Why Are People Busier as AI Takes on More Tasks? Anthropic Research Reveals Attention Bottlenecks in Businesses

Managing multiple AI agents at once might sound like having a team on standby at all times. However, every additional result an agent delivers may also require an extra round of verification, a round of selection, and a decision regarding accountability. While companies are purchasing generative capabilities, what is ultimately put to the test is the organization’s ability to process and utilize those results.

An internal employment study released by Anthropic on October 16, 2026, brought this contradiction to the forefront. It deserves careful consideration by businesses, but should also be viewed with a degree of detachment: this is a study conducted by an AI company about its own organization and cannot be directly viewed as the future for all industries.

Two numbers—read the denominator first.

The report shows that the proportion of entry-level technical positions in technical hiring dropped from 27% to 9%; by August, the working hours of interactive agents managed by employees had reached six times the estimated working hours of employees. The agent metric aggregates sub-agents’ reasoning and tool usage, while the employee metric is estimated based on a 40-hour workweek. The former reflects a change in the hiring structure and cannot be directly interpreted as a decline in the number of entry-level hires; the latter cannot be converted into a six-fold increase in labor value.Source: Anthropic Research

It is important to understand the distinction between these two concepts. Suppose a company increases its overall hiring; even if the proportion of a certain type of position decreases, the absolute number of employees in that role may still increase. Similarly, when six agents operate in parallel for one hour, they may either perform complementary tasks or repeatedly explore the same erroneous path. Only results that have passed quality inspection and been adopted by the business are recorded in the productivity ledger.

Therefore, any discussion of the impact of AI on employment must address at least three distinct issues: how many people companies need, what kind of people they need, and exactly what tasks each person performs. Reducing these three issues to the question of “how many jobs AI will replace” overlooks the changes in job duties—which occur earliest and are most easily overlooked by management.

Why Does Faster Execution Actually Make Management More Difficult?

You can think of an AI team as an assembly line: the front end poses a question, the middle section processes it in batches, and the back end determines whether the results are reliable and worth acting on. If the middle section suddenly speeds up, the back end’s ability to make judgments may not necessarily improve at the same rate.

For example, the marketing team asks an agent to generate ten competitor analyses at the same time. Readers should verify whether the information is outdated, whether the marketing messaging is consistent, and whether each report cites the same unverified piece of news. Generating nine additional reports only makes sense if they help the team make better decisions faster; otherwise, reading, deduplicating, and reworking them become hidden costs.

Even more challenging to manage are interruptions. Whether an agent spots a formatting error or a critical data conflict, both may trigger equally prominent alerts. People must constantly re-establish context to determine which issue warrants immediate attention. The more fragmented the tasks, the more likely the workday is to be broken down into small chunks that leave no room for in-depth judgment.

This also explains why “adding another review agent” may not necessarily solve everything. While reviews can help distribute the workload, questions remain regarding who defines the review criteria, who makes the final decision when multiple agents’ opinions conflict, and how to detect errors in shared data sources—all of which still require careful design. Automation may shift the focus of oversight, but it does not automatically create a system of accountability.

Behind the changes in hiring lies a calculation regarding talent development

For managers, the most appealing arrangement is to have experienced employees make decisions while letting AI handle the basic execution. But if junior employees lose the opportunity to grapple with real-world problems, where will their future judgment come from?

Insights typically come from experience—from what you’ve done, what you’ve missed, and what you’ve corrected. Through repeated cycles of researching, drafting, and troubleshooting, newcomers gradually come to understand which details are prone to inaccuracies. If they’re only asked to read AI conclusions without being required to reconstruct the reasoning process, short-term efficiency may improve, but long-term training may become superficial.

Companies can redesign apprenticeship tasks. For example, they can have new hires independently list acceptance criteria and then compare them with the actual results; require them to identify key assumptions, find counterexamples, and explain why they accept or reject a particular suggestion. The focus of senior employees“ evaluations shifts from ”how much you’ve done“ to ”how you demonstrate that the results are reliable.”

This is not a reason to retain inefficient labor, but rather to clearly define training objectives. Entry-level positions may change, but there must still be alternative training pathways to develop the domain knowledge, sense of responsibility, and ability to identify errors that are acquired while performing basic tasks.

What This Study Does and Does Not Show

The study utilized internal records, surveys, and interviews to compare changes in employee behavior before and after the adoption of the tool; however, adoption was not random. More proactive employees with heavier workloads may have adopted AI earlier, and the company’s rapid growth may also have influenced the results. These are the limitations identified and discussed in the report.Source: Research Methods and Limitations

Furthermore, since the research subjects both develop and sell tools, they have a clear self-interest in the research; the resource conditions and adoption culture within the company are also unique. Therefore, it is more appropriate to view this as a case study of an organization with high-intensity AI adoption—one intended to identify potential issues—rather than to use it to predict how many employees typical companies should lay off.

Based on this, the paper proposes the following conclusion: A company’s automation maturity should include acceptance and coordination capabilities. While this conclusion has managerial explanatory power, it has not yet been validated as a causal conclusion across industries.

Four Things Companies Can Change Right Now

First, categorize tasks by risk level. Tasks that are low-risk and reversible may be automated; tasks involving customer commitments, payments, or changes to critical data should require explicit approval. Authorization levels should vary according to risk to prevent all issues from being funneled to a single person.

Second, reformat reports to facilitate decision-making. Prioritize conclusions, rationale, changes, and pending decisions; provide the full context only as needed. Each delivery should ideally address a single, clear acceptance criterion, rather than a lengthy output that simply says, “Please review.”

Third, measure net收益. In addition to completion volume, track the rework rate, manual verification time, wait time, and actual adoption rate. For a given task, compare the three workflows—“purely manual,” “AI-assisted,” and “executed by an agent followed by manual verification”—to prevent mistaking cost shifting for cost elimination.

Fourth, prioritize training and profit-sharing. The speed gains brought about by new tools should not automatically become an excuse to keep piling on more tasks. Only by investing a portion of the profits in skills training, job reassignments, and appropriate incentives can employees be encouraged to acknowledge mistakes and improve processes.

AI can free up more resources, but human attention remains limited. The metric truly worth optimizing in the next phase is how many reliable outcomes are accepted, adopted, and drive progress each workday. Those who can streamline this process are more likely to achieve sustained efficiency gains.

This article is an industry analysis. Fact-checking was completed as of October 18, 2026; the management recommendations are based on the author’s analysis and do not imply that their effectiveness has been scientifically proven.