Anthropic’s AI agents now log six times as many hours as the people who direct them. The company reported the finding on Oct. 16, in a study of its own staff. For AI adoption in IT services, a business that still sells hours, it is the clearest warning yet.
New here? I publish one agentic AI governance post every weekday. Subscribe to the blog and it lands in your inbox the moment it goes live.
Three days ago, I argued in a post here that skills move and adapters win. The new study shows what that move looks like inside one company. Junior hiring shrank while output rose. Then human attention, not agent capacity, became the scarce resource. Each shift hits firms that bill by the hour.
What Anthropic measured
Four researchers at the Anthropic Institute wrote the report on employment at Anthropic. They used company records on hiring, agent use, output, surveys and calendars. They also talked to 17 staff. So this is usage data, not a vendor survey.
The headline numbers move fast. Agent hours passed human hours for the first time in March 2026. By August, the agents each employee directed worked six hours for every one of theirs. Engineers now commit code at five times the pace of six months ago.
The gains went past code, too. Lawyers got through more contracts, and sales reps closed more deals once they took up agents. One sales rep built a pipeline dashboard in three or four hours. Before, the rep said, that job was out of reach. Quality held, at least by one test. Engineers undid agent code no more often than code they wrote by hand.
The pyramid loses its base
The hiring data is the part an IT firm’s leaders should read twice. Since early 2025, junior staff have dropped from 21 percent to 14 percent of Anthropic’s technical teams. Junior technical hires fell faster, from 27 percent of hires to 9 percent. About a third of machine learning managers said they now hire less because of AI.
Meanwhile, revenue per head runs two to four times what Meta and Google earned at the same age. That is what a firm looks like when agents do the work juniors used to do.
Now map that onto a classic IT firm. Its pyramid puts many junior engineers under a few seniors, and the juniors carry the billable hours. Once agents absorb that work, the base thins and the firm starts to look like a diamond. Hiring shifts too. Anthropic’s hiring managers now weigh judgment and the drive to work alone over narrow skill.
Attention is the new bottleneck
Speed made a new problem. Anthropic found that agents spend about 80 percent of session time waiting for a human. Even for staff who do not code, just reading what their agents write would take over half the work week. One engineer described facing “an army of Claudes, with every one asking for my attention.”
So employees started building the tools to manage the agents. This meta-work rose from 10 percent to 22 percent of sessions. It means writing specs, having one agent check another and watching pipelines. Standing written instructions for agents grew from a few hundred to 8,500 in a year.
For AI adoption in IT services, this is the governance question hiding in the speed story. A client does not pay for agent output. Instead, the client pays for output someone reviewed, signed and will answer for. So the real limit on delivery is review time, and few firms staff for it.
More on agentic AI and IT jobs
- Skills Move: Why typists, bookkeepers and animators show that adapters keep working when a new tool arrives.
- Operating Model: How large enterprises organize their agentic AI teams around a hub that sets the rules.
- Unit of Account: Why I track agent spend as the cost per completed task, not per token.
- Delivery Teams: My five-stage guide for delivery managers moving teams from prompted to autonomous.
How I would run adoption at an IT firm
I think the firms that exit first will regret it. My work in regulated loan origination taught me that software cannot take the blame. Someone still signs. I spent 20 years in Fortune 100 consulting, and the staffing pyramid was always the quiet engine of the margin. So here is how I would run AI adoption in IT services if I sat in a delivery leader’s chair.
First, I would price the outcome, not the hour. When agents work six hours for every human hour, billing time rewards the wrong thing. Then I would staff review as delivery. I would budget the meta-work, and I would name the person who signs each deliverable before the agents start.
Next, I would make decisions visible. Anthropic found that agents miss context when people decide things in unlogged meetings. On a client engagement, the decision belongs in the written record agents read, such as an intent.md file.
Finally, I would keep hiring juniors, but for judgment, and train them as reviewers first. A firm that stops hiring juniors today has no seniors in ten years. I would also run my own early warning system on roles and redeploy.
Intent in, outcomes out
Instructions in, results out was IT; intent in, outcomes out is agentic AI. For three decades, IT firms sold instructions turned into hours. AI adoption in IT services now asks them to sell judgment. That means people who can review an agent’s work and stand behind it.
Anthropic’s data shows the work does not disappear. It moves up, toward review, judgment and the client conversation. The firms that train for that move will keep their clients.
If you run delivery at an IT firm, how much of your team’s week already goes to reviewing agent output? Tell me in the comments.
Was this post useful? One click tells me what to write next.
Subscribe to my LinkedIn Newsletter
- Agentic AI Architect: control design for enterprise agents.
- Agentic AI Case Studies: deployment evidence, one teardown at a time.
- Agentic AI P&L: cost, payback, and risk in a CFO’s voice.
- Agentic AI SDLC: the lifecycle that builds the agents.
- Subscribe to the Blog: I publish a new post every weekday on one agentic AI governance question. Enter your email and each post arrives in your inbox the moment it goes live.
© Dr. Harish Kotadia, Ph.D., All Rights Reserved, 2026
Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.
Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.
Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.