An engineering manager shares how they use LLMs to automate the data-collection layer of performance reviews, while keeping all judgment, ratings, and feedback decisions firmly human. The workflow involves feeding context (competency matrix, rating rubric, review template) and MCP-connected tools (Linear, GitLab, Slack, Notion, Honeycomb) into a model to produce an evidence-backed research brief. The post draws a clear three-layer distinction: data collection (LLM-suitable), analysis (LLM-helpful), and judgment (human-only). It also addresses privacy boundaries, transparency with direct reports, and the ethical asymmetry of managers using AI to evaluate employees.