How much control are you willing to hand over to an artificial intelligence? For OpenAI's latest product, ChatGPT Work, that is the central question. Released last month and available on the company's cheapest subscription tier at $20 per month, the service aims to turn large language models into autonomous helpers for accountants, doctors, investors, and anyone else whose job lives inside a browser or a desktop app.
The pitch is straightforward: instead of merely answering questions, these AI agents can complete entire projects—draft memos, build dashboards, organize calendars, and even send emails—by connecting to the tools you already use. "The world seems to be ready," says Thibault Sottiaux, who leads core product work at OpenAI. "In this new phase, ChatGPT can actually do very complicated tasks autonomously in a way that is delightful and safe."
For the engineers building it, this means granting the AI deep access to their digital lives. Andrew Ambrosino, lead engineer for OpenAI's desktop app, has connected the system to his inbox, Slack, phone, and productivity platforms like Notion and Figma. He acknowledges the risk: "If I'm asking it to write a document, is there a possibility that it's going to pull from a private DM on that subject and not know that it's not supposed to share some info? Yes," he says. "I'll do it for the job. I will take the personal hit here and there if I have to. And I haven't had to."
The product is a modified version of Codex, OpenAI's popular coding assistant. Codex proved that AI agents could transform software engineering by reading codebases, editing files, and running commands with minimal supervision. ChatGPT Work extends that idea to office work, allowing non-programmers to delegate tasks like data analysis, report generation, and meeting preparation.
A wide gap between early adopters and the mainstream
Despite the enthusiasm inside OpenAI, public adoption remains modest. An OpenAI-backed study, based on usage data from June, found that 98% of OpenAI employees used Codex regularly. But only 17% of organizational subscribers and less than 1% of individual subscribers had touched the agentic coding tool. The company says its combined ChatGPT desktop app is used by about 20 million people, out of more than one billion monthly ChatGPT users worldwide.
That chasm is both a huge opportunity and a serious warning. If most people cannot or will not use AI agents, the astronomical costs of training frontier models may never be recovered. "If the labs cannot rapidly get ahold of the key complementary assets needed to scale AI in the market, value will accrue elsewhere," notes Christian Catalini, a technology analyst known for his work with venture firm a16z.
Numerous startups are already circling those valuable niches. Harvey builds AI workflows for law firms, while Clay specializes in sales-focused automation. These model-agnostic companies can plug in whatever AI works best at the moment, whether it comes from OpenAI, Anthropic, or Chinese developers.
From command line to conversation
To make AI agents intuitive for office workers, OpenAI has had to rethink how the technology is delivered. Software developers were content to use a command-line interface (CLI) to interact with Codex, typing text prompts and watching files change. But most people don't speak that language. "There is a reason Windows replaced DOS," Ambrosino explains. An agentic product for the general workforce must be more than a screen with blinking text.
The engineers describe a "harness"—the software that wraps around a language model, giving it access to tools, remembering context, and shaping its outputs. For years, harnesses were designed for programmers. Now OpenAI is trying to make a harness that anyone can drive. That means adding visual buttons for common actions, clicking through permission prompts, and smoothing over the messy reality of third-party apps that haven't been updated in years.
"Without these products in front of the model, experts would know how to get the same results, but you wouldn't get to a billion people using the thing," Ambrosino says. He compares the current strategy to skeuomorphism, when early digital tools mimicked physical objects to help people transition—like a calculator app that looked like a pocket calculator. "That stuff wasn't just cringe design. That actually helped get people into this and make the transition."
Impressive demos, but real limits
When ChatGPT Work functions well, the results feel almost magical. Reporters testing the service have asked the AI to extract children's school calendars from email and insert them into Google Calendar, produce interactive dashboards of financial metrics, or build searchable databases from historical records. The system can also send weekly summaries of new research from academic repositories.
Yet the experience has rough edges. Permissions are confusing and sometimes circular. One repeated problem: trying to grant the AI "read-only" access to a cloud drive would fail, until a dialog box explained the task requires full access. Several settings exist only in the web app, forcing users to bounce between desktop and mobile. And the AI's effort level—how much time it spends on a task—is hard to tune, leaving occasional users with shallow results. Joe Gershenson, an engineering lead for OpenAI's harness, admits the effort setting isn't clear yet. "There are things that we can do better to help them get the right level of reasoning," he says.
Trust is another factor. OpenAI has been reluctant to explain how it handles sensitive data from personal inboxes and corporate tools, and many early users still feel uncomfortable granting an AI full access to their communications. A Wharton School professor who studies AI adoption found that ChatGPT tends to "want to do magic and just do it for you," while competitors like Claude Cowork show their reasoning more clearly and ask for feedback along the way. That can make OpenAI's agents feel like a wild card, especially for new users.
Why coding is easier than white-collar work
One of the biggest obstacles is measurability. Software code either runs or it doesn't; tests pass or fail. A business presentation, a strategic memo, or a sales pitch has no similarly crisp scoreboard. "One of the unique challenges with a product like this is that it can really do anything," Ambrosino says. His team has struggled to define what problems to target, because problems differ across every profession.
OpenAI researchers have created a benchmark called GDPval, drawn from 44 occupations and hundreds of knowledge-work tests. But they rely heavily on feedback from internal users—the 98% of employees who already trust the AI. "We have to always parse out... are we doing the workflow that everybody else will be doing, or are we weird?" Ambrosino asks.
The battle with Anthropic
OpenAI's product design owes a lot to its rival Anthropic. Claude Code, released to wide acclaim, developed a conversational loop that checks back with the user at each stage, offering options and correcting course based on feedback. OpenAI's initial Codex product was more autopilot-oriented, an approach that proved too aggressive for early models. "Our product was a little ahead of where the model and harness was at the time," Ambrosino says now.
Anthropic's method became the industry template. OpenAI eventually incorporated more checkpoints, and now ChatGPT Work resembles that same style of collaboration. Still, engineers at OpenAI are loath to give credit to their competitor. "The 'Mad Men' meme comes to mind," Gershenson says when asked what he thinks of Claude Cowork. "I really don't look at the harnesses that they're building."
Despite the bravado, the similarities are striking. ChatGPT Work even asks users at startup whether they want to import settings and data from Claude Cowork. And an analysis of download statistics suggests Codex only overtook Claude Code in popularity in April of this year, after months of playing catch-up.
The open-source wildcard
Adding to the competition, an open-source harness called Pi has shown impressive results. Built by a software company called Earendil, Pi can outperform Codex even when paired with the same underlying GPT-5.5 model. Databricks and other firms have independently verified that harness choice matters as much as model choice. Pi's creator, Mario Zechner, argues that his minimalist approach works better for software engineering because it relies on the model's own ability to adapt and build its own interfaces.
Zechner also sees OpenAI's aggressive push into harnesses as a way to lock users into its stack. "They need to own the entire stack," he says. "Otherwise, they just become a model provider and then need to compete with Chinese models." But Zechner is skeptical that AI agents will spread quickly beyond coding. "Everything is coding agent shaped... the reason is that they only have training data for coding agent tasks," he notes. "Say I'm in management, I make a decision today, and the outcome happens months later. You cannot capture that in a simple trace of a user and agent back and forth."
Cost is another looming concern for OpenAI's business model. AI agents use far more tokens than simple chat sessions. In a casual four-day test, one reporter consumed more than 80 million tokens, at a cost of $65—just on a $20 subscription. That means every heavy user is being heavily subsidized. "We are working every day to push the frontier on efficiency," Sottiaux says, pointing to recent price cuts for some of OpenAI's most powerful models.
Still a work in progress
Inside OpenAI's headquarters, the mood is a mix of confidence and nervous energy. The engineers constantly check their laptops during interviews, rushing between meetings to fix bugs and review user feedback. Akshay Nathan, who leads product engineering, believes the answer is not just better software but better models. "I still think there's too much complexity," he says. "But I'm very optimistic that we can solve it, with the model and in a truly AI-native way."
The question remains whether non-technical users will ever trust AI agents enough to give them the keys. The technology is rapidly improving, costs are falling, and every major AI company is pouring resources into making these agents feel safer and easier to control. For the average office worker, the trade-off between convenience and privacy will only deepen. OpenAI's engineers have made their choice. The rest of the world is still deciding.
Source: TechCrunch News