AI in the workplace is shifting from a tool for answering questions to a system that receives objectives, breaks down tasks, utilizes various tools, and then returns results for human review. This change is significant because the way we assess productivity is also changing: it is no longer about how quickly someone gets an answer, but rather how well a task is completed from start to finish.
The term AI agent refers to AI systems that can plan multiple steps and execute actions through specific tools, such as reading documents, processing spreadsheets, running code, updating CRMs, or retrieving data from other applications. Unlike typical chatbots that tend to stop after providing a response, agents are designed to work through a series of tasks.
From Short Conversations to Long-Horizon Work
OpenAI reports that the use of Codex, an agentic tool for technical and analytical work, is moving towards longer tasks. In a user sample from May 2026, over 70 percent of users reported having requested work that was estimated to take more than one hour if done by a human. OpenAI also noted growth in usage beyond developer groups, including legal, financial, and recruitment functions.
This figure should be interpreted cautiously. The data comes from the OpenAI ecosystem and products, not from an independent survey of all workers. Task duration estimates are also made using model assessments, so they are better viewed as directional signals of change rather than absolute productivity measurements.
Nonetheless, the shift is quite clear. AI is no longer just used for summarizing emails or correcting sentences. Users are beginning to delegate tasks such as cleaning data, comparing multiple sources, drafting analyses, writing small scripts, and preparing materials that are then reviewed by humans.
Not Just Models Changing, But Job Design
For example, Anthropic has introduced agent templates for work in the financial sector. These templates combine instructions and work knowledge, connectors to controlled data, and subagents for specific tasks. Examples include market research, KYC document checks, pitchbook creation, and monthly book closing processes.
This approach highlights one important point: useful agents are typically not “versatile brains” released without direction. They resemble team members with specific responsibilities, limited access, and clear work procedures.
Imagine a marketing team wanting to understand why sales have declined. A chatbot can help create a list of possible causes. A well-designed agent can pull sales data, group it by region, compare it with ongoing campaigns, create graphs, and prepare a summary. However, decisions about budget changes still require a human who understands the business context.
Why Access Is More Important Than Intelligence
The more tools an agent can use, the greater the impact of its mistakes. An agent that merely drafts has different risks than one that can send emails, alter customer data, or execute transactions.
Therefore, the first question before using an agent is not “which model is the smartest?”, but rather “what actions is this system allowed to take?” Access should be granted based on minimum needs. An agent for generating sales reports may only need read access to specific databases and folders. It does not need permission to delete data or send messages to customers.
Another important control is human approval at high-risk points. The system may collect data and create drafts, but it must stop before sending external emails, altering official records, approving payments, or publishing information publicly.
A Commonly Underestimated Issue: Result Validation
Agents can complete more steps, but that does not mean the results are always correct. Errors can arise when selecting data sources, understanding instructions, executing code, or drawing conclusions from incomplete information.
For work that impacts money, law, customers, or reputation, the results of agents must have a verifiable trail. Keep records of the data sources used, changes made, decisions taken, and who provided approval. In technical contexts, automated testing and isolated environments also help prevent agents from damaging production systems.
The simple principle is: the greater the consequences of an action, the stronger the review process should be. A draft of a blog idea does not require the same level of control as a financial report or server configuration changes.
What This Means for Workers and Business Owners
For workers, increasingly valuable skills are not just the ability to write prompts. More importantly, it is the ability to explain objectives, break down work into stages, determine criteria for good results, and identify errors before results are used.
A person who understands business processes can typically use agents more effectively than someone who only knows their technical features. They know which data is relevant, what exceptions often occur, and when a result looks suspicious.
For business owners, AI investments should start from clear and repetitive workflows. Do not immediately create agents to “handle all tasks.” Choose one process that has measurable inputs, steps, and outcomes. Examples include converting customer request forms into structured tickets, compiling weekly reports from multiple sources, or checking document completeness.
What Can Be Done Now
- Choose repetitive tasks, not the most sensitive ones. Use administrative processes as a starting point, not decisions that directly impact customers or finances.
- Define the agent's authority limits. Differentiate read access, draft creation access, and change execution access.
- Establish a definition of done. For example, reports should include data sources, periods, assumptions, and lists of anomalies.
- Create human checkpoints. Ensure that external actions or permanent changes require approval.
- Measure tangible results. Record time saved, error rates, the amount of work that needs to be redone, and the costs of using the model.
AI agents indeed open opportunities to tackle tasks that were previously too complex or time-consuming. However, their advantages do not come solely from the model's capabilities. The greatest value emerges when agents are placed in workflows that have clear data, access limits, evaluation criteria, and humans who still hold important decisions.
With this approach, AI is not treated as a replacement for human judgment, but rather as a digital coworker capable of handling a series of tasks—as long as we know what it is allowed to do and how to check the results.
Read OpenAI's report on the changing use of agents in work and examples of Anthropic's agent templates for the financial sector.
Sources & Further Reading
– Rio Yotto @rioyotto