How much control are you willing to give an AI model over your digital life? That question sits at the center of OpenAI's latest product push. The company is now building ChatGPT Work, a $20-per-month service that connects its language models to the tools people already use at work: email, calendars, Slack, Notion, Figma, customer records, spreadsheets, and more. Instead of only answering questions, the software is designed to complete multistep projects by operating directly inside these products.
For a control freak or someone hesitant about artificial intelligence, granting that kind of access may feel dangerous. For an engineer leading the desktop app, it is the only way to test where the technology is heading. That engineer has given the app control over his inbox, Slack account, phone, and several productivity platforms. He acknowledges that problems can happen. If the AI is asked to write a document, it might pull from a private direct message and not realize that some information should stay private. He is willing to accept those risks. “I'll do it for the job. I will take the personal hit here and there if I have to,” he says.
The move from coding assistant to workplace agent
ChatGPT Work is a modified version of Codex, OpenAI's coding agent. Codex turned language models into tools that could write code, run commands, and build software with minimal human guidance. That was valuable for software engineers, but OpenAI wants to reach a much larger audience: accountants, doctors, investment analysts, operations teams, marketers, and other white-collar professionals whose work is also done on a computer.
The underlying idea is that an AI agent should not merely answer questions. It should perform whole workflows autonomously. A product leader at OpenAI describes it as the company's mission to bring everyone along. Commercially, that ambition matters. Agents that work for longer stretches consume more tokens, making them more lucrative per user. Reaching new professions is important not just for OpenAI, but for the entire AI industry. If coding has become a profitable niche, it is still a small fraction of the professional tasks needed to justify enormous investments in training and computing.
Some vertical-specific competitors are already chasing those customers. A legal AI startup specializes in work for lawyers, and a sales-focused startup has built agents for revenue teams. These companies are model-agnostic, meaning they can connect their services to whichever AI model performs best at any moment. Industry analysts warn that if large AI labs cannot quickly control the complementary assets needed to scale AI in the market, much of the value will move elsewhere.
The adoption gap inside OpenAI
One surprising challenge is adoption. An OpenAI-backed study found that in June, 98 percent of OpenAI employees used Codex, but only 17 percent of organizational subscribers and less than one percent of individual subscribers were using the agentic coding tool. That gap between total adoption inside the company and negligible adoption outside it illustrates the difficulty of moving from a technical product to a mainstream one.
The company sees a clear opportunity. The more value users get from an agent, the more they may be willing to pay. Executives argue that once people sit down with ChatGPT and see what it can do, paying $20 a month feels easy. But many potential users still struggle to understand what the tool can do and how to configure it safely.
Building software that regular people can use
Every large language model needs what engineers call a harness. A harness is the software wrapped around the model. It decides what information the model sees, which tools the model can use, and how the model presents its answers. For a chatbot, that may simply be a web interface. For an agent, the harness also gives the model instructions, permissions, and the ability to take actions over hours or days.
Developers are comfortable with command-line interfaces. Coders accepted tools that required typing commands and reading technical output. But most workers are not using command lines. There is a reason graphical user interfaces replaced text-based operating systems decades ago. To make agents useful beyond software engineering, OpenAI has to build a harness that can handle “the messy world” of everyday tools and websites, including aging systems built long before modern AI existed.
The first coding agents changed software development through what many call “vibe coding.” Developers could describe a program in plain language and let the AI generate most of the code. Now OpenAI wants to give a similar experience to non-engineers. A user should be able to ask for a report, a dashboard, or a plan, and let the agent carry out the steps.
Product designers at OpenAI say these experiences are vital. Without a friendly product in front of the model, only experts would know how to get the same results. The company has internal debates about simplicity. Some engineers argue that a button is unnecessary when users can simply tell the model what to do. Others say discoverability matters at this early stage. They compare the current transition to skeuomorphism, the design practice of making digital tools look like their physical counterparts. An early calculator app looked like a pocket calculator, and that familiarity helped new users make the switch.
What people are actually doing with agents
OpenAI says ChatGPT Work and the desktop Codex app are used together by about 20 million people. That is a large number, but still small compared with more than one billion users who prompt ChatGPT online. The company declines to break down exactly how many people use the workplace product versus the coding tool.
Early users have found a range of tasks. Some employees use the agent to assemble weekly metrics reports. Others turn spreadsheets into planning tools. Venture capitalists have used agents to gather communications and research about companies into investment memos. Operations teams have created custom dashboards and data visualizations. Even the top OpenAI executive has reportedly used it to plan vacations. An engineer described asking the agent to review a Slack conversation about a technical problem and produce charts, then received a set of helpful graphs.
The value proposition is linked to information overload. Many professional workers have access to enormous amounts of data scattered across email, messaging apps, sales systems, and internal databases. Humans are limited in how much they can read, remember, and act on. An agent that is connected to those systems can surface what matters and turn raw context into usable work.
The hands-on experience remains uneven
A test of the product shows how powerful it can be. One journalist asked ChatGPT Work to pull a child's preschool calendar from an email inbox and transfer the events into Google Calendar. The agent handled the formatting and entry without much trouble. It also produced an auto-updating dashboard of financial metrics and built a searchable database of space launches from public information. These tasks would previously have required either manual data entry or Python scripting.
Yet the product has serious rough edges. Setting up permissions to give the agent access to a cloud drive can be confusing and circular. Trying to grant read-only access occasionally generates error messages, and the model may not explain what went wrong. Some settings are available only in the web app, forcing users to switch between desktop, mobile, and web. The agent can create calendar events but cannot create new calendars. Users are told to select an “effort” level when starting a task, but that setting is not intuitive. Early adopters recommend choosing high effort for anything important. At lower effort, the agent can behave like a poor intern.
An OpenAI engineering lead admitted that the effort settings are not yet clear for new users and said the team is working on better guidance. This kind of friction matters because frustrated users may give up before experiencing the real value of an agent.
Code has clear answers. Work often does not
Another major challenge is evaluation. In software engineering, code either runs or it does not. There are benchmarks, tests, and clear ways to measure whether an AI agent completed a coding task correctly. For a marketing pitch, a business strategy, or a sales presentation, the standards are much messier. It is harder to verify whether an agent's judgment was good, whether it missed important context, or whether a subtle mistake could lead to embarrassment or financial loss.
OpenAI says it uses a benchmark called GDPval, based on 44 occupations and hundreds of knowledge-work tests, to measure how well agents perform. The company also relies on user feedback. But with a product that can “really do anything,” engineers admit they are still learning which workflows matter most to ordinary workers. OpenAI employees often test the product on their own jobs, raising the question of whether their habits will match those of teachers, healthcare workers, and retail managers.
The rivalry that shaped ChatGPT Work
The competitive story around OpenAI's agent tools cannot be separated from Anthropic, the maker of Claude. Anthropic's coding agent, Claude Code, defined the market for AI coding assistants. It became especially popular after OpenAI's first iteration of Codex failed to gain the same traction. The reason was not just the model. Anthropic's product was built around continuous back-and-forth conversation. When given a problem, Claude Code would explore options and present several paths. After the user chose one, it would do a bit of work, then check in again. This reduced the chance of large mistakes and made the agent feel more controllable.
OpenAI's first Codex product was more autonomous in a different way. The company had bet heavily on the model's ability to solve tasks independently. That bet proved premature. One OpenAI engineer later said the product was ahead of where the model and harness were at that time. OpenAI eventually adopted a more interactive approach, and Codex gained ground. By April, download data suggested Codex had taken a slight lead over Claude Code in popularity. Enterprise surveys also suggested OpenAI was catching up.
Part of the shift came from user complaints about Anthropic's safety restrictions and occasional compute shortages. But OpenAI engineers point mainly to model quality. They argue that the latest OpenAI model is both more capable and more cost-effective, and that any harness around it benefits from raw improvements in reasoning, tool use, and instruction following.
The role of the harness
There is an important debate about whether the model or the harness matters more. Some open-source projects have shown that a cleverly designed harness can outperform OpenAI's own tool while using the same model. One independent comparison found that an open-source harness called Pi beat Codex on coding benchmarks while relying on the same underlying GPT model. Pi is intentionally minimal, but it can modify itself and build its own interfaces. Its creator argues that this proves an autonomous approach can still work, at least for coding tasks.
The same engineer also points to a deeper limitation. Most agent training data comes from coding tasks because coding leaves traceable records: commands, diffs, reviews, and finished code. Management decisions, by contrast, can take months to produce an outcome. There is no simple user-and-agent trace that captures the quality of the decision. Many real-world tasks are not digitized well enough for a model to learn from them.
This explains why vertical agents and productivity tools remain such a contested space. If model performance keeps improving, the best harness might eventually be an extremely simple interface. But before that happens, companies may try to lock users into their own stacks through permissions, stored workflows, and integrations. Open-source developers warn that OpenAI needs to own the entire stack. Otherwise, it risks becoming just another model provider competing with low-cost Chinese models on price.
The hidden cost of AI agents
Cost is another concern. In one informal test with a $20 monthly subscription, a journalist used more than 80 million tokens in four days. According to the model's own analysis, that usage would have cost about $65 under normal pricing. This means OpenAI was subsidizing more than three times the subscription price for a few days of experimentation. The company says it is working on efficiency and recently announced an 80 percent price cut for users of a newer model family. Over time, agents should become cheaper to run, allowing similar tasks to be completed with less spending.
Privacy and lock-in also loom over the product. Users who grant the agent access to inboxes, documents, calendars, and business systems may find it painful to switch providers later. Reconfiguring permissions across dozens of integrations is tedious, and the agent accumulates detailed knowledge about a user's habits and priorities. That creates a strong commercial advantage for OpenAI, but it also raises questions about data retention and user autonomy.
From magic box to everyday assistant
OpenAI's headquarters has an atmosphere of focused intensity. Engineers overseeing the agent product monitor live dashboards and often move directly from one meeting to another. The head of product engineering says there is still too much complexity in the experience, but he is optimistic that the model and a “truly AI-native” interface can solve that. He still believes in the promise of the magic box: a place where a user can describe a goal and watch an AI handle the rest.
The question remains how many people are ready to hand over the keys to their digital lives. Some will hesitate, and for good reason. Privacy, accuracy, and cost are all unresolved. Yet OpenAI is betting that the convenience of a tireless agent will eventually outweigh those fears. The company's engineers are willing to accept personal data exposure now to build that future. What is missing is evidence that millions of non-technical workers will make the same leap.
Source:TechCrunch News