Anthropic's October 16, 2026 study of its own employees found that meta-work, meaning the skills, plans and reviews written for agents and the documentation of how work gets done, rose from 10% of the average employee's interactive Claude sessions in fall 2025 to 22% in spring 2026. The researchers also found that interactive Claude sessions amount to around 80% idle time, the agent waiting for human input. Inside the company that builds the tools, AI moved the work toward directing and reviewing, and people's attention became the limit.
It's the part of the job that tells an agent what good looks like.
What did Anthropic find about its own employees' work?
The report, "How employment is changing at Anthropic," comes from the company's economics team and draws on internal usage logs, surveys, calendar data, hiring records and 17 employee interviews. Engineers told the researchers they had become "leaders and managers of agents." One non-technical employee put the shift as "that 20% that was strategic becomes 90% of the job." For some non-technical groups, reading all of their interactive Claude output would take more than half the work week.
Why does your team feel busier after adopting AI?
You brought AI in to take the admin off your plate: the client recaps, the procurement follow-ups, the spec sheet lines nobody enjoys writing. A few months later your designers are reading drafts, fixing a finish code in one, rewording a client email that came out too breezy in another, and re-explaining the project to the tool before each task. If you've felt busier, you've read the pattern correctly, and Anthropic's own staff report the same shape.
That reading and correcting is where your standard lives. A drafted recap that says "brass" where the approved finish is unlacquered brass is the kind of error that turns into a reorder, a delay and an awkward call. The person who catches it is doing the work you hired them for.
Is reviewing AI output the work a creative practice wanted?
For a studio, mostly yes: you started the business to spend your hours on taste and client judgment, and the admin crowded those hours out. When the drafting moves to an agent and the deciding stays with you, the time that remains is the time you wanted.
It pays off only if the review is designed. Left alone, it expands until it fills the week, or it shrinks to a skim and the brass error ships.
How do you keep AI review from taking over the week?
Sort the output before anyone reads it. Anything that leaves the studio carrying a number, a date, a finish or a dimension gets read closely by the person who owns that decision. Internal notes and first drafts get spot-checked. Write the sort down so the team isn't deciding it email by email.
Then turn corrections into instructions. When you fix the same thing twice, write it into a standing instruction the agent reads every time: your finish naming, your vendor terms, the tone you use with a particular client. Anthropic reports its standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later, as the company wrote down how its work should be done. How to Build a Knowledge Base Your AI Agents Can Use covers where those instructions belong.
Expect the first month to feel slower. Set aside time each Friday to go through that week's corrections and decide which ones become instructions, so the same fix doesn't need making a third time.
Related questions
Does AI reduce a team's workload?
In Anthropic's October 2026 study of its own staff, AI shifted work toward directing and reviewing agents. Meta-work rose from 10% to 22% of the average employee's interactive Claude sessions between fall 2025 and spring 2026.
What is meta-work in AI?
In Anthropic's study, meta-work means setting up and checking agents: writing their skills, plans and reviews, building monitors for agent pipelines, and documenting how work is done.
Why are AI agents idle so much of the time?
Anthropic found its employees' interactive Claude sessions amount to around 80% idle time, the agent waiting for human input. The person's attention, more than the agent's speed, set the pace of the work.
How much time does reviewing AI output take?
Anthropic reports that for some of its non-technical teams, reading all of their interactive Claude output would take more than half the work week, which is why deciding what gets read closely matters.
What this changes for a studio
Anthropic's data shows more of each person's work going to directing and reviewing agents once those agents arrived. In a studio, that review is where your designers apply your taste and client judgment to every draft. The AI Productivity Paradox: Where the Saved Hours Go follows the same hours from the other side.
This week, write a one-page review rule for your studio: which outputs get read closely and by whom, which get spot-checked, and where corrections get written down. Then pick the correction you made most often last month and turn it into a standing instruction. Neither depends on running an agent; a new junior designer would need both.
Working together
Flow State Found works with a limited number of businesses to make their best work their baseline. AI was supposed to give your designers time back, and they're spending it fixing drafts. We design the review and the standing instructions around your standard, so the checking gets lighter as that standard gets written down. We take on limited engagements, so it starts with a conversation.
Start a conversation