OpenAI's Agentic Ambitions: ChatGPT Work and the Battle for White-Collar Adoption
OpenAI is making a sweeping bet that artificial intelligence agents will soon become as common in offices as spreadsheets and email. The company's newest product, ChatGPT Work, is designed to let white-collar professionals deploy AI that does not merely answer questions but autonomously completes multistep projects. Released last month at a $20 monthly subscription tier, the product is a modified version of OpenAI's Codex coding tool, repackaged for accountants, investors, doctors, marketers, and anyone else whose workday revolves around a computer.
The core promise is simple: give an LLM access to your digital life—your inbox, your Slack, your calendar, your documents, your project management software—and it can act on your behalf. That means drafting reports, building dashboards, scheduling meetings, analyzing financial data, and coordinating across apps. But the central question is not whether the technology can do these things. It is how much control users are willing to hand over to an AI system, and whether the experience is intuitive enough for mainstream adoption.
What ChatGPT Work Actually Does
ChatGPT Work is built on the same agentic foundation as Codex, which software developers already use to write, test, and deploy code. The difference is that Work aims to be general purpose. It can connect to external services such as email, Slack, Notion, Figma, Google Calendar, and a growing list of SaaS platforms. Once connected, it can read information, perform analysis, and take actions. For example, it can extract a strangely formatted school calendar from an email and insert the events into Google Calendar. It can assemble investment memos from scattered communications. It can turn a Slack conversation about an engineering problem into a set of charts. It can even create an auto-updating dashboard for public company metrics.
OpenAI positions the product as a step toward a world where artificial intelligence goes beyond answering questions to helping everyone turn their biggest ideas into reality. That marketing language masks a more complicated reality. The most valuable uses require deep access to private data. Giving a model the keys to your inbox, your source interviews, or your bank account is a significant trust decision. Even OpenAI's own engineers acknowledge the trade-off. One lead engineer said he gives the app access to his inbox, Slack, Notion, Figma, and phone because it is the only way to test the future. He acknowledged that the model might pull from a private direct message and share information it should not. He said he would take the personal hit if necessary, but so far he has not had to.
Adoption Inside and Outside OpenAI
The gap between internal enthusiasm and external uptake is stark. An OpenAI-backed study found that in June, 98% of OpenAI employees were using Codex. Among organizational subscribers, only 17% were using the agentic coding tool. Among individual subscribers, less than 1% were using it. The joint ChatGPT and Codex app has about 20 million users, compared with more than a billion people who use ChatGPT online. That difference between near-total internal adoption and negligible external adoption represents both the challenge and the opportunity.
OpenAI executives argue that as the value and utility of the product grow, users will be willing to pay for it. They compare it to the broader ChatGPT subscription model: once people experience the utility, they will say of course they want to pay $20 a month because the value is so much greater. But the company has not yet demonstrated that the same willingness to pay will apply to agentic workflows that consume far more compute and require more setup.
The Harness Problem: Making Agents Usable
Every large language model requires a harness—the software wrapped around the model that decides what information it sees, which tools it can use, and how it presents answers. For developers, a command-line interface was enough to change how software is built. But most people do not use command-line interfaces. There is a reason graphical operating systems replaced text-based ones. The same principle applies to AI agents.
An agentic product that goes beyond software engineering must operate in the messy world of everyday life and legacy tools. It must handle websites built in 1995 and never updated. It must navigate inconsistent permissions, confusing settings, and fragmented workflows. OpenAI's engineers say discoverability matters at this early stage. They have added buttons for selecting projects and plug-ins, but the goal is a magic-box interface similar to other OpenAI products. They compare it to skeuomorphism—the practice of making digital tools look like physical objects. That design was not just cringe, one engineer said. It helped people make the transition.
The experience is not always smooth. Setting up permissions for agents to access a cloud drive can be confusing and circular. In one test, a user tried multiple times to grant only read access and received error messages. A dialog box eventually explained that only complete access would work. Many important settings are available only on the web app, forcing users to switch between desktop and web. Some limitations are baffling: linking the app to Google Calendar lets it create events but not new calendars. And unless the effort level is set high, the agent can perform like the worst intern imaginable.
The Competitive Landscape and the Bitter Lesson
OpenAI is not alone in this race. Anthropic's Claude Code and Claude Cowork have defined much of the market for AI coding and conversational agents. Other vertical-specific competitors, such as Harvey for law and Clay for sales, have pursued customers with a model-agnostic approach. They plug in whichever AI works best at the time. Industry analysts have warned that if AI labs cannot rapidly acquire the complementary assets needed to scale in the market, value will accrue elsewhere.
The rivalry between OpenAI and Anthropic has shaped product design. When OpenAI first developed Codex as a web app, engineers bet on the model being smart enough to handle tasks entirely on its own with minimal user input. They were, in one engineer's words, a bit more AGI-pilled. Anthropic's Claude Code took a different approach: it oriented around back-and-forth conversation. If given a problem, it would survey possibilities, offer three or four options, and check back after each step. That approach proved more effective at the time, even though it demanded more work from users. OpenAI eventually followed suit, adding more opportunities for user interaction. That became the Codex known today, with desktop and mobile apps.
The debate over harnesses reflects a deeper principle in AI research known as the bitter lesson: a better general model is more important than specific domain expertise. For true believers, the harness is a temporary crutch, not a moat. One OpenAI engineer said that adding a whole bunch of extras—if-thens and tools—might get good results in the short term, but the next model will come out in a couple of months and make that obsolete. The goal of good harness engineering, he said, is to be more precise about what information the model really needs to solve the problem, because models are getting better at doing that if you simply let them do their thing.
Not everyone agrees that most users or models are ready for that. A Wharton School professor who studies AI tools in the workplace has written that ChatGPT tends to want to do magic and just do it for you, while Claude does comparisons and shows them, repeatedly asking for input and feedback. OpenAI's core product lead disagrees, arguing that the conversational nature of the app is better than learning how to use an application. He says the world seems ready, pointing to incredible adoption.
Cost, Lock-In, and the Road Ahead
Token economics are a critical part of the story. Agents that work for longer stretches burn through more tokens, making them more lucrative on a per-user basis. But they also cost more to run. In one casual test, a user on a $20-a-month subscription consumed more than 80 million tokens in four days. According to the model's analysis, that cost $65. The app has no dashboard to show this. That amounts to a subsidy of more than three times the subscription price for just four days of casual use.
OpenAI says it is working every day to push the frontier on efficiency. It recently announced an 80% price cut for users of its Luna model. If you wake up six months from now, an executive said, you should be able to do all the same tasks with less spend. Still, the question of lock-in remains. These apps create switching costs through data retention and the pain of configuring access to plug-ins and permissions. For non-technical workers, the cost may feel invisible at first, but at scale it will force harder conversations.
OpenAI's headquarters, described as wood-paneled and plant-filled, had a calm but slightly tense atmosphere when visited in July. Engineers monitored laptops during conversations and rushed between meeting rooms. The head of product engineering said the focus remains on the promise of the magic box, but there is still too much complexity. He said he is optimistic that OpenAI can solve it with the model and in a truly AI-native way. That optimism will be tested as the company tries to bring agentic AI to hundreds of millions of white-collar workers who are not software engineers and who may not be willing to hand over the keys to their digital lives without a fight.
Source:TechCrunch News