AI agents and productivity at work: lessons from Anthropic

On October 16, 2026, Anthropic published a report on how work is changing inside its own teams. Since March 2026, the AI agents its employees direct have logged more working hours than the employees themselves, and by August the ratio was six to one. For a business, the most useful lesson is not that headline number. It is what still holds the agents back: they spend most of their time waiting for a human decision, and productivity depends first on how the work around them is organized.

What exactly did Anthropic publish?

The company behind Claude released two studies on the same day, introduced in an overview post. The first, "How employment is changing at Anthropic", analyzes internal usage data and 17 employee interviews. The second, "The AI expectations gap", compares its staff's expectations with those of nearly 11,000 US adults surveyed in August by Morning Consult.

The figure that went viral comes from the first study, and it is worth understanding how it is built. "Agent hours" are measured from session telemetry: the time the model spends reasoning or using a tool, excluding gaps of more than 30 minutes. "Human hours" are simply headcount multiplied by 40 hours a week. So this is not six times more useful work. It is six times more machine time than contracted human time.

In other words, it measures how intensively agents are used, not productivity. Anthropic says as much when it calls code output "a notoriously imperfect measure".

What results does Anthropic report?

The report shares several indicators, all drawn from the company's own data:

On the sales side, the report quotes an account executive who says automations replaced about 20% of their previous tasks and let them talk to 30% more customers. That is one person's account, not an average.

Our reading: these numbers describe a company that has made AI agents an everyday work tool across functions, not only for developers. They also show something we see in the field: the more work you hand to agents, the more human work shifts toward preparation, review and coordination.

Why do AI agents spend 80% of their time waiting?

This is the most useful data point for a business leader. According to Anthropic, Claude sessions spend around 80% of their time idly waiting for human input. The report draws a clear conclusion: human attention has become the bottleneck.

In practice, an AI agent (a program that uses a language model to chain actions together: read a document, query a tool, draft something, then hand back control) moves fast as long as it knows what to do. It stops as soon as it needs an approval, a clarification or an access right. If nobody answers, it waits.

Anthropic's staff responded by building what the report calls meta-work: writing specs for agents, having one agent review another's output, setting up monitoring dashboards, and documenting processes so a machine can follow them. Engineers describe themselves as "managers of Claudes".

For a smaller company, the takeaway is straightforward. The gain does not come from the agent alone but from three things around it:

• Written decision rules. What the agent may do on its own, what needs approval, and who approves.
• Batched checkpoints. Reviewing twenty drafts at once at the end of the morning instead of being interrupted twenty times.
• Accessible information. The report stresses keeping work visible in channels agents can read. A procedure that only lives in one employee's head cannot be followed by an agent.

Do these results apply to smaller businesses?

Only partly, and Anthropic acknowledges it. The report notes that many parts of its practices "cannot be replicated across the economy". The company has tripled in size in a year, builds the very models it uses, and has a deeply technical culture. The report also flags method issues: employees who adopted agents were already on steeper output trajectories, and code quality is hard to assess because AI errors may be harder to spot than human ones.

The expectations gap is striking too. According to the second study, Anthropic staff estimate on average that AI will be able to do 75% of their current work hours by 2030, compared with 28% for the US workers surveyed.

Most companies start from a very different place. In France, for example, a Banque de France study published on September 3, 2026 found that more than two thirds of firms with at least 20 employees reported using generative AI tools in early 2026. But most rely on general-purpose tools for support functions such as administration and sales. More specialized uses, built into core processes, remain concentrated in knowledge-intensive sectors.

That is where the gap lies: most companies use a chat assistant, very few have handed an entire process to an agent. And it is the second kind of use that produces Anthropic's numbers.

Should you expect fewer junior hires?

This is the point that drew the most reactions. The junior share of Anthropic's technical hires fell from 27% to 9%, and their share of technical headcount from 21% to 14% since early 2025. The report does not describe layoffs: the company kept hiring heavily, just differently.

For a smaller business, we would not turn this into a rule. The tasks agents absorb best (first drafts, simple research, repetitive answers) are also the ones that used to train newcomers. Cutting those roles without rethinking training means losing tomorrow's experienced staff. The better question is: which low-value tasks can you take off your teams' plate, and what do you give them instead?

Where should you start to get this kind of gain?

There is no need to aim for six agents per employee. The fastest gains come from one well-chosen process with real volume and clear rules.

Customer support is often the best starting point: questions repeat, documentation exists, and an agent can answer on its own while handing off to a human when a request falls outside its scope. At Peps Digital, the assistant we built on their product documentation now handles 80% of support requests without human intervention, escalating the rest to the team. We describe this approach on our customer support agents page.

Two mistakes come up often. The first is launching a pilot without a baseline: if you do not measure before, you cannot prove a gain after. The second is not planning for human review, which recreates the bottleneck Anthropic describes. We explain how to avoid both in our article on AI pilots that never reach production and in our guide to evaluating an agent in production.

What this means for your business

Measure waiting time, not just time saved. If an agent produces quickly but approvals pile up, the real gain is small. Track the delay between the agent's proposal and the human decision.

Write the rules before you deploy. List what the agent can do alone, what requires approval and who approves. This framing work, more than the choice of model, is what makes the difference.

Start from a process, not a tool. Pick a high-volume flow with clear rules (support, request qualification, document processing), set a baseline, and compare after a few weeks.

Frequently asked questions

Do Anthropic's AI agents really do six times more work than its employees?

Not strictly. The report compares agent activity time, measured from logs, with employees' contracted working hours. It is a usage indicator, not a measure of useful output or quality.

What is the difference between an AI assistant and an AI agent?

An assistant answers a question in a conversation. An agent chains several actions on its own (checking a tool, reading a document, drafting, updating software) to complete a task, then hands back control for approval.

Can a smaller company get comparable gains?

Not at Anthropic's scale, but on a targeted process, yes. The clearest gains show up on repetitive tasks backed by existing documentation, provided human review is organized from day one.

Will AI agents replace employees?

Anthropic's report mainly shows work shifting: fewer execution tasks, more directing, reviewing and coordinating. It also points to fewer junior hires, a trend worth planning for rather than reacting to.