Global payments — 2026
Automating brand compliance for a global payments network
A three-week AI discovery that replaced a 18,000-hour-a-year manual brand review process with an agentic scanner design, across 210 markets.
Stack
The problem
A designer in one of 210 markets submits a campaign for brand review. Two weeks
later they learn the logo clearspace is wrong. Another week: the red is
#E8001B and should be #EB001B. Multiply by roughly seven thousand queries a
year at fifty minutes each — about 18,000 hours, or $600K, spent checking
whether the red is the right red.
Half of all assets bypassed review entirely, because the queue was the slowest part of the process.
What we designed
A scanner that reads a creative asset and returns a graded compliance report, built from deliberately unglamorous parts, each chosen because it is measurable:
- pHash and MD5 against the approved-asset registry, to separate “this is the approved file” from “this looks like it”.
- Delta-E colour distance rather than hex comparison, because brand red is a perceptual target, not a string.
- YOLO detectors custom-trained on the brand symbol, brand mark and partner logos; SAM for shape boundaries; Hough transforms for circularity and rotation.
- OCR plus rule checks for typography, legal text and campaign copy.
The breakthrough was architectural. Instead of one monolithic model asked to “check the brand guidelines”, each rule became a small deterministic check with its own confidence, and an agent layer above them decided what to escalate to a human. Rules are auditable, cheap to add, and can be pointed at a guideline PDF rather than a training run.
What actually happened
The business case for the original tool weakened halfway through — a good outcome, found in week two rather than month six. The scope pivoted from a single review tool to a platform for guideline-driven asset checking, which is what got approved.