As agents do more, human attention becomes the bottleneck

Even when the researchers count only employee-directed interactive sessions, the total reaches 6× human work hours by August. That includes subagents, which an agent starts to handle parts of a task. The chart shows 19× for research organizations. These ratios compare accumulated working time, not the number of employees an agent could replace.

Human hours are estimated at 40 hours per employee per week. Agent hours include model processing and active tool use, but exclude waits longer than 30 minutes. The March comparison includes both autonomous and interactive agents; the August ratio excludes the unattended fleets.

Output per employee also rose across several jobs. Taking December 2025 as 100, the chart shows code commits per engineer reaching 5.1× their earlier level. A commit records a set of code changes; the report's text summarizes the increase as 5x the pace six months earlier. Queries per data scientist and contracts per legal employee both reach 3.1×, deals per salesperson 1.7×, and hires per recruiter and invoices per accounts-payable processor both reach 1.4×.

Those increases cannot all be credited to AI. Anthropic's business was growing rapidly, and selling a stronger model can itself make deals easier to close. To separate these changes as far as possible, the researchers track the same employees before and after they start using agents, and compare them with employees who have not adopted the tools in the same month.

The engineering comparison includes 192 adopters and 74 people who do not adopt during the study. The first month with at least five agent sessions counts as the start of adoption. Engineers add substantially more code in the following months. A separate comparison finds that sales representatives close deals with more clients after adopting agents.

The authors point out that employees choose when to start using these tools. Someone preparing a large project may adopt them in advance; a salesperson with more opportunities may have more reason to automate routine work. Engineers adopting fully autonomous agents were already on a steep upward trajectory beforehand. The comparisons therefore cannot completely separate tool use from other reasons for rising output.

Does more code also mean more mistakes? The researchers compare agent-written code with code written directly by the same employee, in the same month and code directory. They find no difference in how often lines are later removed or reverted, meaning the change is undone. These measures capture only part of code quality, and the authors note that AI errors may be harder to spot.

The interviews show what this looks like day to day. One account executive said Claude took over around 20% of their previous tasks, including drafting emails and integrating with customer-management software. Another used the time saved for face-to-face conversations and could speak to 30% more customers. Their habit was to look for a way to automate a task whenever they found themselves doing it a second time.

A lawyer used Claude to scan for risks, sort legal questions and take a first pass through public-facing copy. A recruiter used it to find candidates, track applications and carry out analysis, leaving more time to update interview questions and meet applicants.

New models also lead employees to attempt harder tasks. The researchers use Claude to score coding sessions with identifying information removed. They set a threshold at the hardest 10% of tasks handled with the old model. After a new model is released, tasks above that threshold rise from 10% to 13% of all tasks, an increase of around 30%, and remain elevated in the following weeks. The easy end changes little.

Employees have high expectations for future capabilities too. As of August, they expected AI to be able to do 75% of their work by 2030, compared with 28% for other Americans. In a separate survey question, agreement that work processes made employees as productive as possible rose from 44% in late 2024 to 58% in November 2025. It reached 75% in April 2026, but that question dropped the words “as possible,” so the results are not directly comparable.

Employees also write reusable instructions explaining how agents should work. These grew from a few hundred in July 2025 to 8,500 a year later. A lawyer described a shift from reviewing individual legal questions toward “SEO for legal advice”: writing legal frameworks that clients' agents can discover. The comparison with search optimization is about making advice findable by agents.

The report calls this preparation and supervision “meta-work.” It includes planning tasks for several agents, having one agent review another's results, and building dashboards to track their progress. Its share of the average employee's interactive sessions rose from 10% in fall 2025 to 22% in spring 2026. Those are shares of sessions, not of total working hours.

The researchers estimate these shares by automatically classifying 2,000 one-line chat summaries. The chart shows infrastructure at 29%, security and trust at 27%, developer tooling and product engineering each at 26%, safety research at 25%, other technical roles at 24%, model research at 22%, data science at 19%, sales at 15%, and finance, legal and marketing combined at 12%.

Managing agents has not removed the need for people to coordinate with one another. Employees describe bug investigations that reveal organizational problems, system issues or older technical problems that people must discuss and resolve. Calendar records show the median employee spending 20% more time in meetings than a year earlier. Company growth and a shift toward meeting-heavy roles explain part of that, but technical employees without management roles also spent 16% more time in meetings.

As code writing, bug finding, progress logging and email drafting become more automated, one non-technical employee described the change as “that 20% that was strategic becomes 90% of the job”. That is an interviewee's account of their own work, not a company-wide measure. Sales calls, interviews and discussions across teams still depend on human interaction.

Slack, the company's messaging platform, shows a similar pattern. Agents eventually appeared in nearly half of active public channels. A chart of one large department identifies 237 employees and 7,922 agent threads and conversations. Yet agents mostly appeared in channels dedicated to human–Claude exchanges. Existing human discussions kept the same number of human responses, with agent replies added alongside them. Agents looked more like a team of assistants around one employee than a replacement for communication between employees.

Agents do not automatically know what people have agreed in person. If a meeting ends with a decision but no notes or written update, they lack information needed to adjust their goals and assumptions. Old documents can mislead them too. One employee's name remained on a months-old document about Claude merchandise, so agents kept bringing them requests for t-shirts and mugs long after their responsibilities had changed.

The report argues that work and decisions need records that agents can read. Unrecorded meetings and private messages leave gaps in what agents know. One researcher described this openness as a major change from how most organizations work: leaders also have to give feedback in the open and explain it there.

They also acknowledge that these approaches will not suit every company. Stock ownership is easier to champion when the company's value is rising rapidly. A commitment to find new roles for existing employees can be costly and may make employers more cautious about hiring newcomers. On the other hand, prioritizing internal moves encourages a company to find new tasks and products, rather than focus only on replacing people to cut costs.

These measures cannot prevent AI from disrupting jobs, and they do not directly address slower junior hiring. Protecting existing employees does not help people who were never hired. The report therefore also calls for government policy on retraining and income support alongside action by employers.