Brand teams shipping 400+ creative variations a quarter have a quiet problem: human QA doesn't scale, but a single off-brand asset in market is worse than three approved-and-mediocre ones. Most teams either over-staff QA or under-ship volume. There's a third option.
The 'middle manager' pattern uses a single Claude or GPT-4 evaluator with three things wired into its system prompt: the brand voice charter (verbatim), the do-not-say list (verbatim), and three reference assets that are explicitly on-brand and three that are explicitly off-brand. The evaluator returns a pass/flag/fail with a one-sentence rationale.
Pass rate in our pilots: 71% of variations approved without human review. Of the 29% flagged, roughly 60% were genuine catches and 40% were over-cautious, both an acceptable failure mode for a system that costs roughly $0.04 per asset to run.
The deployment lift is small. Three Fortune 500 brand teams now run this in production. Two of them retired a contractor row from the budget. The third re-routed the contractor to higher-leverage strategic work.
The prompt template, the do-not-say list format, and the reference-asset selection criteria are all in the appendix.
"The model isn't the creative director. It's the middle manager who reads the brief twice and says 'this isn't quite right' before the work hits the CD's desk."