Sweep thousands of leases for compliance using Amazon Quick and the Adjudicated Query pattern

AWS is pitching a way to sweep tens of thousands of apartment leases without handing the decision to a model. The company’s new pattern, called Adjudicated Query, puts Amazon Quick on top of a deterministic rules engine so business users can ask in chat while the actual compliance call stays machine-checkable and versioned.

The use case is painfully specific, which is why it works as a demo. A portfolio operator with 50,000 leases across multiple states needs to track landlord-tenant rules that change on the legislature’s schedule, not the operator’s. Late-fee caps, notice periods, security-deposit limits — if those shift, compliance teams have to identify what’s now out of line, and later prove exactly how they checked it.

AWS argues that this is where ordinary retrieval and text-to-SQL fall down. Retrieval can surface relevant text, but not prove it checked everything. Generated SQL can look precise while quietly narrowing the population. The proposed pattern keeps the model on a leash: it can translate a question into one of a fixed set of operations, and narrate the answer, but it never writes the compliance query or makes the pass/fail decision.

Behind that boundary sits a rules engine backed by Amazon Aurora Serverless v2, with rulebooks stored as versioned data rather than code. Every sweep produces a completeness receipt, where compliant, in-breach, ambiguous, and unreadable must add up to the scanned total. If that invariant doesn’t hold, the run doesn’t finish. AWS is very explicit that the point is accountability, not cleverness.

The reference architecture routes the chat agent through Amazon Cognito, Amazon API Gateway, AWS Lambda, and then into the rules engine. Amazon Quick Sight reads the same Aurora store directly for the full results table, because chat is bad at showing 10,800 rows and dashboards are bad at pretending they’re conversational. Amazon Bedrock appears only on the exploratory clause-search path, not in the official determination.

There’s also a working sample on GitHub, with synthetic data, deterministic corpus generation, and acceptance tests. AWS says the stack includes the MCP server, the rule engine, Aurora, and Quick Sight wiring, and that the corpus is deterministic so a rebuild reproduces the same results. For anyone stuck between “AI everywhere” and “no AI allowed,” this is the more grown-up answer: use the model at the edge, not at the point of judgment.