The document lands in your inbox looking finished. Clean structure, confident tone, plausible numbers. When a colleague asks where a specific claim came from, you realize you can't explain it. You weren't part of the thinking. You just received the output.
This is the moment most managers don't plan for. They've been told to hand tasks to AI. They hand them over. Something comes back looking polished, and they pass it along. When something later goes wrong, accountability is hard to find, because no one was in the loop while the work was happening.
The problem with starting at the finish line
When you hand over a whole task upfront – write this report, draft this proposal – you skip the part where you learn what the AI missed or quietly assumed. Your judgment about whether the output is correct depends on what you know. That knowledge only gets applied if you're in the conversation.
Anthropic reports that the median chat conversation producing an article involved 13 rounds of back-and-forth, while the median agent-style session began with a single human prompt. [1] This vendor-reported usage data describes how people used two products, not which method produced better work. It does show that product design can push people toward working back and forth or handing over the whole task.
Separately, vendor-reported data from Anthropic shows that users with at least six months on the tool are more likely to work back and forth with it, and less likely to hand over whole tasks at once. Whether this reflects learning over time, early-adopter differences, or something else isn't settled. But the pattern exists and is associated with more successful conversations. [2]
What the daily habit looks like
The iterative rhythm is simple: ask for a draft, review it seriously, explain what missed, refine it together. One person stays responsible for the final call. The first AI draft is not the deliverable – it's the starting point.
When you review, your job isn't to approve or reject. It's to name what's wrong. "This section assumes the old pricing structure." "This misses the vendor exception from last quarter." That explanation lets the next draft address the actual miss instead of guessing.
Someone must still own the outcome. Review can move between people, but responsibility for the final decision needs a name.
Evidence that keeping humans responsible matters
In a non-peer-reviewed study of 5,172 customer support agents, AI assistance increased issues resolved per hour by 15% on average. [3] In that deployment, human agents remained responsible: they could ignore or edit the AI's suggested replies, and they made the final call. The gains were concentrated among less-experienced workers, and the researchers found evidence consistent with worker learning.
The study supports a narrow conclusion. This particular assistant improved performance while people stayed responsible for customer conversations. It did not compare that workflow with end-to-end delegation, and one customer-support setting does not establish what will happen in management or other knowledge work.
What it does offer is a concrete example of AI supporting live work while the employee retains judgment. The manager does not have to choose between ignoring AI and surrendering the task. There is a useful middle: let the assistant propose, then require the employee to decide.
How a task earns delegation
Some tasks will eventually earn full delegation. A narrow, repetitive, well-understood task – one that has been through the review cycle many times – can eventually run without you in the loop. But delegation should be earned, not assumed from day one.
Before a task graduates to end-to-end AI handling, require a written success check. That means several real examples where the AI produced work you would have been comfortable putting your name on without edits – documented with dates, in a place your team can see. Not a general sense that things have been going well. Actual examples, with the task described and the outcome named.
Passing those examples makes the task a candidate to run alone, provided mistakes are easy to detect and reverse. If the task cannot pass, it stays in the review cycle. If a mistake could harm a customer, commit funds, or create a lasting obligation, it also stays there. A shared document with a task description, dated examples, and a note about who approved the graduation is enough to record the decision. The point is that delegation is explicit, not assumed because the work felt fine for a few weeks.
The operating decision
Stop treating the first AI output as a finished product. Treat it as a first draft that requires your explanation of what's wrong. Assign one person to review. Require that person to write down what they corrected and why. Run that loop until you have documented examples to justify stepping back – then, and only then, hand the whole task over.
The default is review. Delegation is the exception you earn.