OpenAI's latest research reveals a fundamental shift in how artificial intelligence is being used at work: instead of asking AI for advice, employees are delegating entire tasks to AI systems that can execute multi-step workflows independently. A new study on Codex, OpenAI's agentic AI platform, shows that the tool has evolved far beyond its original purpose as a coding assistant, becoming the primary work tool across every department at OpenAI, including Legal, Finance, and Recruiting.

What Is Agentic AI, and How Does It Differ From ChatGPT?

The distinction between conversational AI and agentic AI is central to understanding OpenAI's research. ChatGPT operates primarily as a conversational assistant, answering questions and providing information in response to user prompts. Codex, by contrast, is designed as an agentic system, meaning it can inspect files, execute commands, and create or modify artifacts on behalf of the user.

In practical terms, this means users are not asking Codex for advice; they are asking it to do work. The paper describes this shift as moving "from single interactions to delegated, long-horizon tasks." A lawyer might ask Codex to draft a contract and incorporate feedback across multiple revisions. A financial analyst might delegate data processing and report generation. An engineer might ask the system to debug code, refactor it, and validate the changes, all without stopping to ask for permission at each step.

How Quickly Are Non-Developers Adopting Agentic AI?

The most surprising finding in OpenAI's research is the speed at which non-developer adoption is accelerating. Since August 2025, non-developer users of Codex have grown 137 times among individual users, 189 times among organizational users, and 12 times within OpenAI itself. This rapid expansion suggests that agentic AI is not a niche tool for software engineers but a broad workplace technology.

Inside OpenAI, the shift is even more dramatic. As of June 11, 2026, Codex accounted for 99.8 percent of output tokens generated by OpenAI workers, compared to just 63.3 percent among organizational customers and 16.5 percent among individual users. The company noted that Legal, Finance, and Recruiting departments crossed into majority Codex use around April 2026, after engineering had moved first. The average lawyer or recruiter at OpenAI now generates more than 85 percent of their AI work tokens on Codex.

What Types of Work Are Being Delegated to Codex?

While Codex began as a coding tool, the research shows it is now used for a much broader range of tasks. OpenAI's internal data breaks down Codex usage by department and work category, revealing substantial usage across multiple domains:

• Software Development: Debugging, refactoring, validating code changes, and configuring applications remain core use cases.
• Documentation and Knowledge Work: Drafting documents, creating knowledge artifacts, and organizing information across departments.
• Data Analysis and Financial Work: Finance and Business Operations teams use Codex for data processing, analysis, and financial modeling.
• Coordination and Communication: Product, Marketing, and Operations teams delegate communication tasks and workflow coordination to the system.

The paper emphasizes that these activities are better understood as production than as consultation. Users are delegating concrete work, not seeking information or advice.

How Complex Are the Tasks Being Delegated?

OpenAI's research also examined task complexity and duration. By May 2026, 80.6 percent of sampled individual users had made at least one Codex request estimated to exceed 30 minutes of human work. More significantly, 70.2 percent had made requests estimated to exceed one hour, and 25.6 percent had made at least one request estimated to exceed eight hours.

Heavy Codex users operate the system differently from occasional users. Intensive users are more likely to run longer and more complex tasks, operate multiple agents concurrently, and use advanced features. Among the most intensive OpenAI users, Codex functions "less an assistant answering requests and more like a workflow system". This suggests that as organizations mature in their use of agentic AI, the tool transitions from a helper to a core part of the production workflow.

What Are the Limitations of This Research?

OpenAI is transparent about the limitations of its findings. The company notes that OpenAI itself is an unusually favorable environment for agentic AI adoption. Workers are familiar with frontier AI models, usage is inexpensive at the margin, organizational buy-in is high, and many workflows are close to the systems being developed. For these reasons, OpenAI's internal usage patterns are not representative of typical enterprise adoption.

Additionally, the research documents adoption patterns and usage metrics but does not directly measure productivity gains, software quality improvements, or business outcomes. Output tokens can indicate AI-mediated work, but they are not the same as completed work or higher-quality work.

What Does This Mean for the Future of Work?

OpenAI's research points to a broader workplace transformation in which agentic systems are used to delegate technical and knowledge-work tasks across departments. The shift from conversational AI to agentic AI represents a change in how organizations think about AI's role: not as a tool for getting information, but as a system for executing work. As more organizations adopt similar tools, the unit of analysis for AI productivity may shift from individual conversations to delegated workflows and completed tasks.

The research also raises questions about how different organizations will adopt agentic AI and at what pace. While OpenAI's workforce has embraced Codex at scale, broader enterprise adoption remains uneven, with organizational users at 63.3 percent Codex usage and individual users at only 16.5 percent. Understanding these adoption patterns will be critical for companies planning their own AI infrastructure investments.