The coaching team got their week back.

Three people spent their days reviewing student homework. Review cycles took two to three weeks. The rubrics and standards existed, scattered across documents, but the senior reviewers had internalized them so completely they didn't realize how much of what they did was undocumented. New hires took months to train because the knowledge wasn't truly accessible. It lived in people's heads. Every review required manually cross-referencing multiple documents, and the business cost was becoming clear: delays frustrated students, senior staff couldn't move to other work, and growth was capped by reviewer capacity.

We spent the first two weeks watching the reviewers work in real time. Not reading the manual. Watching the actual decisions. We discovered eight implicit rules that weren't in any rubric, three common edge cases that required judgment calls, and two shortcuts the senior reviewers used instinctively. We indexed the entire knowledge base, built a RAG system trained on the real rubrics and standards, and created an agentic review assistant with confidence scoring. Low-confidence reviews get flagged for a human. Everything else runs automatically.

Review cycles went from two to three weeks down to minutes. The three people who had been doing reviews full-time were redeployed to curriculum development and student support. Students started receiving immediate feedback. The system hit 92% accuracy and keeps improving as it calibrates against real outcomes. The biggest change: the company can now scale student capacity without hiring more reviewers. The knowledge that used to live in three people's heads is now queryable by the entire organization.