PM at a F500, leading a 14-person pod (4 eng squads, design, data). Six months ago I started drafting PRDs with an agent loop that pulls from our Jira, Confluence, and a vector index of past launches. Time to first draft went from roughly 6 hours to 40 minutes. I thought I'd reclaimed a day a week. I didn't. What actually happened: my eng leads now get three PRDs where they used to get one, and the quality variance is wider, not narrower. The agent confidently invents dependencies on services that were deprecated in 2024. Twice it cited metrics from a deck that never shipped. My staff engineer started a Slack channel called #prd-fact-check. The bottleneck moved, it didn't disappear. I now spend the saved drafting time defending claims I didn't personally make. The honest framing for my org would be: writing got cheaper, reviewing got more expensive, and we haven't rebalanced headcount or rituals to match. Nobody on my leadership team wants to say that out loud because the productivity number looks good in the QBR.