There’s a fourth symptom of AI adoption amongst managers that is likely the most tempting and damaging - using AI to write for you. It’s the most tempting because you’ve now built a shaped evidence log so clean that “draft the review for me” feels like the obvious next step. Resist it - writing is where you form your opinion. If you outsource the draft, you’ve outsourced the thinking.
Write your draft by hand while referencing the evidence log. Then ask AI to judge it objectively against the evidence log, historical docs, your goals for the task, and your organization’s artifacts (competency maps, level expectations, etc.). You’re the model under evaluation. The rubric is the test suite and the AI is the harness. The rubric targets the things in your own writing you can’t see.
Clarity - am I using couching language, weasel words, hedges that let me avoid saying the thing
Concision - can fewer words better focus and allocate the reader’s attention
Utility - is this actionable; am I celebrating the behaviors I want to see more of, not just flagging gaps
Every claim in your feedback should be traceable to the log. AI’s job is to flag the ones that don’t. What did you assert but didn’t supply evidence for? Now who’s auditing who.
Use AI to prep for the conversation. Talk through the delivery - how to get the engineer in a reflective state of mind, how to weight positive and constructive feedback, how to open the door for feedback on your feedback. Have AI produce a conversation guide that weaves together discussion prompts, evidence, and space to take notes live.
This is what “better” looks like: a review where every sentence is yours, backed by evidence, and something the engineer can act on.