OpenAI is betting that artificial intelligence agents can transform how white-collar workers handle routine tasks, but a massive adoption gap reveals the real challenge: it's not about smarter AI, it's about making people trust it. While 98% of OpenAI employees use the company's agentic tools, less than 1% of individual subscribers have adopted them, according to an internal study cited by the company. This disconnect highlights a critical moment for the AI industry: as models become more capable, the real bottleneck isn't raw intelligence, but making these tools accessible and trustworthy enough for everyday professionals.

What Are AI Agents, and Why Does OpenAI Think They're the Future?

An AI agent is fundamentally different from a chatbot. Instead of answering a single question, an agent can access your digital tools, understand your workflows, and complete multi-step projects autonomously. Think of it as hiring a digital assistant who can read your email, check your Slack messages, pull data from spreadsheets, and synthesize it all into a report without you having to manually coordinate each step.

OpenAI released ChatGPT Work last month as part of this vision. Available on the company's lowest subscription tier for $20 per month, it's designed to give non-engineers the same agentic capabilities that software developers already enjoy through tools like Codex. The product hooks large language models, or LLMs (AI systems trained on vast amounts of text), directly into your workspace tools like email, Slack, Notion, and Figma.

"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, including Work.

Thibault Sottiaux, Product Lead at OpenAI

Why Is the Adoption Gap So Dramatic?

The numbers tell a striking story. An OpenAI-backed study found that in June, 98% of OpenAI employees were using Codex, the company's agentic coding tool. Yet just 17% of organizational subscribers and less than 1% of individual subscribers were using the agentic version. The joint ChatGPT Work and Codex app is used by only 20 million people, compared to more than a billion users prompting ChatGPT online.

Andrew Ambrosino, the lead engineer for OpenAI's desktop app, explained that the problem isn't the AI itself. It's the interface and trust. When OpenAI first tried to deploy Codex to non-engineering teams, the tool was "actively hostile to them," Ambrosino said, asking them about code and showing technical readouts meant for software developers. The company had to rebuild the entire experience from February onward to make it general-purpose.

"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," explained Andrew Ambrosino, Lead Engineer for OpenAI's Desktop App.

Andrew Ambrosino, Lead Engineer for OpenAI's Desktop App

How Are Companies Actually Using AI Agents Today?

Despite the low adoption numbers, early users have found compelling use cases. OpenAI employees are setting up weekly metrics reports and converting spreadsheets into planning tools. Venture capitalists are using agents to assemble communications and analysis about companies into investment memos. Operations teams are spinning up custom dashboards and data visualizations. Even Sam Altman, OpenAI's CEO, is using the tool to plan vacations.

One engineer described asking the program to examine a Slack conversation about an engineering problem and "make some charts," then receiving back a series of insightful plots. The common thread: all these tasks involve synthesizing information scattered across multiple tools and presenting it in a useful format.

What's Blocking Wider Adoption?

The core challenge is trust and usability. Ambrosino acknowledged the real 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 told TechCrunch. He accepts this risk for his own work, but most users aren't willing to take that personal hit.

Ambrosino

There's also a design problem. Most people don't use command-line interfaces; they use graphical interfaces like Windows. Similarly, most professionals aren't comfortable giving an AI system broad access to their digital life without clear, discoverable controls. OpenAI's team is building buttons and interfaces to make agent capabilities visible and manageable, even though some internal voices argue that users should simply ask the model directly.

How to Build Trust in AI Agents for Your Organization

• Start with Transparent Access Controls: Clearly define which systems the AI agent can access and modify. Limiting scope reduces the risk of unintended consequences, such as accidentally sharing confidential information or making unauthorized changes.
• Begin with Low-Risk, Routine Tasks: Deploy agents first on repetitive, data-intensive coordination work that doesn't require creative judgment, such as weekly reporting or data aggregation. These are the tasks agents handle most reliably today.
• Establish Clear Audit Trails: Ensure that all actions taken by the agent are logged and reviewable. This builds confidence and allows teams to catch errors before they cause problems.
• Invest in User Interface Design: Don't rely on command-line prompts or technical documentation. Provide buttons, menus, and visual controls that make agent capabilities discoverable and easy to understand for non-technical users.

"Discoverability matters in this phase, and at some point we won't have the button," noted Ambrosino, comparing the approach to skeuomorphism, the design practice of making digital tools resemble physical objects they replaced.

Andrew Ambrosino, Lead Engineer for OpenAI's Desktop App

Why This Matters for the AI Industry

OpenAI's struggle to scale agents beyond software engineering reflects a broader industry challenge. Coding has proven lucrative territory for AI labs, but it's still a tiny subset of professional work. If AI companies can't rapidly expand into accounting, law, sales, and other fields, they won't justify their massive investments in training and computation.

Vertical-specific competitors like Harvey, which focuses on legal work, and Clay, which targets sales teams, are already chasing these customers with a model-agnostic approach, meaning they'll integrate whichever AI model works best at the time. This creates pressure on OpenAI and other labs to prove that their tools can deliver value across industries, not just in code.

The real test for AI agents isn't whether the underlying models are intelligent enough. It's whether ordinary professionals will trust them with their workflows and whether the interfaces make that trust feel justified. Until OpenAI closes the adoption gap, that question remains unanswered.