Less than two years after analysts were still debating whether AI agents had any real-world applications at scale, OpenAI published research on June 25, 2026, that makes the question obsolete. Its Economic Research team tracked how Codex — the company’s agentic AI platform — spread from a developer tool into the workdays of lawyers, finance analysts, and recruiters across OpenAI’s own organization, while simultaneously recording explosive growth among non-technical users outside the company. The findings are the first large-scale empirical dataset to confirm what the enterprise AI industry had been predicting and debating: the shift from asking AI questions to delegating AI work has already happened, and it has moved well past the engineering department.
The paper, titled “The Shift to Agentic AI: Evidence from Codex,” tracks usage data across three populations: individual Codex users, enterprise account users, and OpenAI’s own workforce. Together they provide what the authors describe as “a view of what agentic AI use may look like in the future, when adoption frictions are minimal.”
What Non-Developer Adoption Actually Looks Like at Scale
The growth figures for non-technical users are the headline finding. By early June 2026, non-developer individual users had grown 137-fold since August 2025. Non-developer organizational users increased 189-fold over the same period. Within OpenAI itself — where workers started from an already above-average baseline — the comparable figure was 12-fold growth.
The Samsung Electronics deployment announced four days before the paper, giving Codex to the company’s entire South Korean workforce and its global Device eXperience division — a group that includes marketing teams, manufacturing staff, and product designers with no programming background — serves as an immediate real-world corroboration. Samsung’s Codex weekly active users in South Korea grew by nearly 800% between February 1 and the June 21 announcement date. More than 5 million people now use Codex weekly across technical and non-technical roles.
Inside OpenAI: From Engineering Tool to Company Platform
The internal adoption pattern at OpenAI tells a more specific story than the headline growth numbers. Engineers were the first movers. By December 2025, the average engineer at OpenAI had already shifted the majority of their AI tool usage to Codex, and today generates 99% of output tokens through Codex rather than ChatGPT.
But what happened next is what the paper’s authors describe as the more significant development. Legal, finance, and recruiting teams at OpenAI crossed the majority-use threshold around April 2026 — adopting faster than engineering did, even though they came later. The average lawyer or recruiter at OpenAI now generates more than 85% of their AI output tokens through Codex. Across all departments combined, Codex accounts for 99.8% of OpenAI’s weekly output tokens.
The paper includes a breakdown of what non-technical workers actually use Codex for. Finance and business operations teams generate their largest category of output in knowledge work — drafting documents, analyzing data, coordinating communication — rather than in coding. But more than a quarter of the work done through Codex by workers in business functions was classified as engineering or coding, even though those workers hold non-technical roles. Agents, the paper argues, reduce the cost of working across job boundaries.
How Agentic AI Works Differently From a Chatbot
The distinction the paper draws between conversational AI and agentic AI is worth understanding precisely, because the difference in architecture is what makes these adoption patterns possible.
A chatbot interaction is self-contained: a user sends a message, the model generates a response. An agentic interaction is a delegated workflow. When a Codex user submits a request — say, building a workflow that extracts last week’s sales data and formats it as a summary report — Codex constructs a multi-layer prompt stacking environment context, repository files, sandbox permission rules, and the user’s message, then sends it to the underlying model for inference. If the model’s response is not a final answer but a tool call to run a shell command, read a file, or execute tests, the agent carries out that call, appends the result to the prompt, and sends the whole thing back to the model again. This cycle can repeat dozens of times before the user sees output.
Each task runs in its own isolated cloud sandbox. The engineering detail that keeps this economically viable at the scale of millions of daily interactions is prompt caching via the prefix property: every new Codex task appends fresh content to the end of an existing prompt, meaning the old prompt is always an exact prefix of the new one. That property lets OpenAI reuse computation from prior inference calls. Even as the raw data sent to the model grows with each tool call, the actual model computation stays closer to linear rather than quadratic. This is not an incidental optimization — it is the mechanism that makes running Codex at enterprise scale financially sustainable. When conversations grow long enough to hit the model’s context window, Codex compacts: it replaces the full history with a compressed representation via an encrypted payload, allowing the agent to continue without losing task context.
How Long Can Agentic AI Work Unsupervised?
The paper introduces a metric for measuring agentic AI use that goes beyond active user counts: task duration as estimated human equivalent. By May 2026, 80.6% of sampled individual Codex users had submitted at least one request estimated to represent more than 30 minutes of equivalent human work. More than 70.2% had crossed the one-hour mark. And 25.6% — roughly one in four — had assigned Codex at least one task estimated to require more than eight hours of human effort.
Nearly a quarter of all Codex requests correspond to tasks a person would need more than an hour to complete. At the extreme, the 99th percentile of daily active users at OpenAI was generating more than 60 hours of agent runtime per day by June 2026, spread across multiple parallel agents running simultaneously. Users are not just asking Codex for more complex tasks — they are orchestrating multiple concurrent workflows the way a manager might assign work to a team. More than 10% of users manage three or more concurrent Codex agents at some point each week.
What These Numbers Mean for Employers and Workers
The practical implications extend in several directions simultaneously.
For employers deciding whether to deploy agentic AI beyond engineering, the paper provides the most specific benchmark currently available: at a well-instrumented organization where adoption frictions are minimal, the transition from chatbot to agent becomes the majority mode of AI interaction — for every department — within roughly a year of the tool reaching general availability. The data also shows that the growth trajectory for non-technical users is steeper, not slower, than for developers.
For workers, the data confirms something prior chatbot productivity research did not fully capture: agentic tools expand what an individual can accomplish beyond their formal job description. Over a quarter of Codex work by employees in business functions fell into engineering or coding categories — work those employees would previously have needed to request from a technical team. That boundary is compressing.
The paper’s authors — drawn from OpenAI, Columbia Business School, the University of Pennsylvania’s Wharton School, and Duke University’s Fuqua School — are careful to note that OpenAI represents a “uniquely favorable environment for agentic AI” and that its internal adoption patterns are not representative of typical organizations today. Workers are highly familiar with AI, usage has no quantity restrictions, and organizational buy-in is high.
That caveat has teeth. The Next Web noted that every metric in the paper comes from OpenAI itself — a company with a direct financial incentive to promote the product it is measuring — and that no independent third party has verified the findings. The paper does not address whether OpenAI employees are encouraged to use Codex, a relevant omission given that near-universal adoption inside a company selling the product is not the same as organic demand.
The external adoption data, however, tells a different story. Among organizational users of Codex outside OpenAI — companies on Business and Enterprise plans — Codex already accounts for 63.3% of output tokens. Individual users remain at 16.5%. The gap between individual and organizational adoption suggests the key variable is not model capability but workflow integration: access to relevant files, management expectations, and the presence of review processes that allow workers to trust what agents produce.
Not every signal is positive. Separate reporting indicates that a growing number of power users describe supervising multiple simultaneous Codex workstreams as cognitively draining — a human cost of parallel delegation that the paper’s output-volume metrics do not capture.
What Agentic AI Still Cannot Do on Its Own
The paper is also explicit about what changes and what does not. Even among OpenAI’s most intensive users, agents do not eliminate the human role — they shift it. Users define tasks, set constraints, review outputs, and own the consequences when an agent produces something wrong. What changes is the unit of delegation: instead of asking an assistant to complete a function, a worker can hand off a multi-hour chore and allocate their time to design and review.
The paper’s research methodology relies on an LLM-as-judge pipeline to estimate task duration — each Codex transcript is evaluated by a language model that estimates how long an experienced human would need to complete the same work. The authors describe these thresholds as “directional rather than exact,” and the sample covers a random 0.1% of individual users.
Frequently Asked QuestionsFrequently Asked Questions
What is the difference between agentic AI and a chatbot, and why does it matter for workers who are not engineers?
A chatbot interaction is a single exchange: the user asks, the model responds. An agentic AI system like Codex runs a loop — receiving a goal, selecting tools to use, executing those tools, inspecting the results, and repeating until the task is done. The user delegates a workflow rather than asking a question. That difference is what makes it useful beyond engineering: a recruiter or lawyer can describe what they need in plain language and let the agent work through the steps, rather than needing to specify each action themselves.
Who is actually using agentic AI tools outside engineering, and what are they using them for?
According to OpenAI’s research paper, the fastest-growing non-technical Codex user categories are data analysis, research, and the creation of reports and work documents. Within OpenAI, knowledge work is the largest output category for finance, business operations, marketing, and product teams. More than a quarter of the output from non-technical roles falls into engineering or coding tasks, meaning workers are using agents to cross into technical work they would previously have needed to request from a developer.
Does OpenAI’s data prove that agentic AI is delivering real productivity gains in the broader economy?
Not conclusively. The paper measures adoption depth and task complexity inside OpenAI — an organization the authors themselves describe as an unusually favorable environment for agentic AI — and tracks user growth among external Codex subscribers. It does not measure final productivity output, revenue impact, or economic value delivered. All figures come from OpenAI itself, which has a commercial interest in promoting Codex. Independent verification has not yet occurred. The paper’s most defensible claim is that agentic AI is being used at scale, for complex, long-horizon tasks, by workers well outside engineering — and that this pattern is accelerating rapidly.
What are the security risks of running agentic AI at work, and should enterprise teams be concerned?
Agentic AI tools introduce risks that chatbots do not. Because Codex agents execute code, read files, and call external tools autonomously, a compromised agent credential gives an attacker persistent, silent access to whatever the account can reach — not just a chat interface. Researchers at Aikido Security documented a malicious npm package that had been quietly exfiltrating Codex authentication tokens for approximately a month before discovery, attracting more than 29,000 weekly downloads. A patched vulnerability disclosed in early 2026 had allowed command injection through GitHub branch names, granting read/write access to victim codebases. Enterprise teams deploying Codex at scale should apply least-privilege and behavioral monitoring to AI agent identities with the same rigor applied to human credentials.