# From Asking Questions to Delegating Work: What the Codex Study Shows > OpenAI's Codex research shows how AI use is shifting from questions to delegated work—and what companies should measure before calling it value. > 💡 **TL;DR: Key Takeaways** > - **In OpenAI's Codex research, the unit of work changes before the volume of use does.** Users are not only asking questions. They are delegating multi-step work across research, code, documentation, and validation. > - **Longer tasks, concurrent agents, and reusable skills indicate that people are building work systems.** None of these measures proves productivity or ROI on its own. > - **The practical issue is not how often someone uses the tool.** It is whether the workflow preserves context, connects tools, defines human decision points, and verifies the output. > - **Companies need to look beyond messages and tokens.** Task complexity and agent activity must be measured alongside accepted output, rework, quality, cycle time, cost, and risk. Asking an AI tool, “What might be causing this bug?” is not the same as asking it to inspect the codebase, find the cause, implement the fix, run the tests, and update the documentation. The first request produces an answer. The second delegates a piece of work. That is the shift documented in the OpenAI Economic Research team's 50-page report, [The Shift to Agentic AI: Evidence from Codex](https://cdn.openai.com/pdf/5d1e1489-21c0-43e4-9d42-f87efdbf0082/the-shift-to-agentic-ai-evidence-from-codex.pdf){.dofollow target="_blank" rel="noopener"}. AI use is moving from conversation toward agentic workflows that use tools and carry out multi-step tasks. The practical significance is not simply that “more people are using Codex.” It is that people are starting to turn AI from a standalone assistant into a system in which research, production, tools, and verification operate together. I use Codex and Claude CLI across source research, document and code preparation, custom MCP development, and image generation. This article is therefore not a translation of the report's findings. I will first separate what the research shows from what it does not, then return to what the same pattern looks like in my own work. ## From asking questions to delegating work OpenAI treats ChatGPT primarily as a conversational tool and Codex as an agentic one. The researchers also acknowledge that the distinction is not absolute. ChatGPT can use tools, while a Codex conversation may remain a simple question-and-answer exchange. An external tool call is therefore not enough to define agentic work. The report's more useful distinction is between **consultation** and **delegated production**. In consultation, the model explains an approach or generates options. In delegated production, it inspects existing files, forms a plan, uses tools, and produces a concrete output or change. The research finds Codex being used across code implementation, debugging, validation, documentation, data analysis, research, planning, and communication. Two years ago, I might have asked an AI tool, “How should I create an Editorial Principles document for my website?” Today, I ask Codex or Claude CLI to research the subject, relevant sources, and best practices, then prepare a draft Editorial Principles document aligned with my Brand Guide. The first request generates an explanation of how to do the work. The second delegates a package that includes both research and writing. You may recognize the same change in your own use: what began as asking AI questions has gradually become asking it to do more of the work. Software is a useful starting point for this shift. The work is digital, can be decomposed, and can often be checked through tests. But the change described in the report is broader than writing code. Instead of working through every step with the model, the user defines the objective and constraints, then delegates execution. ## What is the research based on? The report compares three groups: individual Codex users, people on organizational accounts such as Business and Enterprise, and OpenAI employees. The analysis does not rely on researchers reading conversations one by one. OpenAI says its privacy-protecting automated pipeline aggregates usage telemetry, classifies tasks, and estimates how long some tasks would take an experienced person without AI assistance. Not every finding comes from the same sample. Task distributions use broader samples. Task complexity is estimated from a random 0.1% sample of individual accounts whose users allow their data to be used for model training. The human-time equivalent is a model estimate, not a time sheet. These details do not make the report irrelevant. They limit what it can prove. The study offers strong provider telemetry on how people use the product. It is not an independent impact study and does not establish causal productivity gains or ROI. ## Four findings point in the same direction ### 1. The range of delegated work is expanding Codex use is not limited to producing new code. Understanding existing systems, debugging, validation, repository operations, and documentation all account for meaningful activity. Among OpenAI employees, the work also extends into research, planning, communication, data analysis, product, sales, and recruiting. OpenAI's internal pattern should not be generalized to an ordinary company. Its employees are close to the product, face fewer access constraints, and work in an environment with extensive internal knowledge sharing. Even so, the example indicates that the portfolio of delegable work can broaden as adoption deepens. I saw this expansion in my own work while moving a legacy repository to a completely different technology stack. The agent first inspected the codebase and prepared a migration report. After I decided the approach and scope, it wrote the code, documentation, and tests; created browser tests; prepared a smoke-test list for checks that only I could perform; produced a deployment checklist; and set up the CI/CD pipeline. This does not mean every agent can take every project to production on its own. It is one observed example of how a work package can extend from analysis to delivery preparation when access, decision points, automated tests, and human checks are clearly defined. ### 2. The task horizon is getting longer The share of active individual users who submitted at least one task that the model estimated would take a human more than one hour rose from 35.4% in December 2025 to 70.2% in May 2026. The share who submitted at least one task estimated at more than eight hours rose from 2.1% to 25.6%. These figures do not show that the tasks ran successfully for eight hours. A request may fail, stop halfway, or require substantial correction. What they show is that users are starting to delegate larger pieces of work at once. ### 3. Some users manage agents in parallel In the week ending June 11, 2026, more than 10% of users managed at least three concurrent Codex agents. Most external users still stayed within one workflow or ran no more than two tasks in parallel. Concurrency was much higher inside OpenAI. The human role changes in this pattern. The user must separate the work, provide the right context, monitor progress, intervene when necessary, and combine the results. High concurrency is not a success measure by itself. In a poorly designed workflow, it may simply create a larger review queue. When appropriate, I design workflows so that tasks can run in parallel and delegate bounded work to subagents. More recently, the parallelism has extended beyond a single task. While one workflow runs in one CLI window, I open another CLI window and move a separate topic forward rather than waiting. This is not a recommendation that everyone should open as many simultaneous tasks as possible. In my experience, concurrency changes the user's work rhythm as well as the number of agents: one process continues while the user advances another piece of work. ### 4. Repeated work is being encoded in skills and plugins The share of active Codex users who invoked at least one skill rose from 5.4% on March 1 to 26.6% on June 11, 2026. In the report, a skill is a reusable set of instructions and a work recipe for a specific job. A plugin can package skills with application integrations, Model Context Protocol configuration, hooks, and other assets. Custom skills can preserve local context such as an organization's writing standards, reporting routines, or control steps. This is central to turning AI into a work system. Instead of explaining the same context in every conversation, the user turns the way the work should be done into a durable operating layer. Reuse is not proof of quality, however. A bad rule can also scale consistently. ## When does AI become a work system? A long prompt or a large number of conversations does not create a work system. Five connected elements do: - The boundary of the delegated work is explicit. - Access to the required files, data, and tools is controlled. - Repeated context is preserved in a skill, rule, or another system record. - The point where the agent stops and a human decides is defined. - The output is verified through a test, source, record, or accountable person. That is why [the distinction between an AI agent and a fixed workflow](/when-do-you-actually-need-an-ai-agent) matters. Not every task requires open-ended judgment; a more constrained automation may be more reliable when the path is predictable. When an agent is appropriate, [the authority given to the AI system](/how-much-authority-should-ai-have) should still be limited according to the risk of the work. ## What does this pattern look like in my own work? I can make it concrete through the content workflow for this website, which has two different starting points. In the first, the idea and often the initial draft come from me. I ask Codex or Claude CLI to research the subject, supporting evidence, and best practices, then enrich and edit the draft under my Brand Guide and editorial rules. Claims are checked, existing articles are reviewed, and relevant internal links are proposed. The second starting point is source monitoring. This workflow helps me follow developments, classify new material, and notice a possible article. I still make the decision that something is worth writing, and I provide the central point, angle, or initial draft. The item that surfaced the topic does not have to remain the only source; it can be strengthened or bounded by additional research. Once the subject is clear, the same work system continues through Turkish and English drafts, documents, and code when needed. If existing tools are insufficient, custom MCPs can be designed and implemented. Codex has assisted with parts of many of my recent blog articles. I have also generated all the blog's visuals through ImageGen in Codex, and an agent prepares the ALT description for the generated image. This is not a Codex-only product review or an ROI claim. The editorial thesis, source decisions, language, and publication decision remain mine. ## Why tokens and messages are not enough One of the report's most useful warnings is the limit it places on its own usage figures. In an agentic system, active users, chats, messages, and token volume are inadequate measures of the value of the work. The 23.2 billion tokens I have used since mid-July make that limitation concrete. The same screen shows a peak of 1.47 billion, a 33-day streak, a longest task of 7 hours and 18 minutes, and a highest weekly bar of roughly 6.13 billion. These figures describe the intensity and duration of use. They do not show that every token became valuable or completed work, that the outputs were correct, or that the activity produced business value, ROI, or quality. One user may write dozens of short messages. Another may give one instruction and manage agents that work for several hours. The activity patterns are different, but neither proves that the result is correct or valuable. OpenAI therefore proposes measures such as task complexity, runtime, workflow reuse, concurrency, and production output. Companies need to connect those measures to operating results: - How much of the output was accepted? - How much human correction or rework was required? - Was the error found before delivery or after it? - Did cycle time, capacity, quality, cost, or risk change? This is where [using AI and creating business value from AI](/how-can-companies-create-value-from-ai) diverge. Without connecting usage intensity to operating results, a company can mistake high activity for success. ## The implication for companies OpenAI's research makes a convincing case that AI use is shifting from question-and-answer interaction toward delegated production. Longer tasks, parallel agents, and reusable workflows are separate signals of that change. The report does not show that these systems automatically make companies more productive. That decision requires task selection, access design, controls, and outcome measurement. A company's first objective should not be more agents or higher token consumption. It should make explicit which work is being delegated and why, what context and tools the agent can use, where a human makes the decision, and how the output will be verified. That is when an AI assistant becomes a work system: not when people talk to it more, but when it is connected to the way work is done in a controlled and measurable way. --- Language: English License: CC BY 4.0 License URL: https://creativecommons.org/licenses/by/4.0/ Scope: Evren Bal-authored text, unless this article expressly states otherwise. Excluded: Third-party material, quoted excerpts, logos, and separately marked images retain their own rights. Attribution: Credit Evren Bal, link to the canonical source and license, and indicate changes. Source: https://evrenbal.com/from-asking-questions-to-delegating-work