Why the review step is the job, not overhead
The suggested-reply loop is: AI drafts from the case and the pack, you review against the cited article, you edit, and it sends under your name. The customer never learns the tool exists from the signature. On a busy shift, the review step is the one you'll be tempted to skim. Here's the evidence for why you shouldn't — and it's the best evidence any occupation in this course has.
Brynjolfsson, Li and Raymond, "Generative AI at Work" (Quarterly Journal of Economics, May 2025), studied a staggered rollout of an AI assistant across roughly 5,200 support agents at a real company — not a lab task, actual queues. Findings:
- 15% average productivity gain in issues resolved per hour.
- The newest agents gained most — faster and higher quality. The assistant was, in effect, whispering the moves of the strongest agents to people still learning them.
- The most experienced agents saw small speed gains and small quality declines. Read that twice. For the best people on the team, accepting the machine's suggestion as-is was a downgrade from what they'd have written themselves.
Both halves matter for how you run the loop:
- If you're new: the drafts are genuinely good scaffolding — they encode practices it would take you months to absorb from shadowing. Reviewing them against the article is also how you learn the policy, the way copying out a good reply used to be.
- If you're experienced: your edit is the value. The seniors in the study who lost quality were the ones for whom the suggestion was worse than their instinct. The review step is where you put your edge back in — tightening the promise the draft almost made, catching the article that changed last month, hearing the anger the model read as a status question.
Review with teeth
A review pass that approves everything isn't review; it's latency. In practice a healthy loop rejects or heavily edits a meaningful slice of drafts, and — this is the part almost nobody does — logs the edits. "Removed 'guaranteed by Friday', added the actual carrier window" is not paperwork. Each logged correction is a real case where the drafter went wrong, with the known-good answer attached.
Hold that thought. In the next lesson those logged corrections become golden eval cases — the raw material a macro has to pass before it earns any autonomy at all. Today's annoyed edit is tomorrow's automated gate.
First, though: build the gate for single drafts. Next two steps you write the linter, then wire it to the citation check from the fix you just did.