The Bottleneck Moved

Anthropic just published a report on how AI has changed work inside its own company. It's the best data we have on what happens when an organization goes all in. By August, the AI agents its employees directed were logging six hours of work for every hour a human worked.

The headlines will focus on engineers. Most SMB owners don’t care about that. They run accounting firms, agencies, contractors and distributors, often without a single engineer. So I read the report looking for something else: what happened to the lawyers, salespeople, recruiters, assistants and AP clerks. That's where the clues for small business are.

As you can see, coding experienced the most dramatic increase in productivity, but output rose on the non-technical side too. Lawyers processed more contracts. Accounts payable moved more invoices. Salespeople closed deals with more clients after they adopted agents.

The details are what interest me.

And the volume: coding-agent sessions among non-technical employees went from 2 a month in August 2025 to 90 a month in May 2026. People who had never written code were building their own tools.

Anthropic tripled in size in a year, and its administrative functions barely grew. The report calls the company "clerical-light." In most small businesses, admin grows right alongside revenue: another coordinator, another AP clerk, another assistant. That relationship is breaking. Before you post your next admin job, ask whether someone on your team could build the agent first.

The most useful rule in the report came from a salesperson: when you do something a second time, figure out how to make Claude do the third. I call these people Idea Programmers. They shape AI through instructions and feedback instead of code. Anthropic's salespeople, lawyers and assistants are Idea Programmers. They know the work, so they're the right people to automate it.

For a small business, this is the real opportunity. You already have the builders: your best ops person, your best salesperson, your bookkeeper. Give them protected time each week to build.

Goldratt taught us to find the constraint and manage everything around it. At Anthropic the constraint is now human attention. Agent sessions spend about 80% of their time waiting on a person. Reading everything the agents produce would take non-technical employees more than half their work week. As one engineer put it, people are "bottlenecking the agents."

So employees started doing what the report calls meta-work: writing instructions, building dashboards, setting one agent to review another. That grew from 10% of their agent sessions to 22% in six months.

In most small businesses the constraint is already the owner. Every decision runs through you. Add agents and the line outside your door gets longer. The fix is the one it's always been: push decisions down, build review into the system, and only look at what needs your judgment. The difference is speed. You'll hit this wall in months.

Standing written instructions for agents at Anthropic went from a few hundred to 8,500 in a year. Employees are pushed to make decisions in the open, in Slack threads and meeting notes, because a decision made in a hallway is invisible to the agents. One employee kept getting agent requests for T-shirts and mugs because their name was still on an old swag document.

That's a problem for small businesses, which run on what's in the owner's head and what got said in the truck. The firms that get the most out of agents will be the ones that write things down: decisions, processes, customer context. I call this knowledge programming. Your SOPs are about to become instructions for your agents, so they'd better be current.

Hiring managers at Anthropic now comment far more on a candidate's autonomy and judgment, and less on domain expertise and execution. Comments on how candidates use AI tripled for technical roles and quadrupled for non-technical ones. Jobs are broadening. One employee described it this way: "that 20% that was strategic becomes 90% of the job."

For owners, that changes the job description. Look for people who can decide what the task should be and check the agent's work. The agents bring a lot of the execution.

The hard part is juniors. If agents do entry-level work, where do your future senior people come from? The report doesn't solve that, and most small businesses won't either. Someone still has to train the next generation of judgment, and in a small firm that someone is you.

Anthropic built an early warning system to spot jobs where the work is shifting and move people into new roles before those jobs shrink. Employees were 50% more likely to switch departments each month in 2026 than in 2025. Whole teams moved, like the copywriters who now write instructions for Claude.

Small businesses can do this more easily than anyone, because the owner knows everyone. When a role starts to shrink, look at who in it has been building. Move your best people into the new work before you hire from outside. The AP clerk who builds the invoice agent is your new operations analyst.

Anthropic brings in two to four times as much revenue per employee as Google and Meta did at the same age. That’s the AI difference. Google and Meta already made multiples of revenue per employee than normal firms and Anthropic has moved the bar further.

But alas, we aren’t Anthropic so for regular owners, the better gauge is labor efficiency: gross margin divided by what you spend on labor, the measure Greg Crabtree teaches in Simple Numbers. If agents are working, that number should climb. If it doesn't, you're adding tools without changing how the work gets done.

Anthropic's engineers got the headlines. Its lawyers, salespeople and assistants are the preview for everyone else. They spend less time doing the work and more time deciding what the agents should do, checking it, and talking to people. That’s coming to all our businesses eventually. I suggest getting prepared now.

Sources