What challenges did Amazon.com face before AI implementation?

Extract and enrich structured product attributes and generate accurate product titles at massive scale. Improve performance on edge cases, evolving terminology, and ambiguity without slow manual prompt updates and redeployments.

What was the impact of the AI solution?

The article says error and disagreement rates fell continuously as learnings accumulated. Costs decrease because supervisor calls are selective, while quality increases across millions of products. The system scales catalog accuracy improvement without retraining.

Source

Evidence: Low25/100

Amazon.com Catalog Team built a self-learning generative AI system using Amazon Bedrock AgentCore (self-improving attribute extraction)

Use case typeAI platformUpdated Jan 23, 2026

Amazon.com (Amazon Selection and Catalog Systems) built a self-learning system for catalog enrichment that extracts structured product attributes and generates titles at massive scale. The system uses multiple smaller worker models, a supervisor agent, and a hierarchical knowledge base that captures learnings from disagreements and feedback signals to continuously improve accuracy without retraining.

Organization: Amazon.com
Industry: Retail
Location: United States
Published: January 2026

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks

Customer: Amazon.com
Provider: AWS
Maturity: Unknown
Linked source: AWS Machine Learning Blog

No explicit deployment-stage evidence found.

Customer identity supportedSource describes one deploymentMaturity evidence evaluated

Primary read

Use case focus

Showing 3 of 4

1Catalog enrichment
2Product attribute extraction
3Content generation

Implemented a generator-evaluator multi-model architecture where lightweight worker models handle routine product cases through consensus.
When workers disagree, a supervisor agent investigates with specialized tools, generates reusable learnings, and stores them in a dynamic hierarchical knowledge base that is injected back into worker prompts.
The system also ingests seller listing updates, appeals, customer returns, and negative reviews to reinforce the learning loop.

Technologies

Amazon Bedrock Amazon Bedrock AgentCore Amazon Bedrock Runtime Amazon EC2 Amazon SQS

The article says error and disagreement rates fell continuously as learnings accumulated.
Costs decrease because supervisor calls are selective, while quality increases across millions of products.
The system scales catalog accuracy improvement without retraining.

Architecture

A self-learning generator-evaluator architecture uses parallel worker models for routine extraction. Disagreements are routed to a supervisor agent running on Amazon Bedrock AgentCore. The supervisor uses specialized tools, generates generalized learnings, stores them in a hierarchical knowledge base, and feeds those learnings back into worker prompts. The system also incorporates human review via Amazon SQS, runs worker models on Amazon EC2 GPU instances or Amazon Bedrock Runtime, and uses Amazon Bedrock for model access.

Sources & evidence1

Evidence: Low25/100Evidence strength