AWS shows how an AI assistant can answer questions across two databases

AWS has published a worked example of an AI assistant that answers business questions using two separate databases, with access rules enforced by the data layer instead of the assistant. It is the third post in an AWS series on a "semantic layer," a shared set of business definitions that sits between the AI and the data. This one uses Strategy Mosaic.

What the example does

A retail company keeps customer records, addresses, credit cards and rewards enrollment in Amazon Aurora. Orders and products sit in Amazon Redshift. The sample question: which of our top 20 customers by lifetime spend have not joined the rewards program? Answering it needs both systems. The two link through a shared customer ID called cid, and Mosaic handles that match.

How it works

An AI agent built with Strands runs on Amazon Bedrock AgentCore and uses Anthropic Claude Sonnet 4.6 to understand the question. It then reaches Mosaic through MCP, a standard way for AI tools to connect to data tools. The joins, metric definitions and per-user filters live in Mosaic, not in the instructions given to the agent.

Mosaic does not merge rows across databases on the fly. It sends one query to each source, then combines the results in its own engine using the shared attributes the modeler has linked.

Who sees what

Each request carries the identity of the person asking. Mosaic checks that identity and a permission to use its MCP server. It also checks that the model is certified, then applies row-level and column-level security for that user. In plain terms, two people can ask the same question and get answers limited to what each is allowed to see.

Why it matters

Many businesses want an AI assistant over customer and billing data. Putting access rules in the data layer means they do not depend on the assistant behaving well. Mosaic can also serve the same definitions to Power BI, Excel, SQL tools and custom apps, so numbers can stay consistent.

The example builds on the Strategy One release from March 2026, when Mosaic access through MCP became available by default. Earlier posts in the series did the same job with Stardog and Denodo.

What is not covered here

The material is a technical walkthrough, not a product review. It gives no cost, speed or accuracy figures. Anyone considering this should test it on their own data and systems.