Should You Hand AI the Whole Task? The People Who Know It Best Don't
The intuitive theory of getting good with AI goes like this. At first you check everything, because you do not trust it. Then you learn what it can handle, you hand over more, and eventually you are delegating whole jobs while you get on with something else.
Anthropic sampled a million conversations and found the opposite pattern. The people who had been using the tool longest were handing over less.
Two ways to use the same chat box
The research separates how you work with a model into modes, and two of them matter here.
The first is Anthropic’s term for a conversation where the human “delegates complete task execution to AI with minimal interaction.” You describe the job, the model does the whole thing, and you take what comes back. Most people’s mental image of using AI is this mode. use: you ask for “an agenda for Thursday’s team meeting,” read what comes back, and send it. The second is A conversation where “human and AI engage in iterative dialogue to complete a task with the human refining the AI outputs.” You get something back, you react to it, and the thing improves. The work is a collaboration rather than an errand. : you ask for the same agenda, then follow up with “cut the middle section” and “make the tone less formal” until it is right. Same tool, same request, and about four minutes of difference. It is the second one that turns out to correlate with experience.
Experience pushes people toward iteration
In the study, high-tenure users were people who had signed up at least six months before the data was collected. The report’s own summary of what separates them is worth quoting directly: high-tenure users “are more likely to use Claude to iterate on their work, and much less likely to delegate greater responsibility through directive use patterns.”
The obvious reading is that experienced people use AI less. That is not what the data says, and the distinction matters. The same report finds high-tenure users are seven percentage points more likely to be using it for work at all. They use it more, on more of their job, and they hand over fewer complete tasks while doing so.
Put those two together and a picture emerges that is not “trust grows over time.” It is closer to: as you learn the tool, you get more specific about which parts of a job it should touch.
Two explanations, and the data cannot separate them
Being honest about what a finding like this can support matters more than the finding itself.
One reading is that experience teaches you where the tool fails. You have been burned by a confident, wrong, complete answer, so you keep your hand on the work and check it in stages.
The other is selection. Maybe the people who naturally work iteratively are also the people who found AI useful enough to still be around six months later, and the tenure is a consequence rather than a cause.
This is Data that photographs many people at one moment, rather than following the same people over time. It can show that two things occur together, like long-standing users and iterative habits, but it cannot show which came first. To separate cause from selection you would need to watch the same users change, or run an experiment. data, so it cannot tell you which of those is doing the work. It is probably some of both. What survives either way is the practical part: iteration is what sustained, work-heavy use actually looks like, so it is worth being deliberate about rather than treating delegation as the goal you graduate toward.
How this squares with “long chats get worse”
There is an apparent contradiction here with something else on this site, and it is worth resolving rather than hoping nobody notices.
AI Gave You a Bad Answer reports research finding that models score around 39% worse when a task is spread across a conversation instead of given in one message, and that when a model takes a wrong turn it does not recover. That sounds like an argument for delegating in one clean shot and never iterating.
They are not, and the difference is the whole practical lesson. It comes down to what the extra turns are carrying.
- Turns that carry the requirement make things worse. You give half the brief, the model guesses at the rest, you correct the guess, and it is now building on an interpretation it committed to three turns ago. That is the situation the 39% was measured in.
- Turns that carry feedback on the output do not. The brief was complete the first time. Now you are saying this section is too long, this example is wrong for my audience, cut the third point. You are not filling in a specification the model already guessed at, you are steering something that already understands the job.
So the two findings combine into one rule. Specify in one message, then refine. Which is also why the fix for a genuinely wrong answer is a fresh chat with a better brief rather than an argument in the existing thread.
What to do differently on Monday
- Notice which mode you are in. If you are typing a one-line request and accepting whatever returns, that is directive use, and it is the mode with the worst error profile for anything that matters.
- Delegate where checking is cheap. Formatting, first drafts of things you will rewrite anyway, converting one format into another. If a wrong answer is obvious at a glance, hand the whole thing over.
- Iterate where judgment matters. Anything where being subtly wrong is expensive, or where the answer depends on things only you know about your organization. That is the filter for choosing a first task applied at the level of a single request.
- Front-load, then steer. Everything you would have said in a follow-up goes in the first message. After that, react to what comes back rather than adding requirements you left out.
The takeaway
The people who have been at this longest are not the ones handing over the most. They use AI on more of their work and give it fewer complete jobs, which is a more useful model of expertise than “learn to trust it.”
Give the whole brief in one message, then refine the output rather than re-specifying the task. Delegation is a choice about a particular job with a particular cost of being wrong, not a level you unlock.
Sources: the tenure finding and the definitions are from Anthropic’s Economic Index reports, specifically the Learning curves (opens in a new tab) report, which sampled a million conversations over 5 to 12 February 2026 and defines high-tenure users as those who signed up at least six months earlier, and the Cadences (opens in a new tab) report for the mode definitions quoted. Two caveats that matter. This is Anthropic measuring how people use Anthropic’s own product, so it describes Claude users rather than AI users in general. And secondary coverage puts specific numbers on the tenure gap, roughly 38% directive use for newer users against 29% for long-standing ones, but those come from a table I could not read at the source, so they are not quoted above and the qualitative finding is what the argument rests on. The reconciliation with the multi-turn research, and the specify-then-refine rule, are mine.
Related: AI Gave You a Bad Answer: Why Correcting It in the Same Chat Usually Fails is the other half of this, on what to do when refinement is not the right move. The First Job to Give an AI Agent, and the One to Keep applies the same delegate-or-not question to whole workflows rather than single requests.