When AI Does the Work, Verification Becomes the Job
On October 16, 2026, Anthropic’s Economics team published a report called “How employment is changing at Anthropic.” It is the most detailed primary account I have read of what happens inside a company that has reorganised itself around AI agents. The authors are Nathan Wilmers, Regina Catipon, Maxim Massenkoff, and Peter McCrory, and the data covers employment records, AI use telemetry, output metrics, employee surveys, internal communication patterns, calendar data, and hiring records, supplemented by 17 employee interviews.
The headline number is that by August 2026, interactive AI agent hours at Anthropic outnumbered human employee hours by six to one. Agent hours first surpassed human hours in March 2026 and the ratio has been widening since.
For HR leaders, the report carries a specific implication that I have not seen surface in the broader conversation about AI adoption. When agents produce the work, the human’s job shifts from doing to verifying. And verification has a cost that most job descriptions and most performance frameworks were not built to measure.
What the report actually found
The report documents several shifts happening simultaneously inside Anthropic.
Junior technical hiring has contracted. The share of junior-level employees in technical departments fell from 21% to 14% since early 2025, and the proportion of junior technical hires fell from 27% to 9%. Postings for junior roles dropped sharply after December 2025. A majority of managers surveyed in April 2026 said they were hiring for a different set of skills because of AI. Around a third of machine learning managers and a fifth of software engineering managers reported hiring less due to AI.
Output rose across functions. Code commits per engineer rose roughly fivefold over six months. Sales representatives who adopted agents closed deals with more clients. Lawyers processed more contracts. Accounts payable moved through more invoices. The report tracked 192 engineers who adopted coding agents against 74 who did not, and found that code output rose sharply after adoption while code quality held steady by several measures including rates of lines subsequently removed or reverted.
Hiring criteria changed. Using automated analysis of hiring manager interview scorecards, the report found that emphasis on autonomy and judgment increased substantially for technical candidates between 2024 and 2026, while discussion of domain expertise shrank as a share of assessment. For non-technical candidates, emphasis on AI tool usage rose from roughly 7% of assessments to around 28%, and emphasis on customer and mission orientation increased while discussion of execution skills fell.
The bottleneck the report names
Here is the part I think HR leaders should read most carefully.
The report states that human attention has become the bottleneck for agent output. Interactive Claude sessions spend roughly 80% of their wall-clock time idle, waiting for human input. The authors converted agent output tokens to word counts and estimated reading time at standard speeds. Even for non-technical employees, the volume of Claude output from interactive sessions would take more than half the work week to read. For technical functions it is higher.
The report describes employees responding by building what the authors call “meta-work”: tools to monitor agent pipelines, standing instructions for agents, dashboards to surface problems. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. The share of meta-work in the average employee’s interactive Claude sessions rose from 10% in autumn 2025 to 22% in spring 2026.
One engineer described the experience as babysitting: agents bring requests with equal urgency, and the human has to triage which ones actually matter. Another wrote a performance review for their Claude team. A lawyer shifted from reviewing legal questions to doing “SEO for legal advice,” writing frameworks that clients’ agents would discover.
What this means for job design
The Anthropic report is a case study of one company, and the authors are careful about generalising from it. But the pattern it describes aligns with a framework from organisational psychology that predates AI by decades.
In 1976, Hackman and Oldham published their model of job characteristics, identifying five dimensions that shape how people experience work: skill variety, task identity, task significance, autonomy, and feedback. The model has been cited thousands of times and remains a standard reference in job design. Their paper is available at https://doi.org/10.1016/0030-5073(76)90016-7.
What the Anthropic data suggests is that AI delegation reshuffles which of those dimensions dominates a role. When an agent handles implementation, the human’s skill variety narrows toward judgment and triage. Task identity fragments, because the person who initiates the work and the person who reviews it are increasingly the same person operating at different stages rather than owning the full cycle. Autonomy stays high, but the object of autonomy shifts from “how should I do this?” to “is this worth doing, and did the agent do it correctly?”
This is where I would place a caveat. I am reading the Anthropic findings through the lens of a job design model, and that reading is mine. The report itself stays with the employment and productivity data. I am extending it toward a framework the authors leave uninvoked, because the connection gives HR teams a concrete vocabulary for the changes they are already seeing.
If verification becomes the job, several things follow that most HR functions are not yet set up to handle.
Performance criteria need to shift. When output is measured by lines written or deals closed, the human doing the verification work disappears from the metric. The report notes that 20% of surveyed employees described their role as shifting from doing work to managing Claude. Performance management built around throughput of finished work will undervalue the person who catches the agent’s mistakes and redirect its effort.
Selection criteria need to shift with them. Anthropic’s own hiring data shows this happening at the frontier: managers are weighting judgment and autonomy over execution and domain expertise. For roles where AI handles the production, the selection question becomes whether a candidate can tell whether the work was done correctly and intervene when it was not. Doing the work matters less than auditing it.
Job descriptions need to acknowledge the verification burden. The report’s estimate that reading agent output would consume over half a work week for non-technical staff comes from measured output volume from agents already running inside one company. If your organisation is moving agents into production workflows, someone is already spending time checking what they produce, and that time is probably not in their job description.
The reallocation question
The report also describes Anthropic standing up an “early warning system” to track jobs facing AI-driven transformation, with the goal of reallocating affected employees internally rather than turning to the external market for every new task. Employees who had been at the company for at least a year were 50% more likely to move across departments each month in 2026 than in 2025.
This is the part of the report with the most direct relevance to workforce planning. The question it raises is whether internal mobility systems are ready for a pace of reallocation that tracks AI capability releases rather than annual planning cycles. The report notes that hiring managers flagged a decline in demand for junior technical employees in January 2026, and by April the shift was visible in hiring data. That is a quarter, not a planning cycle.
What I would take from this
The Anthropic report is one company’s internal data, published by researchers with access to employment records and telemetry that most organisations will not have. What it reports is what happened there, in the data they collected.
What happened is that agents produced the work, and the humans spent their time checking it, organising it, and building tools to manage the flow. The job moved from production to verification, and the cost of verification, in time and attention, became the constraint on output.
For HR leaders, the practical question is whether your performance frameworks, your selection criteria, and your job descriptions have caught up with that shift. Most have not.
Sources: Wilmers, N., Catipon, R., Massenkoff, M., & McCrory, P. (2026). How employment is changing at Anthropic. The Anthropic Institute. https://www.anthropic.com/research/employment-at-anthropic · Hitzig, Z., Massenkoff, M., Lyubich, E., et al. (2026). Agentic Coding and Persistent Returns to Expertise. Anthropic. https://www.anthropic.com/research/claude-code-expertise · Huang, S., Seethor, B., Durmus, E., et al. (2025). How AI Is Transforming Work at Anthropic. Anthropic. https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic · Hackman, J. R., & Oldham, G. R. (1976). Motivation through the design of work: Test of a theory. Organizational Behavior and Human Performance, 16(2), 250-279. https://doi.org/10.1016/0030-5073(76)90016-7