Open-ended feedback is among the most common forms of consequential communication in

organizations, yet the messages people write are often sparse and generic even when they

hold detailed private evaluations. We propose that this reflects a production

constraint---turning a private judgment into an informative and socially appropriate

message is effortful---and test whether generative AI can relax it. Rather than composing

on the writer's behalf, the AI asks targeted follow-up questions and assembles the

writer's own answers into a revised message. Across three preregistered experiments

(N = 4,011), AI assistance made messages longer, more concrete, and substantially more

useful and credible to independent expert evaluators, for both course feedback (Study 1)

and feedback to a difficult coworker (Study 2). A rewrite-only condition produced about a

tenth of the gain: the value lies in the questions the AI asks, not the polish (Study 2).

In an incentivized hiring experiment, AI-assisted self-promotion conferred a large

advantage on whichever candidate adopted it---an advantage that dissipated when both did,

even as participants remained willing to pay for the assistance (Study 3).