How much control are you willing to give an AI model over your digital life? That is the central question as OpenAI pushes its agentic tools beyond software developers and into the broader white-collar workforce.
Getting the most value from a large language model means giving it access to your inbox, calendar, cloud drives, and business applications. For Andrew Ambrosino, the lead engineer for OpenAI's desktop app, that trade-off is necessary to test the future. His ChatGPT app can now read and control his Slack, email, phone, Notion, Figma, and more.
“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,” Ambrosino said. “I’ll do it for the job. I will take the personal hit here and there if I have to.”
That willingness is not universal. OpenAI’s biggest bet in this direction is ChatGPT Work, released last month and available on the company’s lowest subscription tier for $20 a month. The product is designed to let accountants, investors, doctors, and other white-collar professionals deploy AI agents that can complete multistep projects autonomously. Instead of merely answering questions, the model is meant to draft documents, build spreadsheets, analyze data, and coordinate across software platforms.
OpenAI's marketing describes a world where “intelligence goes beyond answering questions to helping everyone turn their biggest ideas into reality.” For software developers, that shift is already underway. Codex, OpenAI's coding agent, has changed how many engineers write code. The challenge is to bring the same capability to non-engineers.
ChatGPT Work is a modified version of Codex. It is designed to give non-engineers a version of the same functionality that software engineers already get: an AI tool that doesn't just answer questions but completes complex tasks on its own.
“In this new factor, ChatGPT can actually do entire, very complicated tasks for you all autonomously in a way that is delightful and safe,” said Thibault Sottiaux, who leads OpenAI's core product work. “It’s the very mission of OpenAI — to bring everyone along.”
Commercial stakes and adoption gaps
Commercially, agentic AI is important because longer tasks consume more tokens, making each user more valuable to the company. Reaching new professions is critical for OpenAI and the broader AI industry. Coding is a lucrative but small slice of professional work. If AI labs cannot expand to other knowledge work, their massive investments in training and computation may not pay off.
While the big labs have focused on software engineers, vertical-specific competitors like Harvey for law and Clay for sales have been chasing those customers with a model-agnostic approach. They plug in whichever AI works best at any given time. Industry analysts see this as a major challenge for OpenAI.
“If the labs cannot rapidly get ahold of the key complementary assets needed to scale AI in the market, value will accrue elsewhere,” wrote Christian Catalini on a16z's blog.
Making AI apps work for non-engineers requires more hand-holding. OpenAI's non-engineering workforce, such as communications and finance teams, initially used Codex when it was “actively hostile to them,” Ambrosino said. The interface asked them about code and showed empty diffs. Between February and now, the company made it more general purpose.
An OpenAI-backed study found that in June, 98% of OpenAI employees were using Codex, but only 17% of organizational subscribers and less than 1% of individual subscribers were using the agentic coding tool. That gap between near-universal adoption inside the company and negligible adoption outside it is both the challenge and the opportunity.
“The more value and the more utility that we generate for users, the more they will be willing to also pay for some part of that utility,” Sottiaux said. “You sit there and you’re like, ‘of course I want to pay $20 bucks a month for this,’ because the value that you get is so much more.”
Making agentic AI intuitive
To understand the adoption gap, it helps to understand what engineers are building. 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 an agent to do tasks, the harness gives the model tools and instructions for long-term projects.
For developers, a command-line interface was enough to transform software development. But most professionals are not using CLIs. Windows replaced DOS for a reason. An agentic product for non-engineers must work with the messy reality of everyday tools, including websites that were built decades ago and never updated.
Apps like Claude Code and Codex made “vibe coding” possible by abstracting away the actual software writing. Users could just describe what they wanted. Now OpenAI wants to bring similar functionality to general knowledge work. The company is looking at tools like OpenClaw, which coders use to put LLMs to work, and make them as easy as prompting.
“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 said.
There is an internal debate at OpenAI about whether buttons and interface elements are necessary. Some employees argue that users can simply ask the model. Ambrosino pushes back because “discoverability matters in this phase, and at some point we won’t have the button.”
ChatGPT Work has a few more buttons for selecting projects and plug-ins, but it still aims for the same “magic box” interface as other OpenAI products. Ambrosino compares this to skeuomorphism, when digital tools were designed to look like their physical counterparts. “That stuff wasn’t just cringe design. That actually helped get people into this and make the transition.”
OpenAI would not say how many people use Work versus Codex. The joint app is used by just 20 million people, compared to more than a billion users the company says are prompting ChatGPT online.
Giving ChatGPT a license to skill
For now, OpenAI is pitching ChatGPT Work for routine, data-intensive coordination tasks. Employees are setting up weekly metrics reports, turning spreadsheets into planning tools, and building dashboards. VCs are using agents to assemble relevant communications and analysis into investment memos. Ops teams are creating bespoke data visualizations. Sam Altman is using the tool to plan his vacations. One OpenAI engineer described asking the program to look at a Slack conversation about an engineering problem and “make some charts,” then receiving a series of insightful plots.
“There is a deluge of information for the average worker or employee of any of these companies, including myself,” said Akshay Nathan, who leads the product engineering team at OpenAI. “We’re actually quite limited by our ability to parse everything that’s available to us, and then take action on it. That information lives in all these system records tools [like Salesforce]… the value of ChatGPT is you already have access to this, but now you truly have access to it.”
This could become the digital personal assistant that AI evangelists have dreamed about. Like Claude Cowork or Perplexity's browsing agent, ChatGPT Work links agents to your existing workspace — email, browser, SaaS platforms — and puts that context to work.
When it works, the system can be impressive. One user asked ChatGPT Work to take a strangely formatted preschool calendar from email and put it into Google Calendar, saving repetitive data entry. Another asked it to do financial analysis on publicly traded companies and received an auto-updating dashboard. It also built a queryable database of space launches, a task that previously required writing Python scripts.
But setting up permissions is not simple. Granting access to a cloud drive can be confusing and circular. The model may not help much, and some settings are only available on the web app. Users often need to work in both the desktop and mobile apps at the same time. Some limitations are baffling: linked to Google Calendar, ChatGPT Work can create events but not new calendars. And unless the effort level is set to high, the output can feel like the worst intern you’ve ever worked with.
Joe Gershenson, engineering lead for OpenAI's harness, acknowledged that effort settings are not intuitive for new users. “There are things that we can do better to help them get the right level of reasoning,” he said, adding, “Watch this space.”
Another challenge is that most knowledge work is not as measurable as code. A software program either works or it doesn't. But a good presentation, business strategy, or sales pitch is harder to evaluate. OpenAI uses its benchmark GDPval, drawn from 44 occupations and hundreds of knowledge work tests, plus user feedback. Less officially, the product direction often comes from OpenAI employees themselves. As Ambrosino said, “We have to always parse out … are we doing the workflow that everybody else will be doing, or are we weird?”
The rivalry that drove product design
OpenAI's engineers were reluctant to say what separates ChatGPT Work from Claude Cowork or other competing harnesses. Gershenson said he does not look at what Anthropic is building. “The Mad Men ‘I don’t think about you at all’ meme comes to mind here.”
Still, the similarities between the user interfaces are striking. ChatGPT Work even prompts new users to port over their Claude Cowork data. Claude Code defined the market for AI coding and launched a revolution in software engineering. That is a sensitive topic at OpenAI because the company had the idea first but did not harness it correctly.
When OpenAI first developed Codex as a web app, the engineers bet on the model being smart enough to handle a task on its own with minimal user input. Ambrosino describes it as “a bit more AGI-pilled.” Anthropic's Claude Code, built shortly afterward, was oriented around back-and-forth conversation. If you gave it a problem, it would survey possibilities and offer three or four options. After you chose, it would go further and then check again. This approach left less room for mistakes.
Anthropic's approach proved more effective. “[Our] product was a little ahead of where the model and harness was at the time,” Ambrosino said. OpenAI eventually added more interaction points, creating the Codex used today. Download statistics suggest Claude Code was more in demand until April of this year, when Codex took a slight lead. Surveys of enterprise use also suggest OpenAI is catching up.
Part of that lead comes from getting product-market fit right. Part comes from complaints about safety restrictions on Anthropic's models and compute shortages. OpenAI's engineers insist that the key differentiator is the strength of the underlying model. “The frustrating answer is that a lot of times it is the model,” Ambrosino said.
What makes a good harness?
That explanation returns to the “bitter lesson” in AI research: a better general model is more important than specific domain experience. For true believers, the harness is a temporary crutch, not a competitive moat.
“You could get good results in the short term by adding a whole bunch of extras — if and thens and tools — but like, come on, the next model is going to come out in a couple of months and make that obsolete,” Gershenson said. His team focuses on exposing the model to the necessary tools and context in the simplest way possible.
“The goal of good harness engineering is to be more precise about what information the model really needs to solve your problem, because the models are getting better and better at doing that if you simply let them do their thing,” he said.
Still, it is not clear that a model-specific harness is the right approach. Comparisons run by companies like Composio and Databricks show that different harness and model combinations deliver different performance on coding benchmarks. Databricks found that Pi, an open-source harness published by the software company Earendil, outperformed Codex while using the same GPT 5.5 model. Pi has been used to build projects like OpenClaw and Cloudflare OS.
Pi's creator, Mario Zechner, says the minimalist harness works because it can modify itself and build its own interfaces. But he sympathizes with the challenge OpenAI faces in expanding beyond engineers. “Everything is coding agent shaped… the reason is that they only have training data for coding agent tasks,” he said. “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.”
He sees big labs pushing their harnesses as a way to lock in users. “They need to own the entire stack; otherwise, they just become a model provider and then need to compete with Chinese models.”
There are also cost concerns. One reporter using ChatGPT Work on a $20-per-month subscription used more than 80 million tokens in four days, which cost $65, according to the model's own analysis. That is a subsidy of more than three times the subscription price for casual use alone. Sottiaux said efficiency is a priority, pointing to an 80% price cut for users of OpenAI's Luna model. “If you wake up six months from now, you should be able to do all of the same [tasks] with less spend.”
Other relevant questions include lock-in through data retention and the painful process of configuring access to plug-ins and permissions. Inside OpenAI's headquarters, the atmosphere is calm but tense. Engineers constantly monitor their laptops and rush between meetings. Nathan, the product engineering lead, says the focus remains on “the promise of the magic box, but I still think there’s too much complexity… I’m very optimistic that we can solve it, with the model and in a truly AI-native way.”
The future of agentic AI depends on whether companies like OpenAI can make their tools trustworthy, intuitive, and cost-effective enough for the millions of workers who have never written a line of code.
Source:TechCrunch News