action across the tools you already use. The hard part is convincing people to hand over the keys.
ChatGPT Work launched last month on OpenAI's lowest subscription tier at $20 a month. It is built on Codex, a tool originally created for software developers, and is being repositioned for accountants, investors, doctors, marketers and other professionals who spend their days in browsers and productivity apps. The company describes the shift as moving from answering questions toward a world in which artificial intelligence helps everyone turn ideas into reality.
The commercial stakes are large. AI agents that run for long periods consume more tokens than simple chat sessions, making them more valuable to the company on a per-user basis. But OpenAI also needs to prove that agents can do more than write code. Coders have been the clearest early market for agentic AI, but they remain a small slice of the broader workforce. If the tools do not reach doctors, bankers, marketers and administrators, the enormous investment in training and computing may never be justified.
That helps explain why OpenAI, in private and public comments, has been talking less about model benchmarks and more about what it calls harness engineering. A harness is the software wrapped around a model that controls what information the model sees, what tools it can use and how it delivers results. For a yes/no chat interface, this is straightforward. For an agent that must open files, read spreadsheets, send messages and navigate websites, the harness determines whether the experience feels magical or maddening.
OpenAI's own workforce has become the first test bed. The lead engineer for the desktop app said he connects the product to his personal inbox, Slack, phone, Notion and Figma files. In an interview, he admitted there is some risk that the model might pull information from a private message without realizing it should not. He said he takes that risk as part of building the product. The company says many of its employees use the tool to automate routine reporting and even plan vacations, but the gap between internal use and outside adoption is stark.
An OpenAI-backed study found that in June, 98% of OpenAI employees used Codex. Among organizational subscribers, only 17% had used the agentic coding tool. Among individual subscribers, the figure was less than 1%. ChatGPT Work and Codex combined are used by about 20 million people, compared with more than a billion users who prompt ChatGPT online. The discrepancy shows how much work remains before agentic AI becomes a mainstream habit rather than an engineering niche.
OpenAI's product team describes the problem in terms of usability. The company's earliest agentic tools were designed for developers, with a command-line interface that felt actively hostile to nontechnical employees. Employees in communications and finance were asked to look at code diffs and error messages. So the team started making the tool more general purpose, adding buttons and conversation steps to help users understand what the model can do. One engineer argued that discoverability matters in this phase, and that a button can be removed later once users learn to ask the model directly.
Outside OpenAI, early adopters have found genuinely useful workflows. Venture investors have used agents to compile communications and analysis into investment memos. Operations teams have used them to build dashboards and visualizations. A reporter trying the tool successfully had it pull a preschool calendar from an inbox and add the events to Google Calendar, saving significant manual data entry. Another request produced an auto-updating dashboard of financial metrics from public documents. Those wins make the product's remaining friction more frustrating.
Permissions are a common point of confusion. In one test, the reporter tried to give the agent read-only access to a cloud drive and hit repeated error messages. Eventually a mobile app dialog box explained that full access was required. Important settings also appeared only on the web app, forcing users to move between interfaces. Some limitations are arbitrary: the tool can create calendar events when connected to Google Calendar, but it cannot create new calendars. And if the effort level is not set high enough, the output can feel like the best result from an unreliable intern.
Even OpenAI engineers admit
Source: TechCrunch News