On October 16th, Anthropic published a study on work and employment using itself as a case study. The figure that is most likely to be intercepted and circulated is that as of August, the ratio of the operating time of the company's internal interactive AI agents to the working hours of employees was approximately six to one. However, interpreting this as "one employee being replaced by six agents" is neither in line with the research methodology nor does it accurately represent the findings. The operating time of these agents does not equate to the total human working hours; model generation, waiting times, rework, and acceptance processes are all part of the actual production workflow. What is truly worth noting in the report is that after AI was introduced into the organization, there were changes in job structures, work rhythms, and management costs simultaneously.

This study utilized employees from within Anthropic, as well as materials related to usage, output, questionnaires, communication, and recruitment from AI. It was also supplemented by interviews with 17 employees. This is a highly special sample: the subjects of study are both model developers and among the most active users of these models; it can reflect the early changes in cutting-edge companies, but it cannot be directly used to estimate the unemployment rate across society. The company noted in the report that the total number of employees continued to grow, and recruitment in technical departments did not cease. However, the proportion of junior technical positions within technical departments decreased from 21% at the beginning of 2025 to 14%, and the proportion of junior technical positions among newly hired technical staff dropped from 27% to 9%. These are data reflecting "structural changes," rather than statistics on "how many people have been laid off."

Shift from the number of positions to the breakdown of tasks

Changes in recruitment structures are partly due to companies moving from the research and development phase to commercialization: functions such as sales, finance, and recruitment typically expand as business grows; another part is related to programming and review tasks being taken over by intelligent agents. The report states that about one-third of machine learning managers and one-fifth of software engineering managers indicated in a survey in April that they had reduced recruitment due to AI. At the same time, non-technical employees have begun to build their own tools, and some managers are also writing code. These phenomena indicate that the boundaries once defined by job titles are starting to soften: a salesperson no longer has to wait in line for engineering resources, and an engineering manager is no longer limited to just coordinating tasks.

However, "tasks can be automated" does not equate to "responsibilities can be automated." Customer communication, legal judgment, changes to production systems, and verification of model outputs still require human involvement to bear the consequences. Sales, legal, and recruitment cases under study are more about entrusting repetitive tasks to tools, allowing employees to focus their time on interviews, risk assessment, or process redesign. The result is not that machines take over entire positions in the traditional sense, but rather that the tasks within a position are reorganized, rearranged, and then recombined. If companies only consider how much writing time is saved without accounting for review, approval, and rework, they may overestimate their net profits.

The report also shows that cross-departmental mobility of employees was more frequent in 2026 than in 2025. The study observed that for employees who had been in their positions for at least one year, the likelihood of cross-departmental transfers each month was about 50% higher. It cannot be proven that each transfer is caused by AI, but it does highlight a new human resources issue: when old tasks diminish and new ones emerge, can companies provide internal retraining and job reassignment instead of simply stopping recruitment? This is especially true for junior positions. In the past, many newcomers learned the business by fixing small errors, writing simple code, and organizing materials. If these introductory tasks are taken over by intelligent agents, companies will need to design new mentoring and practical training programs; otherwise, the pipeline for cultivating senior talent in the future will narrow.

Human attention is becoming a bottleneck.

Another observation that is easily overlooked in the report is the waiting time. The study subtracts the active working time from the total duration of interactive intelligent conversations and estimates that about 80% of the conversation time is spent waiting for human input. This number does not mean that “only 20% of the time is effectively working”; rather, within the context of a specific interactive conversation, it is still the human instructions, choices, and confirmations that determine whether the process can continue. The faster the model produces results, the more items that need to be reviewed accumulate on the reviewer’s screen, and thus the bottleneck shifts from writing the initial draft to allocating attention.

As a result, employees have developed a new type of "meta-work": writing agent skills, setting up monitoring panels, planning task queues, and reviewing multiple agent processes. It may not sound like traditional output, but it constitutes the infrastructure necessary to transform model capabilities into stable business results. A system that can generate ten plans overnight may not be more effective than one that carefully completes just one plan, especially in the absence of evidence tracking, version management, and clear division of responsibilities. Teams also need to decide which outputs must be reviewed item by item, which can be checked sample-wise, and which errors could cause irreparable losses.

For other companies, the value of this report does not lie in replicating a six-to-one ratio. Industry factors, regulatory constraints, data quality, and customer tolerance for errors are all completely different from those of Anthropic. More relevant questions are: Which tasks have actually shortened delivery times? Has the saved time been translated into better service or more orders? Have review costs increased accordingly? Do junior employees have new learning pathways? Do managers possess the ability to interpret model outputs? These questions need to be measured within the company's actual processes and cannot be answered by product demonstrations alone.

The study also reminds readers to be cautious when interpreting causal relationships. Companies that are growing rapidly will naturally adjust their job structures, and a decrease in the proportion of technical staff may also be related to business expansion; employee interviews can explain some of these mechanisms, but they cannot replace statistical evidence across different companies and over different time periods. Therefore, the claim that “AI has led to a large-scale disappearance of jobs” goes beyond what this material can support. It is more prudent to say that in companies that make the most extensive use of AI, there have been changes in the relative demand for junior technical positions, the way employees work, and the burden of review tasks that deserve continued monitoring.

Source: Anthropic, October 16, 2026, " How employment is changing at Anthropic ": https :// www.anthropic.com / research / employment-at-anthropic