The corpus covers most of these one at a time: the multimodal ones, code generation, and the cases where a managed AI service wins. This card is the whole list on one page. A scenario usually names the use case and nothing else, so the service has to come from the name alone.
Flash card
The GenAI use cases Domain 2 lists (image, video, and audio generation; summarization; AI assistants; translation; code generation; customer service agents; search; recommendation engines), and where each one lands on AWS.
- Image, video, and audio generation run on Amazon Bedrock. In the Amazon Nova family that is Nova Canvas for images, Nova Reel for video, and Nova Sonic for speech.
- Image generation commonly runs on diffusion models, a different family from the transformer-based LLMs behind text. A diffusion model starts from noise and subtracts a predicted amount of it over repeated steps.
- Summarization is a generative task a foundation model does well out of the box. Translation is too. Amazon Translate is the purpose-built option, charged per character at USD$15.00 per million for standard text.
- AI assistants and customer service agents are Amazon Bedrock when you build the conversation yourself, with Amazon Bedrock AgentCore as the runtime. Amazon Quick, which took over from Amazon Q Business, is the managed front end over your own data. Amazon Lex still fits a bot answering a fixed set of defined intents.
- Code generation is Kiro, which runs as an IDE, a CLI, a web app and a mobile app over one shared agent. When the generating happens inside your own application, it is a code-capable model called directly on Bedrock.
- Search here means retrieval. Amazon Bedrock Knowledge Bases embeds your documents, finds the passages, and has the model answer over what came back. Amazon Kendra moved to maintenance in July 2026, so do not reach for it. Recommendation engines are different again. That is Amazon Personalize, trained on item interaction data, not a foundation model handed a browsing history.
Pick it when
Pick a generative model when the output is new content (prose, an image, audio, working code), or an open-ended conversation with no script to follow.
It's the wrong answer when
It is the wrong fit when a purpose-built service covers the task more cheaply and more consistently. Sentiment and entity extraction belong to Amazon Comprehend, transcription to Amazon Transcribe, straight translation to Amazon Translate, and ranked recommendations to Amazon Personalize.