The most significant change in AI work today is not just faster responding models. What has changed is the duration and nature of the work. AI agents are beginning to be assigned tasks that last from dozens of minutes to hours, executing code, reading documents, calling external services, and then compiling results for human review.
OpenAI reported that by mid-August 2026, its research organization was using about 3.1 agent-workdays for every human workday. This term does not mean that AI is actually working like an employee, but rather a way to estimate the total work time of agents running sequentially or in parallel. This figure indicates one thing: the execution capacity of AI is growing faster than many people's habits of supervising it.
This is where new issues arise. When AI only helps summarize emails, mistakes are usually easy to spot. However, when AI works on analysis, transforms data, writes code, or makes initial decisions over several hours, errors can hide among many steps.
From chatbots to long-duration work
Traditional chatbots operate in a simple pattern: users ask, AI answers, then the conversation stops or continues. AI agents have a longer pattern. Users provide goals, agents break them down into several steps, use tools, evaluate interim results, and try again if necessary.
In a report from OpenAI on research acceleration, the use of coding agents among researchers sharply increased throughout 2026. By mid-August, the median researcher was using an agent daily, while some users were running multiple agents simultaneously.
A simple analogy is like having several assistants working at different desks. One checks data, one runs experiments, one writes code, and another compares results. Humans no longer complete every step manually but set goals, provide constraints, and then assess whether the results are trustworthy.
Big numbers do not automatically mean high productivity
Fact: the duration and volume of work performed by AI agents have indeed increased. Anthropic also noted that the longest-running Claude Code sessions nearly doubled in three months, from less than 25 minutes to over 45 minutes, based on research on agent autonomy.
Analysis: AI work time does not equate to work value. Agents can execute thousands of steps that do not yield useful decisions. They can also spend time fixing mistakes they made, choosing incorrect sources, or generating code that looks neat but is insecure.
Therefore, metrics like the number of tokens, runtime duration, or the number of files created are insufficient. Productivity should be assessed based on the final results: whether reports are more accurate, whether completion times have genuinely decreased, whether risks have diminished, and whether humans can make better decisions.
The biggest problems shift to the review stage
When AI only generates drafts, human work is still centered on creating something. As AI can produce many drafts, codes, and analyses simultaneously, human work shifts to reviewing.
The problem is that reviewing is often more challenging than creating. To validate financial analysis results, for example, one needs to understand the data sources, assumptions, formulas, and business context. To review code, one must test normal behavior, extreme conditions, security, and its impact on other systems.
This is why essential skills in the era of agentic AI are not just about writing prompts. Users need to be able to:
- define what results are considered correct before work begins;
- break large tasks into checkable stages;
- provide clear data sources and access permissions;
- test results with normal examples and deliberately difficult examples;
- know when work should be halted and handed over to humans.
What this means for non-technical workers?
This development is not only relevant for programmers. OpenAI reports that the use of Codex is growing across various functions, including legal, finance, recruitment, marketing, and operations. Agents can help consolidate information from multiple sources, compile reports, transform raw data, or prepare communication drafts.
However, non-technical work does not mean free from technical needs. In fact, the more tasks are delegated to agents, the more important a basic understanding of data, access, source quality, and how systems work becomes.
A marketing staff member does not need to be a programmer to use an agent. But they need to know whether customer data can be processed, whether the figures in reports come from the correct sources, and whether the emails drafted by the agent require approval before being sent.
What can be done now
- Start with low-risk tasks. Use agents to draft, categorize information, or create initial summaries before giving them access to critical systems.
- Define human approval points. Agents may prepare emails, but sending them still awaits approval. Agents may suggest code changes, but merging into the main system must go through testing.
- Keep a work trail. Document data sources, main instructions, tools used, changes made, and who approved the results.
- Measure outcomes, not activities. Compare completion times, error rates, costs, and quality of results before and after using agents.
- Limit permissions from the start. Provide the minimum necessary access. An agent that only needs to read documents should not have permission to delete files or send messages out.
Not replacing judgment, but expanding its reach
AI agents enable one person to handle more work, especially tasks that were previously delayed due to being too technical, repetitive, or requiring many small steps. However, this capability also amplifies the impact of mistakes.
In the coming years, the advantage will not only belong to those who use AI most frequently. The advantage will belong to those who can design workflows with clear boundaries, choose the right tasks to delegate, and build review processes that are just as rigorous as the creation processes.
AI may be able to generate its own “workdays.” The responsibility to determine whether those workdays produce something correct, safe, and useful remains in human hands.
Sources & further reading
- Research acceleration: the view inside OpenAI
- Measuring AI agent autonomy in practice
- How agents are transforming work
- From assistance to execution: How enterprises put AI to work
– Rio Yotto @rioyotto