What has AWS published?
AWS describes how to build a semantic layer using EDB Postgres AI, with a Strands Agents agent running on Amazon Bedrock AgentCore. The example answers customer 360 questions, such as which customers spend the most and which products sell the most units. This is a technical walkthrough, not a product launch announcement.
According to the AWS walkthrough, the AIDB extension keeps business meaning inside the operational Postgres database. Comments explain fields, foreign keys describe relationships, and views hold shared business definitions. A semantic knowledge base, or SemKB, makes that information searchable by meaning.
The architecture connects the agent to AIDB through its Model Context Protocol (MCP) endpoint. Postgres permissions govern database access. AWS also describes tracing agent activity and recording SQL execution traces for review.
Why should a smaller business care about a semantic layer?
An AI-generated query can be technically valid and still answer the wrong business question. AWS illustrates this with a price field that contains an entire order’s total, rather than a unit price. Treating it as a unit price and multiplying it by quantity produces an incorrect customer-spend calculation.
For an owner or manager, the broader issue is consistency. Terms such as active customer, overdue invoice or profitable order need agreed definitions before an AI assistant uses them. Otherwise, different questions may produce answers built on different assumptions.
AWS’s example stores its definition of a big spender in a database view. The practical lesson is to maintain important definitions in a shared, controlled place rather than asking each employee or AI prompt to recreate them. That preparation is useful even if you choose a different technical approach.
How should access and answer quality be checked?
AWS places access enforcement in Postgres rather than relying on instructions to the model. Its example defines separate database roles, including one intended to exclude access to customer Social Security numbers. Column-level permissions and row-level security determine which data a caller can retrieve.
For a small or medium business, start by deciding what each role genuinely needs to see. A person asking for a sales summary should not automatically receive access to every customer field. Map those needs to database permissions with a qualified technical lead, and explicitly test requests for restricted information.
Check correctness separately from access. Compare AI answers with approved queries, including awkward cases such as missing addresses, repeated purchases and ambiguous business terms. AWS describes SQL execution tracing through AIDB; use available records to investigate how an answer was produced, not just whether it sounded plausible.
What practical steps should a business take next?
Choose one narrow reporting task before attempting a broad AI rollout. Write down the questions it must answer, the approved calculations and the fields it may use. Ask a technical lead to document confusing columns, verify table relationships and create shared definitions for recurring terms.
Run a read-only pilot with limited permissions. Keep a set of expected answers and test both permitted and forbidden requests. Before committing, ask for an assessment of setup work, ongoing maintenance and total cost. The AWS walkthrough does not establish what this deployment would cost your business.
If your wider goal includes doing work in applications rather than only querying data, assess that as a separate requirement. Terabot gives businesses digital workers: AI workers with their own private virtual machine that work in the apps a team already uses, around the clock. It is in private preview, by invitation only.
Key takeaways
- AWS’s walkthrough combines business definitions in EDB Postgres AI with an agent running on Amazon Bedrock AgentCore.
- Clear data definitions matter as much as query generation: a misunderstood field can produce a convincing but incorrect answer.
- Start with a narrow, read-only pilot, enforce database permissions and compare answers against approved calculations.
Written with AI assistance from the source linked above, and checked against it before publishing. Product names belong to their owners. Check the original source before relying on details.