OpenAI is building AI agents for everything. Will everyone use them?

How much of your digital autonomy are you prepared to surrender to a Large Language Model (LLM)? For many, the prospect of granting an AI the "keys to the kingdom"—access to private communications, financial records, and professional workflows—is a daunting proposition. Yet, for Andrew Ambrosino, lead engineer for OpenAI’s desktop application, this level of integration is the only viable path toward testing the future of human-computer interaction.

Currently, Ambrosino’s own desktop environment is fully tethered to OpenAI’s agentic systems. His inbox, Slack workspace, mobile device, and productivity suites like Notion and Figma are all under the purview of an AI agent.

“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 admits. “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.”

This philosophy underpins ChatGPT Work, a product launched last month as part of OpenAI’s $20-per-month subscription tier. The goal is ambitious: to transition AI from a passive chatbot into an active agent capable of executing complex, multi-step workflows for accountants, investors, medical professionals, and the broader white-collar workforce.

The Shift from Chatbot to Agent

OpenAI’s marketing vision is clear: they want to move beyond simple question-and-answer interactions toward a reality where AI helps individuals transform abstract ideas into tangible outcomes. While this "agentic" shift has already revolutionized software engineering, its migration into other professional sectors has been sluggish.

ChatGPT Work is essentially a refined iteration of the company’s Codex coding tool. By adapting the logic used by software developers—who have long relied on AI to manage complex, multi-stage projects—OpenAI hopes to provide non-technical users with the same autonomous capabilities.

“In this new factor, ChatGPT can actually do entire, very complicated tasks for you all autonomously in a way that is delightful and safe,” says Thibault Sottiaux, who oversees OpenAI’s core product development. “It’s the very mission of OpenAI—to bring everyone along.”

From a commercial standpoint, the stakes are high. Agentic tasks, which require longer processing times and more complex reasoning, consume significantly more tokens than simple queries. This makes them a more lucrative revenue stream for OpenAI. Furthermore, if the company is to justify the massive capital expenditure required for model training and compute, it must expand its utility far beyond the niche of software development.

The Adoption Gap

The challenge lies in the "hand-holding" required for non-engineers. Early internal testing at OpenAI revealed that when the company’s finance and communications teams were first introduced to Codex, the experience was jarring. They were met with technical jargon and code-centric outputs that felt, in the words of Ambrosino, “actively hostile.”

Data from the study “The Shift to Agentic AI: Evidence from Codex” highlights the disparity:

• 98% of OpenAI employees were utilizing the agentic coding tool by June.
• 17% of organizational subscribers were using the tool.
• Less than 1% of individual subscribers had adopted it.

This gap between internal ubiquity and external hesitation is the primary hurdle OpenAI must clear. Sottiaux remains optimistic, noting that as the utility of these agents grows, the $20 monthly fee will increasingly be viewed as a bargain.

Designing for the "Messy World"

To bridge the gap between power users and the general public, OpenAI is focusing on the "harness"—the software architecture that surrounds an LLM, dictating which tools it can access and how it interprets user intent.

For developers, a command-line interface (CLI) was sufficient. However, the average user requires a more intuitive experience. As Ambrosino notes, the goal is to build an interface that can navigate the "messy world" of modern digital life, including legacy websites and fragmented software tools that have not been updated in decades.

OpenAI is currently debating the necessity of "buttons" and UI elements. While some purists argue that users should simply prompt the model for everything, the engineering team believes that in this early phase of adoption, discoverability is paramount. They are leaning into skeuomorphism—designing digital interfaces that mirror familiar physical tools—to lower the barrier to entry for mainstream users.

Giving ChatGPT a License to Skill

Currently, ChatGPT Work is being positioned as a powerhouse for data-intensive coordination. Use cases are already emerging:

• Operations teams are automating the creation of bespoke dashboards and data visualizations.
• Venture Capitalists are utilizing agents to synthesize communications and research into investment memos.
• Personal productivity is seeing a boost, with users automating calendar management and email summaries.

During testing, the system demonstrated the ability to extract complex, poorly formatted data from emails and integrate it into a Google Calendar—a task that previously required tedious manual entry. However, the experience is not without friction. Permissions management remains a pain point, with users often forced to navigate between web and mobile apps to grant the correct level of access.

The Competitive Landscape

OpenAI faces stiff competition from vertical-specific players like Harvey (legal) and Clay (sales), which are model-agnostic and prioritize deep integration into specific industry workflows.

Furthermore, the shadow of Anthropic looms large. While OpenAI engineers publicly downplay the influence of competitors, the market reality suggests otherwise. Anthropic’s Claude Code set the standard for user-centric agentic design by emphasizing a "back-and-forth" collaboration model.

Initially, OpenAI’s approach was more "AGI-pilled," betting that the model could handle tasks entirely on its own. Anthropic’s more cautious, iterative approach proved more effective, forcing OpenAI to pivot toward a more collaborative, human-in-the-loop design.

The Future of the Harness

Is a model-specific harness the right long-term bet? Some industry analysts and open-source developers argue that the "bitter lesson" of AI—that general model performance eventually eclipses specific domain engineering—suggests that complex harnesses may be temporary crutches.

Mario Zechner, creator of the open-source harness Pi, argues that minimalist, self-modifying architectures are the future. He warns that by locking users into proprietary harnesses, labs like OpenAI risk creating a "walled garden" that prioritizes data retention over user flexibility.

As OpenAI continues to iterate, the focus remains on the "magic box" experience. Despite the current complexity and the occasional "worst intern" performance, the team is confident that the combination of increasingly capable models and more intuitive UI will eventually make AI agents an indispensable part of the professional landscape.

For now, the experiment continues. Whether the average worker is ready to hand over the keys to their digital life remains the ultimate, unanswered question.