OpenAI's Landmark Study: The Shift to Agentic AI Is Already Here

OpenAI has released a major new research paper, "The Shift to Agentic AI: Evidence from Codex," in collaboration with economists from Wharton (University of Pennsylvania), Columbia Business School, and Duke University. The study draws on real-world usage data from millions of interactions with Codex, OpenAI’s agentic AI tool, across three distinct groups: individual users, organizational users, and OpenAI’s own employees.

OpenAI employees serve as a "showcase of the future"—a low-friction environment where adoption barriers are minimal and tools are used at the frontier of capability. The findings reveal a profound transition: AI is moving from a conversational assistant (ChatGPT-style) to autonomous agents that handle delegated, long-horizon tasks.

Tokens Tell the Real Story: Agents Are Dominating Work

The most striking metric is the share of output tokens generated through Codex rather than traditional chat interfaces:

OpenAI employees: 99.8%;
Organizations: 63.3%;
Individuals: 16.5%.

This shows that at the cutting edge, agents have almost completely replaced chatbots as the primary interface for knowledge work.

People are no longer just asking for advice — they are delegating real work. Task complexity has surged. Among individual users, the share submitting requests estimated to require more than 8 hours of human effort grew nearly tenfold, from 2.1% to 25.6% since the start of the year.

Users typically front-load the most complex parts of a task at the beginning of a session, followed by iterative refinements.

Development Leads, But It’s Far More Than Code Generation

OpenAI's Landmark Study: The Shift to Agentic AI Is Already Here Engineering and software development remain the core use case, but the scope is broad. Agents help engineers understand unfamiliar codebases, set up environments, verify changes, and produce documentation. Among organizational engineers, 26.8% of tokens now flow through Codex — five times higher than in January. For active users in this group, that figure reaches 88.3%.

Non-developers are adopting rapidly too. Codex enables business, legal, finance, and operations teams to handle technical tasks that once required specialized support, expanding what individuals can accomplish.

Parallel Workflows and Skill Usage on the Rise

Intensive users are building sophisticated, parallel agent systems:

67.4% of organizations and 63.9% of individuals still rarely (or never) run agents in parallel.
At OpenAI, 28.6% of users managed 5+ agents simultaneously at some point during a recent week.

OpenAI's Landmark Study: The Shift to Agentic AI Is Already Here The use of "skills" — pre-defined instructions for complex, reusable workflows — has grown sharply. The share of users invoking at least one skill rose from 5.4% in March to 26.6% in June (25.7% for individuals, 30.4% for organizations, and 96.2% at OpenAI).

The productivity impact is dramatic. From November 2025 to June 2026, median token output per OpenAI employee increased at least 10x across functions. Lawyers saw a 13x rise; researchers more than 50x.

Lessons from Electrification: Rebuild the Workflow

The authors draw a compelling analogy to factory electrification. Productivity gains didn’t come merely from swapping steam engines for electric motors, but from entirely redesigning factory layouts around the new technology. With agentic AI, the "rebuilding" costs are negligible—restructuring workflows requires little more than time and imagination. This suggests the transition will happen faster than previous industrial shifts.

New Metrics for the Agentic Era

Traditional metrics like Daily Active Users (DAU), chat volume, or message counts are becoming obsolete.

The study argues we should instead measure:

Complexity of delegated tasks;
Duration of agent runtime;
Degree of parallelism.

Human value is shifting from "being able to do the work" to "setting the right tasks, verifying outputs, and coordinating multiple agents." The most valuable organizations and individuals will be those ready to redesign processes for delegation and robust verification.

What This Means

OpenAI’s internal transformation offers a preview of broader economic change. When capable agents are readily available with minimal friction, people quickly move beyond short interactions to long-running, multi-step work. Non-technical roles are expanding their reach, and entire workflows are being re-architected around autonomous AI labor.

The agentic future isn’t coming — it’s already underway for those at the frontier. For everyone else, the question is no longer whether to adopt agents, but how quickly they can rebuild their processes to harness them. The companies and professionals who thrive will be those who treat agents not as fancy chatbots, but as tireless digital colleagues capable of handling hours — or days — of coordinated work.