A study of 27,778 messages on a crisis counseling platform gave researchers a large-scale look at how AI-generated draft replies change counselor behavior. Counselors who used suggested drafts sent faster replies and covered more recommended topics. Those same counselors said the drafts sometimes felt inauthentic. Several said they worried their own skills were atrophying from leaning on suggestions. The mixed-methods data surfaces a tension that matters beyond crisis counseling. Approving a full draft is fast, but it is a different act than writing from scratch. Counselors described skipping reflection they would otherwise do. Some felt the AI voice was flattening the personal quality that makes the conversation effective. The benefit and the harm come from the same source: the draft does the work, and the counselor decides whether to send it. For teams designing AI assistance in high-stakes roles, this study is direct evidence that speed and coverage gains are real and measurable. Skill erosion and inauthenticity are reported but harder to quantify. The design question is not whether to offer drafts but how much the person still shapes the reply before it goes. Partial drafts, suggested phrases, or structured prompts may preserve the counselor's voice and judgment in ways that full-draft approval does not.