Advertising & regulated industries
Catching non-compliant ads before they go live.
A multimodal AI system that reads advertising creatives the way a reviewer does, combining images, text and layout, to flag compliance issues automatically.
- Project
- Multimodal creative compliance
- Client
- International media agency network
- Capabilities
- Computer vision · NLP · Multimodal classification · Workflow automation
- of non-compliant content flagged automatically
- 92%
- reduction in manual compliance checks
- 50%+
See it work
interactive demo · fictional brands
Vision + text + layout, analysed together.
The problem
Every advertising creative has to meet brand guidelines, platform policies and, in regulated categories like food, alcohol and finance, legal requirements. Checking them was a manual job: reviewers looking at each asset against long checklists, under deadline pressure, at growing volume.
Text-only tools miss most of it. A compliance issue is often about what is shown, what is said and how it is laid out together: a disclaimer that exists but is unreadable, a claim placed next to a product it doesn’t apply to, a logo that is present but obstructed.
Our approach
We built a system that treats an ad as a document with visual, textual and spatial structure:
- Computer vision and CLIP-based models detect logos, products and visual elements and understand what the image depicts.
- OCR and LayoutLM read the text and where it sits: its size, position and relationship to everything else on the creative.
- A multimodal classification layer combines these signals and maps them to specific policy rules, so each flag comes with the region and the reason, not just a score.
Reviewers stay in control. The system clears the obvious cases, flags the risky ones with evidence, and sends the ambiguous ones to a person.
The result
92% of non-compliant content was flagged automatically, and manual compliance checks fell by more than half. Reviewers spent their time on the creatives that actually needed judgement.
Client details are anonymised. Demos are illustrative, built on synthetic data and fictional brands.