AWS SMGS built NarrateAI, a conversational agentic AI assistant powered by Amazon Bedrock AgentCore, to deliver on-demand business intelligence to leaders across the organization. The system uses a two-layer architecture combining batch narrative generation with real-time multi-agent query processing, enabling natural language access to complex business data while maintaining row-level security and role-based personalization.
Results at a glance · every figure cited
Time-intensive data preparation: AWS leaders spent hours manually gathering data from fragmented dashboards before business reviews, leaving insufficient time for strategic reasoning.
Data fragmentation and accessibility: Business insights scattered across multiple systems created inconsistencies. Complex dashboards required specialized knowledge, forcing dependency on intermediary reporting teams and delaying critical decisions.
Two-layer architecture: Batch narrative generation (SQL extraction → JSON transformation → Jinja templating) pre-processes business data into persona-specific narratives stored in Amazon S3. Real-time conversational layer uses Amazon Bedrock AgentCore with six specialized agents for question classification, persona identification, knowledge extraction, relevancy evaluation, answer generation, and response validation.
Key design patterns: Table-of-contents-based retrieval, hierarchical chunking, row-level security through full data isolation, and parallel sub-task decomposition for complex queries. Supervisor Agent orchestrates workflow; Claude Sonnet 4 powers reasoning.
Deployment accelerated from months to weeks via serverless architecture and built-in orchestration. Leaders gain immediate, context-rich answers to natural language business questions. System maintains production-quality observability through native CloudWatch integration and automated session management.
NarrateAI supports three core capabilities: natural language queries across multi-dimensional data (regions, products, segments, time), inherent row-level security through user-specific narratives, and role-tailored responses (CEO strategic insights vs. regional manager operational metrics).
Separating batch narrative generation from real-time agent interaction enables secure, scalable conversational BI without building custom orchestration infrastructure.
AWS SMGS's architecture solves a real organizational problem—executive data access bottlenecks—by moving complex ETL and security enforcement into the batch layer, freeing the real-time agent to focus on reasoning and orchestration. This design pattern transfers well to enterprises with stable data schemas and user hierarchies; the table-of-contents retrieval strategy is a practical alternative to RAG-based chunking when source data is structured and role-based access is non-negotiable.
The case study emphasizes deployment velocity and operational simplicity (serverless, managed agent orchestration, CloudWatch integration) over algorithmic innovation. The six-agent decomposition (question classification, persona ID, knowledge extraction, relevancy, generation, validation) mirrors production patterns AWS promotes across other Bedrock AgentCore implementations. Notably absent: quantified latency improvements, cost comparisons to legacy BI tooling, or ground-truth accuracy validation metrics—suggesting this is a proof-of-concept showcase rather than a mature production case study.
Response times vary based on query complexity and underlying data sources (not quantified). No published metrics on query accuracy, latency targets, or production error rates. User adoption, usage frequency, and business impact (cost savings, decision speed improvements) not disclosed. System tested within AWS SMGS org only; transferability to external customers not validated.
NarrateAI is an AI-powered conversational assistant built by AWS SMGS using Amazon Bedrock AgentCore. It delivers on-demand business intelligence to organizational leaders by answering natural language questions about business performance, replacing time-consuming manual report preparation and fragmented dashboard navigation.
The batch layer (narrative generation) extracts data via parameterized SQL, transforms it to JSON, and renders human-readable narratives stored per-user in Amazon S3. The real-time layer (conversational interface) uses six specialized agents orchestrated by Amazon Bedrock AgentCore to retrieve user-specific narratives and synthesize answers with Claude Sonnet 4.
Row-level security is applied during batch narrative generation. User permissions are enforced when extracting data from Amazon Redshift, and each user's narrative file in S3 is fully isolated, preventing cross-user data leakage. This design ensures leaders only access authorized business data.
Question classifier (intent analysis, query complexity detection), Persona knowledge identifier (user role and permissions), Knowledge extractor (TOC-based S3 retrieval), Relevancy evaluator (content filtering), Answer generator (Claude Sonnet 4-powered synthesis), and Online evaluator (response validation against source data).
The Question classifier automatically breaks complex, multi-part queries into parallel sub-tasks. AgentCore's native multi-agent coordination framework orchestrates these tasks in parallel, enabling comprehensive answers to questions like 'Compare my top 5 accounts' growth rates across all product lines.'
Amazon Bedrock AgentCore (agent orchestration and foundation model hosting), Amazon Redshift (data source), AWS Lambda (data transformation), Amazon S3 (narrative storage and retrieval), Amazon Quick (conversational interface), and Amazon CloudWatch (observability and monitoring).
Analysis by
Twarx Research Team · Applied AI Research
Twarx researches and deploys enterprise AI agents with a measurement-first method: every metric traced to a primary source, projections labelled as estimates, and limitations stated up front.
From evidence to deployment
We deploy and audit AI agents the way we research them: real baselines, cited outcomes, and no inflated ROI. Twarx maps where an agent actually pays off for your team — and, just as honestly, where it won't.
The teams that win with AI agents aren't the ones with the boldest claims — they're the ones who measured honestly and deployed where the evidence actually pointed.
— Twarx Research
