OpenAI's latest research shows that agentic AI tools like Codex are moving far beyond coding, with lawyers, recruiters, and finance teams now delegating complex multi-step tasks to AI systems rather than using them for simple advice. A new study titled "The Shift to Agentic AI: Evidence from Codex" reveals that inside OpenAI itself, every department from Legal to Finance to Recruiting now relies on Codex as its primary AI work tool, marking a fundamental shift in how knowledge workers interact with artificial intelligence.

The distinction matters because it signals a move away from conversational AI, where users ask questions and receive answers, toward agentic AI, where users delegate entire workflows to systems that can inspect files, execute commands, and create or modify documents autonomously. OpenAI framed this shift clearly: "Agentic AI changes the unit of knowledge work from single interactions to delegated, long-horizon tasks".

What Makes Codex Different From ChatGPT?

While both tools come from OpenAI, the research treats them as distinct systems serving different purposes. ChatGPT functions primarily as a conversational assistant, answering questions and providing information. Codex, by contrast, is designed for production work, where users ask it to perform concrete tasks like debugging code, drafting legal documents, analyzing financial data, or refactoring applications.

The usage patterns tell the story. Among OpenAI employees, Codex accounted for 99.8 percent of output tokens generated across both tools as of June 11, 2026. Among organizational users on paid plans, Codex represented 63.3 percent of tokens. Even among individual users, Codex usage reached 16.5 percent, despite being less mature than ChatGPT in the broader market.

How Are Non-Developers Using Agentic AI?

• Legal Department Work: Lawyers at OpenAI now generate more than 85 percent of their AI-mediated output tokens through Codex, using it for document drafting, contract analysis, and legal research tasks that previously required manual effort.
• Finance and Business Operations: Finance teams use Codex for data analysis, financial modeling, and report generation, with the department crossing into majority Codex usage around April 2026.
• Recruiting and Talent Acquisition: Recruiters delegate candidate screening, job description writing, and interview coordination to Codex, with non-developer adoption rising 137x among individual users and 189x among organizational users since August 2025.

The surprise is not that Codex writes code. The surprise is that OpenAI is presenting evidence of a workplace-wide shift in which agentic systems are being used to delegate technical and knowledge-work tasks across departments that have nothing to do with software development.

How Long Are These AI-Delegated Tasks?

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 dramatically, 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 of work.

These duration figures are model-estimated and based on a sample of users who opted into training data sharing, so OpenAI cautioned they should be treated as directional rather than exact. However, the pattern is clear: heavy Codex users are delegating increasingly complex, multi-step workflows rather than asking for quick answers. Among the most intensive OpenAI users, Codex functions "less an assistant answering requests and more like a workflow system".

What Does This Mean for Job Search and Hiring?

The shift toward agentic AI is already reshaping how companies connect with job seekers. Appcast, a major recruitment marketing platform, has announced a strategic partnership with OpenAI to integrate real-time job listings directly into ChatGPT. When job seekers ask ChatGPT about career opportunities matching their skills or location, relevant job postings from Appcast's employer clients will surface automatically within the conversation.

The partnership addresses a longstanding frustration in recruitment: candidates wasting time applying for positions that have already been filled. By syncing directly with Applicant Tracking Systems (ATS), the integration ensures that job listings stay continuously updated and accurate, rather than relying on outdated web scraping that often pulls expired information.

For employers, the stakes are high. As a significant portion of the workforce migrates toward generative AI for daily information gathering, companies risk losing visibility if their job listings are not formatted for large language models (LLMs), which are AI systems trained on vast amounts of text data. By cementing a direct link with OpenAI, Appcast is positioning its clients to maintain competitive advantage and drive high-intent candidates directly to their career sites.

Why Should Organizations Care About This Shift?

OpenAI's research does not directly prove productivity gains or business outcomes. Output tokens, which measure the amount of text generated by the AI, indicate AI-mediated work but are not the same as completed work, higher-quality work, or economic value. However, the scale of adoption inside OpenAI itself suggests that knowledge workers across departments are finding genuine utility in delegating tasks to agentic systems.

The company cautioned that OpenAI is an unusually favorable environment for agentic AI adoption. Workers are familiar with frontier AI models, usage costs are low, organizational buy-in is high, and many workflows are close to the systems being developed. "OpenAI usage is therefore not representative of the typical organization today," the researchers noted.

Still, the research points to a broader workplace transformation. As agentic AI tools mature and become more accessible, organizations across industries will likely face pressure to integrate them into their workflows. The question is no longer whether agentic AI will reshape knowledge work, but how quickly and in which domains the shift will accelerate.