Anthropic's AI Agents Spend 80% of Their Time Waiting for a Human

Last Friday, Anthropic, the company that makes Claude, published a study of its own employees. It draws on internal data on hiring, output, calendars and agent use, which makes it the most detailed look yet at a company that runs on AI agents.

By August, the agents its employees ran were working six hours for every hour their human handlers worked. Salespeople who adopted agents closed deals with more clients. One account executive said automations had taken over about 20% of their tasks. Another said they could now talk to 30% more of their customers.

The hiring changed too. Junior roles fell from 27% of technical hires to 9%. At five years old, Anthropic brings in two to four times as much revenue per employee as Meta and Google did at the same age.

That is the part everyone shared. The finding a CMO can use is further in this newsletter.

In interactive sessions, Claude spends about 80% of its time idle, waiting for a person to reply. Even for non-technical teams, reading everything their agents produce would take more than half the work week. One engineer described the job as facing "an army of Claudes, with every one asking for my attention."

So employees started spending more of their time managing agents instead of doing the work. That kind of work grew from 10% of employee sessions with Claude in autumn 2025 to 22% by spring 2026. Written standing instructions for agents went from a few hundred in July 2025 to 8,500 a year later.

The habit that helped was refusing to do the same thing twice. As one of those account executives put it: "anytime you do anything a second time, you're kind of like, okay, how can I use Claude to make it so that I don't have to do it a third time?"

My favourite detail in the study is about t-shirts. One employee's name was on a months-old document about Claude merchandise. Agents kept sending them requests for t-shirts and mugs long after their responsibilities had changed.

The study found that when a decision is made in a meeting and nobody writes it down, the agents never hear about it. They keep working from the old version. Anthropic's response was a default of deciding in the open, in writing, where agents can read it. In the report's words, "The need to feed AI context requires observability of human work and decisions."

Your target accounts changed in a quarterly review, and a campaign was paused on a call. The pricing message was rewritten in a Slack thread. Your people know about those changes, but your agents are still reading last year's persona deck.

The study shows where the job is going. Hiring managers at Anthropic now mention a candidate's use of AI four times as often for non-technical roles as they did in 2024. They talk more about judgement and autonomy, and less about execution skills. One employee summed it up:

Marketing already has its own example inside the study. A team that wrote the product copy inside Claude built what the study calls a context graph, which guided agents to write that copy well. The team then moved into writing instructions and tests for Claude. The study sums up the change: "Writers who used language to drive human behavior are now using it to drive Claude behavior."

In a Fortune column dated 22 Oct, the CEO of a sales software company wrote that his customer EliseAI plans to double its sales team over the next 12 months. Its VP of enterprise sales, Ben Rey, says AE productivity has tripled. "That productivity has us hiring faster than ever," he said. Those are the company's own numbers.

Most marketing teams aren't there yet. In BCG's survey of 300 CMOs this year, only 20% of B2B CMOs said agentic marketing was already having a measurable impact on revenue.

In almost every GTM stack I look at, teams bought agents but kept deciding things where agents can't see them. The CMO's job is moving from approving the work to writing down the decisions the work runs on.

BCG's exhibit shows what the shift looks like. Today, marketing is organised by channel and domain: web and SEO, email, paid media and social. Large human teams with overlapping ownership sit between the product owner and the end customer, and the coordination between them slows execution.

BCG's near-future picture is much leaner. Full-stack marketers, backed by deep discipline experts and AI agents, own a whole solution area from end to end. Product owners sit closer to the end customer, and there is less coordination overhead.

Takeaway: that structure only works if the agents can read what the team has decided. A full-stack marketer can't own a solution area if the targeting, messaging, pricing and budget decisions behind it live in meetings and DMs. The CRO has the same problem at the handoff between marketing and sales, where lead owners, stages and next steps change every week. The organisation BCG describes runs on written decisions, so the teams that start writing them down now will be ready to restructure when agents take on more of the work.

This takes 30 minutes on a Friday, with your marketing ops lead and whoever runs your agents.

Adapt this decision note and standing instruction for your team.

The sweep keeps your agents working from current decisions. It won't fix a CRM with the wrong owner on half your accounts, and it doesn't decide who should be allowed to see what. It also won't tell you whether your agents create pipeline.

To check that, split your accounts at random. Compare the half the agents worked against the half they didn't. That is a rough read, not incrementality testing with a proper statistical design, statistically derived multi-touch attribution or marketing mix modelling.

Anthropic's numbers come from one AI company studying itself, and the authors note that early adopters of agents may already have been ramping up.

Anthropic's agents got faster with every model release. The people around them became the limit, so decisions needed to be written where agents could read them. One researcher called it a big shift "because leaders have to defend feedback out in the open."

That is the part of the CMO's job that is growing. Agents will do more of the execution every quarter, but they can't know what you decided on Tuesday unless you write it down.

Forward this to whoever runs your agents, and hold your first decision sweep this Friday.

P.S. In July we launched RevSure's context layer, one connected record of your GTM data. Last week we launched Digital Workers that sit on top of it. We built them in that order on purpose: a digital worker is only as good as the context it reads. See how it works.