Generative AI is revolutionizing business operations across various applications, including conversational assistants like Rufus and Amazon Seller Assistant, both under the Amazon umbrella. Beyond these, some of the most impactful generative AI applications operate autonomously in the background — an essential capability enabling businesses to transform their operations, data processing, and content creation at scale. These non-conversational implementations often take the form of agent workflows powered by large language models (LLMs), performing specific business objectives across industries without direct user interaction.
Unlike conversational applications that benefit from real-time user feedback and supervision, non-conversational applications exhibit unique advantages such as higher latency tolerance, batch processing, and caching. However, their autonomous nature necessitates stronger security measures and thorough quality assurance.
This article explores four distinct generative AI use cases within Amazon:
- Amazon Product Listing Creation and Catalog Data Quality Improvement – Showcases how LLMs assist selling partners and Amazon in creating higher-quality product listings at scale.
- Prescription Processing in Amazon Pharmacy – Demonstrates implementation in a highly-regulated environment and task decomposition for agent workflows.
- Review Highlights – Illustrates large-scale batch processing, traditional machine learning (ML) integration, the use of smaller LLMs, and cost-effective solutions.
- Amazon Ads Creative Image and Video Generation – Highlights multi-modal generative AI and responsible AI practices in creative work.
Each case study reveals different facets of implementing non-conversational generative AI applications, from technical architecture to operational considerations. Through these examples, you’ll see how a comprehensive suite of AWS services, including Amazon Bedrock and Amazon SageMaker, are critical to success. Finally, we outline key lessons learned commonly shared across the various use cases.
Creating High-Quality Product Listings on Amazon
Creating high-quality product listings with comprehensive details helps customers make informed purchasing decisions. Traditionally, selling partners would manually enter numerous attributes for each product. A new generative AI solution, launched in 2024, is transforming this process by proactively sourcing product information from brand websites and other sources to improve the customer experience.
Generative AI streamlines the experience for selling partners by supporting information inputs in various formats, such as URLs, product images, or spreadsheets, and automatically converting it into the required structure and format. Over 900,000 selling partners have used it, with nearly 80% of generated product listing drafts being accepted with minimal edits. The AI-generated content provides comprehensive product details, which helps enhance clarity and accuracy, thus aiding product discoverability in customer searches.
For new product listings, the workflow begins with the selling partner providing initial information. The system then generates a comprehensive product listing, including title, description, and detailed attributes, using multiple sources of information. The generated listing is shared with the selling partner for approval or editing.
For existing product listings, the system identifies products where additional data can be used to enrich the existing information.
Data Integration and Processing for Large Output
The Amazon team built robust internal and external source connectors for LLM-friendly APIs using Amazon Bedrock and other AWS services, enabling seamless integration with Amazon.com backend systems.
A major challenge was synthesizing varied data into a coherent product listing across 50+ attributes, including both text and numerical values. LLMs require specific control mechanisms and instructions to accurately interpret e-commerce concepts, as they may not perform optimally with such complex, diverse data. For instance, an LLM might mistake “capacity” in a knife block as dimensions instead of the number of slots, or interpret “Fit Wear” as a style description rather than a brand name. Prompt engineering and fine-tuning were extensively used to address these cases.
Generation and Validation Using LLMs
Generated product listings should be complete and correct. To help achieve this, the solution implements a multi-step workflow using LLMs for both attribute generation and validation. This dual-LLM approach helps prevent hallucinations, which is crucial when dealing with safety implications or technical specifications. The team developed advanced self-reflection techniques to ensure the generation and validation processes effectively complemented each other.
Multi-Layer Quality Assurance with Human Feedback
Human feedback is central to the solution’s quality assurance. The process includes initial assessments by Amazon.com experts, as well as input from selling partners for approval or editing. This provides high-quality outputs and enables continuous enhancement of the AI models.
The quality assurance process includes automated testing methods incorporating ML, algorithms, or LLM-based evaluations. Failed listings are regenerated, and successful listings proceed to further testing. Using [causal inference modeling], we identify the underlying characteristics that influence product listing performance as well as enrichment opportunities. Ultimately, product listings that pass quality checks and are accepted by selling partners are published, ensuring customers receive accurate and comprehensive product information.
Application-Level System Optimization for Accuracy and Cost
Given the high standards for accuracy and completeness, the team employs a comprehensive experimentation approach equipped with an automated optimization system. This system explores various combinations of LLMs, prompts, playbooks, workflows, and AI tools to improve higher-level business metrics, including cost. Through continuous evaluation and automated testing, the product listing generator can effectively balance performance, cost, and efficiency while adapting to new AI advancements. This approach means customers benefit from high-quality product information, and selling partners gain access to cutting-edge tools for efficient listing creation.
Generative AI-Based Prescription Processing in Amazon Pharmacy
Building on the human-in-the-loop workflow foundations of the seller listing example discussed previously, Amazon Pharmacy demonstrates how to apply these principles in a Health Insurance Portability and Accountability Act regulated industry. In the [Learning how Amazon Pharmacy creates an LLM-based chatbot using Amazon SageMaker] article, we shared a conversational assistant for patient care specialists, and now we focus on the automated prescription processing.
At Amazon Pharmacy, we developed an AI system built on Amazon Bedrock and SageMaker to help pharmacist technicians process medication instructions more accurately and efficiently. The solution integrates human experts with LLMs in creation and validation roles to enhance the accuracy of patient medication instructions.
Delegated Workflow Design for Healthcare Accuracy
The prescription processing system blends human expertise – data entry specialists and pharmacists – with AI assistance to provide direction suggestions and feedback. The workflow begins with a pharmacy knowledge base preprocessor, which standardizes raw prescription text from [Amazon DynamoDB] before a fine-tuned small language model (SLM) on SageMaker is used to identify key components (dosage, frequency).
The system seamlessly integrates experts such as data entry specialists and pharmacists, where generative AI complements the overall workflow to enhance agility and accuracy, thereby better serving our patients. A directions assembly system with guardrails then generates instructions for data entry specialists to create their typed directions via a recommendation module. A flagging module flags or corrects errors and enforces additional guardrails as feedback for the data entry specialist. Technicians finalize a highly accurate, safely typed direction for pharmacist feedback or downstream service execution of the direction.
A highlight of the solution is the use of task decomposition, which allowed engineers and scientists to break down the entire process into multiple steps, including individual modules composed of sub-steps. The team made extensive use of fine-tuned SLMs. Additionally, the process incorporates traditional ML routines, such as [named entity recognition (NER)] or using [regression models ]to estimate final confidence scores. The use of SLMs and traditional ML in this controlled, well-defined process significantly improves processing speed while maintaining strict safety standards due to the incorporation of appropriate guardrails at specific steps.
The system includes multiple well-defined sub-steps, with each sub-process operating as a specialized component, working in a semi-autonomous but collaborative manner within the workflow towards an overall goal. This decomposed approach, with specific validations at each stage, proves more effective than an end-to-end solution, while allowing for the use of fine-tuned SLMs. The team uses [AWS Fargate ]to orchestrate the workflow due to its current integration into existing backend systems.
During the team’s product development, they leaned into Amazon Bedrock, which offers high-performing LLMs with ease-of-use features tailored for generative AI applications. SageMaker supports further LLM selection, deeper customization, and traditional ML approaches. To learn more about this technology, see [How task decomposition and smaller LLMs can make AI more affordable], and read the [Amazon Pharmacy business case study].
Building a Reliable Application with Guardrails and HITL
To comply with HIPAA standards and provide patient privacy, we implemented stringent data governance practices, while adopting a hybrid approach that combines fine-tuned LLMs using the Amazon Bedrock API with retrieval-augmented generation using [Amazon OpenSearch Service]. This combination allows for efficient knowledge retrieval while maintaining high accuracy for specific sub-tasks.
Managing LLM hallucinations, which is critical in the healthcare domain, requires more than just fine-tuning on large datasets. Our solution implemented domain-specific guardrails built on [Amazon Bedrock Guardrails ]and supplemented by human-in-the-loop (HITL) oversight to enhance the system’s reliability.
The Amazon Pharmacy team continues to enhance the system with real-time pharmacist feedback and expanded prescription format capabilities. This balanced approach of innovation, domain expertise, advanced AI services, and human oversight not only enhances operational efficiencies, but also means that the AI system correctly augments healthcare professionals in providing optimal patient care.
Generative AI-Powered Customer Review Highlights
Our previous example showed how Amazon Pharmacy integrates LLMs into a real-time workflow for prescription processing, while this use case shows how similar technologies – SLMs, traditional ML, and thoughtful workflow design – can be applied to large-scale [offline batch inference].
Amazon introduced [AI-generated customer review highlights ]to process over 200 million annual product reviews and ratings. This feature distills shared customer opinions into concise paragraphs, highlighting positive, neutral, and negative feedback about products and their features. Shoppers can quickly grasp consensus while maintaining transparency by providing access to the relevant customer reviews and retaining original reviews.
The system enhances shopping decisions through an interface where customers can explore review highlights by selecting specific features—such as picture quality, remote control capabilities, or ease of installation for a Fire TV. These features are marked with a green checkmark to indicate positive sentiment, an orange minus sign to indicate negative sentiment, and gray to indicate neutral sentiment—meaning shoppers can quickly identify a product’s strengths and weaknesses based on verified purchase reviews.
Using LLMs Cost-Effectively for Offline Use Cases
The team developed a cost-effective hybrid architecture combining traditional ML methods with specialized SLMs. This approach assigns sentiment analysis and keyword extraction to traditional ML, while using optimized SLMs for complex text generation tasks, enhancing accuracy and processing efficiency.
The feature employs [SageMaker batch transform ]for asynchronous processing, enabling significant cost reductions compared to real-time endpoints. To deliver a near-zero latency experience, the solution [caches ]extracted insights and existing reviews, reducing wait times and allowing multiple customers to access concurrently without additional computations. The system incrementally processes new reviews, updating insights without re-processing the complete dataset. For optimal performance and cost-effectiveness, the feature uses [Amazon Elastic Compute Cloud ](Amazon EC2) [Inf2 instances ]for batch transform jobs, [providing up to 40% better price performance compared to alternatives].
By following this comprehensive approach, the team effectively manages costs while handling a large volume of reviews and products, making the solution efficient and scalable.
Amazon Ads AI-Driven Creative Image and Video Generation
In previous examples, we primarily explored text-centric generative AI applications, and now we will shift to multi-modal generative AI with [Amazon Ads Sponsored Ads Creative Content Generation]. This solution has [image ]and [video ]generation capabilities, and we will share detail about these features in this section. Overall, at the heart of the solution is the use of the [Amazon Nova] creative content generation model.
Working backwards from customer needs, an Amazon survey in March 2023 revealed that nearly 75% of advertisers cited creative content generation as their top challenge when striving for campaign success. Many advertisers, particularly those without in-house capabilities or agency support, face significant hurdles due to the expertise and cost of producing high-quality visuals. The Amazon Ads solution democratizes visual content creation, making it accessible and efficient for advertisers of all sizes. The impact is significant: advertisers using AI-generated images in their [Sponsored Brands ]campaigns saw a nearly 8% increase in [click-through rate (CTR)], and submitted campaigns 88% more often than non-users.
Last year, the AWS Machine Learning Blog published an article [detailing the image generation solution]. Since that time, Amazon has embraced [Amazon Nova Canvas ]as the foundation for creative image generation. Leveraging text or image prompts, combined with text-based editing features and controls for color scheme and layout adjustments, it creates professional-grade images.
In September 2024, the Amazon Ads team added the ability to create [short video ads ]from product images. This feature uses [foundation models available on Amazon Bedrock ], giving customers control via natural language controls over visual style, pacing, camera movement, rotation, and zooming. It uses an agent workflow to first describe the video storyboard and then generates the story content.
As discussed in the original article, [Responsible AI ]is at the heart of the solution, and the Amazon Nova creative model comes with built-in controls to support safe and responsible AI use, including watermarking and content moderation.
The solution uses [AWS Step Functions ]and [AWS Lambda ]functions to orchestrate serverless coordination of the image- and video-generation processes. Generated content is stored in [Amazon Simple Storage Service ](Amazon S3), metadata is stored in DynamoDB, and [Amazon API Gateway ]provides customer access to the generation functionality. The solution now incorporates Amazon Bedrock Guardrails in addition to maintaining the [Amazon Rekognition ]and [Amazon Comprehend ]integration in various steps for further safety checks.
Creating high-quality ad creatives at scale presents complex challenges. Generative AI models need to produce engaging and brand-appropriate visuals across diverse product categories and advertising contexts, while remaining easily accessible to advertisers of all technical skill levels. Quality assurance and improvement are foundational to the image and video generation capabilities. The system continuously enhances through extensive HITL processes facilitated by [Amazon SageMaker Ground Truth]. The implementation provides a powerful tool to transform advertisers’ creative processes, driving increased ease of high-quality visual content creation across diverse product categories and contexts.
This is just the beginning of how Amazon Ads uses generative AI to help advertisers who need to create content suited for advertising goals. The solution demonstrates how reducing creation friction can directly improve advertising campaigns while maintaining the high standards of responsible AI use.
Key Technical Learnings and Discussion
Non-conversational applications benefit from higher latency tolerance, enabling batch processing and caching, but require robust validation mechanisms and stronger security measures due to their autonomy. These insights are applicable to both non-conversational and conversational AI implementations:
- Task Decomposition and Agent Workflows – Breaking down complex problems into smaller components has proven valuable across various implementations. This thoughtful decomposition by domain experts enables the creation of specialized models for specific sub-tasks, as exemplified by prescription processing in Amazon Pharmacy, where fine-tuned SLMs handle discrete tasks such as dosage identification. This strategy allows for the creation of specialized agents with clear validation steps, enhancing reliability and simplifying maintenance. The Amazon seller product listing use case exemplified this through its multi-step workflow with separate generation and validation processes. Additionally, the review highlights use case showed the cost-effective and controlled LLM usage, namely, by using traditional ML for preprocessing and performant parts that can be associated with the LLM tasks.
- Hybrid Architectures and Model Selection – Combining traditional ML with LLMs offers better control and cost-effectiveness compared to purely LLM-driven approaches. Traditional ML excels at well-defined tasks, as shown by the review highlights system’s use for sentiment analysis and information extraction. Amazon teams have strategically deployed large and small language models based on requirements, combining RAG with fine-tuning for effective domain-specific applications, like the Amazon Pharmacy implementation.
- Cost Optimization Strategies – Amazon teams have achieved efficiencies through batch processing, caching mechanisms for high-volume operations, specialized instance types like [AWS Inferentia ]and [AWS Trainium], and optimized model selection. Review highlights demonstrated how incremental processing reduces computational demands, while Amazon Ads utilized Amazon Nova foundation models for cost-effective creative content creation.
- Quality Assurance and Control Mechanisms – Quality control relies on domain-specific guardrails via Amazon Bedrock Guardrails and multi-layered validation combining automated testing with human evaluation. Dual-LLM approaches for generation and validation help prevent hallucinations in Amazon seller product listings, while self-reflection techniques enhance accuracy. Amazon Nova creative FMs provide inherent responsible AI controls and are supplemented with continuous A/B testing and performance measurement.
- HITL Implementation – HITL approaches span multiple layers, from expert assessments by pharmacists to end-user feedback from selling partners. Amazon teams established structured improvement workflows, balancing automation with human oversight based on specific domain requirements and risk profiles.
- Responsible AI and Compliance – Responsible AI practices encompass content intake safeguards for regulated environments and compliance with regulations like HIPAA. Amazon teams integrated content moderation for user-facing applications, maintained transparency in review highlights by providing access to source information, and implemented data governance with monitoring to enhance quality and compliance.
These patterns enable scalable, reliable, and cost-effective generative AI solutions while maintaining quality and responsibility standards. These implementations show that effective solutions require not just advanced models but also careful attention to architecture, operations, and governance, supported by AWS services and established practices.
Next Steps
The Amazon examples shared in this article illustrate how generative AI can create value beyond traditional conversational assistants. We invite you to follow these examples, or create your own, to learn how generative AI can reshape your business or even your industry. You can visit the [AWS generative AI use case ]page to launch your ideation process.
These examples demonstrate that effective generative AI implementations often benefit from combining different types of models and workflows. To understand which FMs are supported by AWS services, refer to [Foundation models supported in Amazon Bedrock ]and [Amazon SageMaker JumpStart Foundation Models]. We also recommend exploring [Amazon Bedrock Flows ], which can simplify the path to building workflows. Additionally, we remind you that Trainium and Inferentia accelerators offer important cost savings across these applications.
As illustrated in the examples we have shown, agent workflows have proven particularly valuable. We recommend you browse [Amazon Bedrock Agents ]to quickly build agent workflows.
Successful generative AI implementation is more than model selection – it represents a comprehensive software development process from experimentation through application monitoring. To begin building your foundation across these essential services, we invite you to browse [Amazon QuickStart].
To learn more about how Amazon uses AI, see [AI at Amazon] in Amazon News.