AWS re:Invent 2025 - Amazon Ads Creative Agent uses AWS to democratize ad creation (IND3335)

AWS re:Invent 2025 - Democratizing Ad Creation with Amazon Ads Creative Agent

Overview of Amazon Ads

• Amazon Ads is a full-funnel advertising solution available to businesses of all sizes
• Serves over 300 million average monthly reach across the US, nearly the entire adult population
• Powers advertising across Amazon.com, Prime Video, Twitch, Audible, Fire TV, Music, and third-party publishers

Introduction to Amazon Ads Creative Agent

• Creative Agent is a complete production-agnostic system running on AWS, not just a text-to-image or text-to-video model
• Enables advertisers to generate images, audio, videos, and headlines across a variety of ad formats and surfaces

Key Challenges Addressed by Creative Agent

1. Scale: Advertisers need to generate 10-50 variations for every campaign, across hundreds of products and events
2. Automation vs. Control: Advertisers want automation but also the ability to control brand guidance and creative direction
3. Time-to-Market: Advertising trends shift quickly, so content needs to be generated and optimized rapidly
4. Cost: Small advertisers cannot afford expensive content creation services
5. Expertise: Not all small businesses have access to technical creators, content writers, and optimization experts

How Creative Agent Works

• Integrates multiple models and features for image generation, audio generation, video generation, and more
• Provides a natural language interface and full-fledged image/video editing capabilities
• Uses planning, reasoning, and multiple agents to act on behalf of advertisers intelligently

Technical Innovations in Creative Agent

1. Agent Architecture: Modular, specialized agents and tools enable parallel development and execution of long-running tasks
2. Parallel Execution: Leverages parallel tool calling and stateful tools to optimize for speed and cost
3. Multimodal Context Management: Intelligently selects and forwards relevant information to the language model, avoiding context overload
4. Evaluation and Observability: Uses adaptive benchmarks, automated testing, and advanced logging/diagnostics to ensure quality and reliability

Business Impact and Results

• 338% increase in CTR compared to non-Genai campaigns
• 89% of offers were new-to-brand, unlocking new customer segments
• 121% return on ad spend, lowering costs while increasing revenue

Lessons Learned and Future Roadmap

• Productionalize building blocks, not just prototypes
• Keep agent scope small, optimize for context, and delegate through an orchestrator
• Build for parallelism and retry from the start to handle nondeterministic agents
• Make evaluation and observability a core part of development

AWS AI/ML Landscape for Agentic Systems

• Infrastructure layer: Amazon SageMaker, Trinium, Inferentia for model training and hosting
• Development services: Amazon Bedrock Agent Core, Amazon Kite for purpose-built agentic systems
• Key capabilities: Agent runtime, agent gateway, agent memory, foundation models, and open-source SDKs