You're Delegating More to AI Than Ever. Here's What You're Losing.

A quarter of OpenAI employees now run four or more AI agents simultaneously each week. Those agents are not writing their emails — they are producing strategy documents, analysing data sets, and completing work that would have taken a colleague an afternoon. The people doing it most effectively are not software engineers. A study of Claude Code users, reported by Ethan Mollick in "The Twilight of the Chatbots" (June 30), found that a lawyer performed as effectively as software engineers on coding tasks. Domain expertise — not technical background — predicts success.

The data does not show what comes next: how many of those professionals can tell when the agent got it wrong.

What happened this week

Mollick's June 30 piece drew on OpenAI internal data showing a surge in agentic usage across the company. Separate OpenAI data from June 25 showed that 80.6% of individual users now delegate tasks corresponding to 30 or more minutes of human work, with non-developer user growth at 137x since August 2025.

Two days later, Simon Willison highlighted Geoffrey Litt's concept of "cognitive debt" — the risk that as AI handles increasingly complex work, professionals lose the understanding needed to direct it well. Litt's warning: "you need a rich set of concepts in your mind to think creatively and fluently about how to move something forward." The more you delegate, the more you need to actively maintain the judgment that makes delegation safe.

This is not a new problem in a new form. The management equivalent is a team lead who stopped doing hands-on work so long ago they can no longer evaluate whether their team's output is good. With AI, that gap can open in months, not years.

What to do differently on Monday

Delegation is the new prompting skill. Effective delegation means scoping what the agent can decide, what it must flag, and what it must not touch — before it starts. The professionals generating the most output are not writing more elaborate prompts. They are constraining the task clearly enough that reviewing the output takes minutes, not a second round of work.

Cognitive debt accumulates silently. If you cannot explain why the AI's output is right — not just that it looks right — you have a debt problem. The test: can you identify where the agent's reasoning could have gone wrong, even if it didn't? If the answer is no, you are evaluating output on aesthetics, not accuracy.

Cross-model review is becoming a practical quality gate. On July 5, Willison reported that having different AI models check each other's work "really does work." It catches errors the original model missed — not because the second model is smarter, but because it approaches the task from a different direction.

Tariq: the variance report that looked right

Tariq is a finance manager at a 180-person manufacturer. He uses AI to produce the monthly variance report — pulling from ERP exports, calculating ratios, writing the narrative. The output is clean. He stopped rebuilding the model himself eight months ago.

Last month, the AI applied the prior year's margin targets to this year's cost structure. The numbers were internally consistent. The conclusion was wrong. The CFO caught it in the review meeting.

The problem was not the AI. Tariq had delegated enough that he no longer had a working model of where errors were likely to hide. Scoping the task more precisely — specifying which cost structure, which targets — would have prevented the error. Keeping the mental model of the report logic would have caught it in review.

Priya: the operations lead who stayed sharp

Priya is a senior operations lead at a 60-person logistics company. She runs three agents in parallel: one monitoring supplier quotes, one summarising exception reports, one drafting responses to client escalations. She reviews every output before it leaves her control.

She does more than proofread. She has a mental model of what each agent is likely to miss. Supplier quote summaries drop split-delivery terms. Exception reports de-emphasise single-order anomalies that are actually early signals of a larger problem. She built that model by running the same tasks manually for five weeks before handing them to agents. The investment meant the agents made her faster — and that she kept the judgment to catch them when they were wrong.

The one-sentence version

The professionals staying ahead of this are not delegating less — they are actively protecting the judgment that makes their delegation worth trusting.

<BlogPracticeSection />