The more you hand work to AI, the more you want it stronger, and the less you fear replacement

On June 26, 2026, Anthropic released a report called Cadences. Using real usage data from thousands of Claude users plus a survey of about 9,700 people, it looks at how AI enters a day of work and life. This piece skips the big "will AI replace people" debate and picks up something more specific and more counterintuitive: the people who hand more of their work to AI are less likely to see a stronger AI as a threat. They actually want it stronger, and they fear replacement less. I will unpack where this pattern comes from, what the data can and cannot say, and finish with a method you can follow.

They changed how they look this time

The Anthropic Economic Index is a series that uses real Claude usage data to watch how AI enters economic activity. This is the sixth report, titled Cadences, which you can read as "rhythms." Compared with earlier reports, three method changes are worth knowing first, because they set what this report can see.

1. Sampling down to the hour. Earlier reports looked at seven-day aggregates. This one sees usage across different times within a single day. It is the first time AI usage can be lined up against the rhythm of a person's day.
2. Classifying the outputs. A new system sorts conversation outputs into 30-plus categories to see what Claude actually helps people produce. The result: 93% of conversations produce an identifiable output, most often explanations, documents, and guides.
3. Connecting a survey for the first time. A survey of about 9,700 people is linked, using privacy-preserving methods, to pair "what users think" with "how they actually use it."

In plain terms: before, we only knew how much people used it. Now we start to see what time, producing what, and what they think about it.

AI is already woven into the rhythm of a day

First, the usage behavior the report observed. This layer is actual data, the facts. Across a day, what people ask AI follows the rhythm of life.

Now weekday versus weekend. Personal-use conversations are about 35% of the total on weekdays and rise to about 50% on weekends; weekend content shifts from business correspondence and marketing toward emotional support and health questions. One counterintuitive detail: startup-related conversations run higher on Saturdays and Sundays, but job hunting and sending resumes drop on weekends. In plain terms, weekends are more the time for daydreaming about being your own boss, and applying for jobs gets pushed to the back.

Put these pictures together and they say one thing: AI is already woven into the rhythm of work and life, part of the day. Asking how to sleep before dawn, what to cook at dusk, how to file taxes at the deadline. It shows up in the gaps of daily life.

First, be clear on who this report is looking at

Before we get to "what people think," we need to be clear on who this report's "people" are. This step goes first because it sets how far every later number can generalize.

This is not a survey of the whole labor market. It is a Claude-user sample, and the occupational mix is clearly skewed. Computer and mathematical roles make up about 30% of respondents, but only about 4% of total US employment; management makes up 23% of respondents versus 7% of US employment. Physical occupations are badly underrepresented, and women are only about 12% of the sample.

Once you have seen this chart, you will add a line to every number that follows: this is how a group of mostly knowledge-worker, heavy AI users think, not how all workers think.

The real point: the more people delegate, the more optimistic they are

This next part comes from the survey, so it is respondents' stated expectation, a belief, not a fact. Keep that layer in mind. The report shows a positive correlation: the more someone uses Claude in automated ways, the more optimistic they are about how AI will affect their own job prospects next year. Specifically, the higher the automation share, the more upbeat they are across six dimensions: pay, job security, ability to find new work, sense of meaning, autonomy, and human interaction.

Here is the productivity gain these heavy delegators report for themselves.

The perception numbers line up too: 68% feel they learn more with AI, and 57% feel AI makes their skills more valuable in the market. Looking ahead, more than a third expect AI to handle most or nearly all of their work tasks within 12 months. But one contrast is worth a look. Asked about job loss, only 10% think they themselves are likely or very likely to lose their job in the next 12 months; the same group worries far more about junior colleagues, with about a third rating a junior's job-loss risk above 60%.

Here is how I read this part. This group has already changed the question. They are not asking "will AI take my job." They are testing "what part of my work can AI take next." Being optimistic about themselves while worried about junior colleagues also reveals something: the difference may not be about title, but about whether you can break work into pieces and hand them off.

So, what this report cannot be used to claim

Chart 3 already gave the warning; here is the full set of limits.

• People willing to fill out a survey and who are heavy users may differ from everyone else to begin with. That is selection bias.
• The "optimism correlates with automation" link is likely partly driven by selection: people already bullish on AI are more willing to hand work off. Correlation is not causation, and the report itself flags this caveat.
• "Self-reported learning more" does not rule out actual skill erosion.
• The report puts it bluntly: they cannot conclusively identify the jobs of the people making requests.

How I see it

Pulling the first half together, here is my read. On the surface this report is about AI usage rhythms; what it actually reveals is a capability gap. When AI gets stronger, the upside does not land evenly on everyone. It flows first to the people who have already broken their work into pieces they can hand off.

The logic is plain. If work can be handed off, then when AI gets stronger you gain an extra pair of hands and take on one more stretch of the process. If work cannot be handed off, all tangled together and living only in your head, then no matter how strong AI gets it can barely reach in to help, and all you feel is anxiety.

That is why "delegators want AI stronger" matters so much for companies of one and knowledge workers. It points to a skill you can practice: breaking work into pieces you can hand off, and knowing which piece needs review and who does the reviewing. This is not about blind optimism. That skill decides which side you stand on next time AI gets stronger, the benefiting side or the anxious side.

One more line from the same report belongs here: experienced workers think AI can take on about 10 percentage points fewer of their tasks than first-year workers do, and they stress that what AI struggles to replace is judgment, context awareness, and situational reasoning. The report uses a conditional: if the human stays on the highest-value tasks, the pattern looks more labor-augmenting than labor-displacing. The key is that "if." Only those who stay on the highest-value tasks capture the gain.

The strategy I recommend: sort tasks into three tiers

Do not stop at a vague "so I should use more AI." The approach I recommend is one action you can do today: take this week's work and sort every item into three tiers.

Tier one: tasks I only trust AI to draft

For these, AI starts it, gives a draft, expands ideas, but you always revise line by line. Things like a first pass of outward copy, the skeleton of a long piece, the draft of an important email. You hand off the output speed and keep the final shape.

Tier two: tasks I let AI run end to end, but I review at the finish

For these, AI runs the whole thing and you do not watch every step, you only check the result at the end. Things like turning a batch of transcripts into a fixed format, pulling data into a table, batch-renaming files by your rules. You hand off the whole process and keep the review gate. This tier captures the most upside as AI gets stronger, because the stronger AI is, the more tasks you can safely put here.

Tier three: tasks that still have to be my judgment

Do not hand these off yet, usually for one of four reasons: context (only you know the backstory), relationships (judgment about people, trust, tact), responsibility (you carry the fallout), or missing data (AI does not have enough to make this call). Things like whether to take on a partnership, how to reply to a sensitive message, whether to spend a sum of money.

Turning your workflow into a skill, so what can be handed off actually gets handed off

Once you have sorted the tiers, you notice something: for tier-two tasks that AI can finish and you review, if you re-explain the rules and steps to AI every single time, it gets tiring and the handoff is never clean. What truly makes it "instantly handed off" is writing that workflow down as a skill. In plain terms, a skill is a clear work manual: how this task is done, in how many steps, the judgment standard at each step, and when to stop and ask you. You write it once, and afterward the AI agent reads it and just does it, without you starting from scratch each time.

Once the three tiers are sorted you get two things. One is a clear view of which work can already be handed off and which cannot. The other, more important: next time AI gets stronger, you only ask one question, which tier-one tasks can move up to tier two, and which tier-three ones can move up once the data is filled in. That question is how you actually land the upside of a stronger AI in your own hands.

If you are building a company of one: the parties you delegate to are more than just AI and you

Those three tiers are all about how "I and AI" divide the work. If you are starting up, building a company of one, you can widen the set of parties you delegate to. When something comes in, I ask three questions in order:

1. Can this be handed to AI? If yes, hand it to AI.
2. What AI cannot do, can it be handed to someone else? If yes, outsource it or find a partner.
3. Only what neither AI nor anyone else can do is left for me to do myself.

Follow that order and what is left in your hands is the thing AI cannot replace and no one else can take over. That is usually the true core of your company, the highest-value task the report talks about. A company of one is about narrowing yourself down to the one thing neither others nor AI can do, and finding ways to hand everything else off, one layer at a time.

How to start: three steps you can do this week

With the framework covered, here is the smallest starting routine. No need to wait for free time, no need to learn a tool first.

1. List 10 things. Open up the work you actually did this week and jot down 10 items. They do not need to be tidy; a plain running list is fine.
2. Mark them 1, 2, 3. Give each a number: 1 for those you only trust AI to draft, 2 for those you let AI finish but you review, 3 for those that still need your own judgment. The few you get stuck on are usually where your sense of AI's boundary is blurriest, which is normal; just mark a gut number.
3. Move only one thing, and write it down once it runs. Pick one item from the 2s that you are most confident about, and this week actually hand it to AI to run end to end, doing only the final review yourself. Once it runs smoothly, jot the steps into a one-page note, which is the seed of a skill for this task. Next time the same task comes up, AI reads it and takes over, and you do not explain from scratch.

After that, keep one habit: every time you see news that AI got stronger, look back at this list and ask, which tier-one task can move up to tier two, and which tier-three one can move up once the data is filled in. That question is how you land each round of a stronger AI in your own hands.

Further reading

This piece is about breaking work up and handing it to AI. For how it gets finished and closed out after you hand it off, I have written a few pieces you can read next.

• Loop engineering, explained through my workflow for writing an article: after you hand something off once, how to turn it into a loop that finishes itself. This one explains loops, and picks up right at the tier-two review gate.
• When I started to understand loops, I designed three workflows as loops: three loops I run myself, turning repeated handoffs into a system.
• I used an AI agent workflow to build a program-based automation workflow: turning a one-off handoff into something that runs itself.
• Docs are the system: how non-engineers design an agent framework: how skills, written bit by bit, grow into your whole system.
• A decision ladder that non-programmers can use too: facing a pile of tasks, how to decide which to hand to AI and which to carry yourself.
• Where you stand in the AI world: what the capability gap between people looks like as AI gets stronger.