AWS details design pattern to verify compliance across thousands of leases using Amazon Quick and a deterministic rules engine

AWS describes a design pattern called Adjudicated Query that pairs Amazon Quick, an AI-powered chat assistant, with a deterministic rules engine to address compliance challenges at scale. The pattern is designed to ensure two critical properties: provable completeness and defensibility. Provable completeness means that claims like 'we checked all 22,910 Texas leases' must be verifiably true, with no leases silently omitted. Defensibility requires that any compliance finding can be traced to the specific version of the rule applied, the clause text evaluated, the method used, the date of assessment, and the responsible party.

The post highlights the limitations of traditional approaches such as Retrieval Augmented Generation (RAG) and text-to-SQL in meeting these requirements. RAG, while improving accessibility, cannot guarantee completeness because similarity search lacks a mechanism to ensure all relevant documents are included. Text-to-SQL, while narrowing the gap, risks hallucinated predicates that silently reduce the scope of evaluation, producing exact-looking numbers that are methodologically unsound.

The Adjudicated Query pattern solves these issues by introducing a bounded conversational layer over a deterministic rules engine. In this setup, the model’s role is strictly limited to translating natural-language compliance questions into calls on a fixed set of typed operations and narrating the results returned by the rules engine. The model does not write queries, alter the scope of evaluation, or perform determinations. The rules engine itself operates on versioned rules data, using generic comparison operators to evaluate lease clauses against current regulations.

The pattern is illustrated through a reference architecture and a deployable AWS CDK sample, with lease compliance serving as the primary use case. However, AWS notes that the pattern is applicable to other high-stakes compliance domains, including sanctions screening, insurance claims adjudication, and export control. The post provides a concrete example of a portfolio operator managing 50,000 leases across multiple states, where state-specific landlord-tenant laws frequently change, necessitating a scalable and verifiable compliance process.