Do managers delegate to machines even when human judgment is better?

Agrawal, Gans and Goldfarb's model argues that decision-makers will sometimes give machines real authority over a decision even though human judgment, if applied, would produce a better result. As the abstract states it: "humans will delegate some decisions to machines even when the decision would be superior with human input." The paper locates this result in what it calls "span of attention issues for human judgment," worked out in the model's Section 5, where it "demonstrat[es] that humans may give the machines real authority even when human input would lead to better decisions."

The underlying mechanism is that judgment is a scarce, deliberate activity — the paper insists "judgment is not a passive activity. Instead, it requires deliberative cognitive application" — and a decision-maker with many decisions to oversee cannot apply that deliberate attention to all of them. Rather than treating judgment as free to use whenever it is available, the authors treat it as something a human must choose where to spend, and under a limited span of attention, delegating a given decision to the machine's prediction-driven default can be optimal in aggregate even when, decision by decision, a human who looked closely would have chosen better.

This gives an economic, attention-scarcity explanation for a pattern Why do AI-delegated firms stop exploring new business models? documents from a different angle: managers who could exercise superior judgment decline to, and here the reason offered is that attention itself is the binding constraint, not inattentiveness or incentive failure. It also qualifies What makes accountable judgment scarce when AI cognition is cheap?: judgment being the scarce, valuable asset does not guarantee it gets used — the same scarcity that makes it valuable is what crowds it out of any given decision.

The excerpt gives the qualitative result — that under-delegation of judgment can be rational given limited attention — without the full quantitative conditions under which it holds generally, since the surrounding equations in the excerpt are garbled by extraction (the probability terms and payoff expressions are not fully legible). It is also framed around a single stylized decision rather than a portfolio of decisions competing for the same attention, so the excerpt does not establish how the result scales once many decisions compete for one decision-maker's limited judgment capacity — only that the single-decision case can tip toward delegation even when human input would have been better.