It's tempting to sell an AI grader on automation: feed it essays, get grades, done. That product fails the moment a professor catches one wrong grade in a student's record. So BUC Populi sells the opposite - the human stays in charge.
Make the gate the feature
Claude does the first pass against a rubric and returns per-criterion scores and feedback. Every result is written as a draft at pending_review. Nothing reaches a student until a human approves it. That constraint is what got the tool adopted by faculty - it removed the busywork without asking anyone to trust a model with the part that matters.
Automating the draft is a time-saver. Automating the decision is a liability. The product is knowing the difference.
Sell the outcome, not the model
What professors actually bought wasn't "AI grading":
- A first draft that's consistent across a stack of essays, so they're editing, not starting cold.
- Confidence flags (plagiarism, off-topic, missing citations) that tell them where to look first.