AWS has published a reference design for checking large collections of leases against changing regulations, using AI for the conversation but not for the compliance verdict. Its key safeguard is a deterministic rules engine that accounts for every record, including unreadable or ambiguous ones, rather than letting a model quietly narrow the field.
AWS AI Watch analysis
What happened
The Adjudicated Query pattern pairs Amazon Quick’s chat interface with a fixed set of operations backed by a rules engine. AWS uses a hypothetical operator with 50,000 leases as its example, and provides a working sample implementation using synthetic data.
A user can ask compliance questions in natural language, but the model translates them into calls on six bounded tools. It cannot write SQL, choose which leases count or make a pass-or-fail determination. Rules are stored as versioned data, so a change to a legal requirement is represented as a rulebook edit rather than a code change.
For a formal sweep, the system assigns each record to one of four buckets: compliant, in breach, ambiguous or unreadable. A completeness receipt checks that those counts add up to the total scanned before results are saved. The full results sit in a dashboard; the chat response carries the counts and receipt. AWS says Amazon Bedrock is used only for exploratory clause search, not for compliance sweeps or official determinations. Read AWS’s design and sample.
Why it matters
This is an attempt to give people a conversational way into a high-stakes workflow without asking a generative model to decide what the rules mean or which records to count. That distinction matters: a fluent answer is no comfort if the system silently missed leases or cannot show which rule version it applied.
The pattern also makes the limits of the chat interface explicit. Exploratory search can return a ranked sample, but it cannot answer an official question about how many records comply. For a formal result, the rules engine does the accounting and the dashboard exposes the underlying findings.
Our read
The useful idea here is not “AI checks leases”. It is that AI can help users ask questions while a deliberately narrower system handles the decision and proves what it counted. That is a more credible division of labour than giving a chatbot a database connection and hoping the important rows survive the conversation.
AWS presents this as a reference pattern and runnable sample, not evidence of a production deployment or independently tested compliance outcomes. Teams considering it should examine the rulebook, extraction process and audit trail against their own obligations before treating the receipt as a guarantee.
What to watch
- Whether teams can independently inspect and validate the rule versions and findings behind a receipt.
- How the design handles changes to regulations, lease extraction errors and disputed clauses in real deployments.
- Whether the separation between exploratory search and formal determinations remains clear to users.
Discussion spark: For high-stakes compliance, is a deterministic rules engine with a chat interface the right balance, or should the conversational model have a larger role in interpreting the rules?
Sources and evidence
- Sweep thousands of leases for compliance using Amazon Quick and the Adjudicated Query pattern (2 October 2026, 15:48 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)