A business that runs advertising with major partners has a quiet bottleneck. Every piece of creative has to be approved before it runs. A logo is the wrong size. A required disclaimer is missing. A competitor's mark is visible in the background of a video. Someone has to catch all of it, by eye, across hundreds of assets.
We built an agentic approval system for a client in exactly that spot, reviewing creative that had to meet the rules of major distributors and rights holders. The client isn't named here. The design is worth sharing because it applies to any high-volume review.
The problem with manual review
Review by eye is slow and inconsistent. Two reviewers apply the same rule differently. Video is the worst case, because a violation can sit in three seconds of a ninety-second spot. And when a partner rejects something after submission, the whole campaign waits.
The goal wasn't to replace reviewers. It was to have them look at the right things, with the evidence already found.
Step 1: turn the rulebook into checks
Brand and partner guidelines arrive as long PDFs. We broke them into individual rules a system can test. Each rule got:
- A plain statement of the requirement.
- What kind of asset it applies to.
- How to detect it: a visual check, a text check, an audio check or a judgment call.
- What happens on a miss: block, or flag for review.
This is the unglamorous step, and it decides whether everything after it works. Vague rules get vague results.
Step 2: use the right detector for each signal
No single model handles creative well. The pipeline runs several, each on what it's good at.
- Object detection finds logos and marks in images and video frames, with a box around each one. That gives position and size, which is what most placement rules are about.
- Text recognition reads on-screen copy, so required wording and prohibited claims can be checked exactly.
- Speech-to-text turns audio and voiceover into a transcript with timestamps.
- A language model handles the judgment calls the others can't: tone, implied endorsement, whether a claim needs a disclaimer.
For video, frames are sampled across the whole runtime, more densely around scene changes, so a brief violation isn't missed.
Step 3: check with rules before asking a model
Once detectors have produced structured findings, most rules are simple comparisons. Is the mark at least the minimum size? Is the disclaimer present and on screen long enough? Is a blocked term in the transcript?
Code answers those the same way every time. The language model is saved for what's left. That keeps results consistent and costs down.
Step 4: show the evidence
Every finding carries its proof: the frame, the box around the mark, the timestamp, the line of transcript, and the rule it relates to.
A reviewer doesn't get "possible logo issue." They get the exact frame with the mark outlined and the rule beside it. They can decide in seconds, and the person who made the creative knows precisely what to fix.
Step 5: route by confidence
Clean assets pass. Clear violations go back to the creator with the evidence. Uncertain cases go to a named reviewer.
Reviewer decisions are logged with their reasons. Those decisions become new test cases, so the system gets better exactly where it was unsure.
What made it work
- Detection models trained on the actual marks. Generic models miss partner logos, especially small or partly hidden ones.
- A test set of past decisions. Approved and rejected creative with the reason for each, run on every change. See how to test an AI agent before production.
- Rules kept as data. When a partner updates its guidelines, the rule changes in one place. Nobody retrains anything.
- People kept in charge. The system recommends. A person approves anything uncertain.
Where else this applies
The same pattern fits any review of mixed media against written rules: marketing compliance, packaging and label checks, training content, partner co-branding. Break the rulebook into checks, use a detector per signal, compare in code, attach the evidence and route the exceptions.
If you have a review queue like this, see how we approach document and content review or tell us about it.