Markdown edition

# How AWS's lease review sample separates AI from compliance decisions

> AWS's sample routes lease questions through Amazon Quick to fixed tools and a rules engine. The guide explains the receipt, evidence trail, setup limits and what adopters must validate.

By BIG CHANGE Editorial

Published: 2026-10-03T08:22:25.737Z
Updated: 2026-10-03T08:22:25.737Z
Canonical: https://bigchange.ai/blog/aws-lease-compliance-adjudicated-query-guide

![Conceptual charcoal illustration of an open archive drawer filled with lease folders, one orange divider and a blank slip clipped to the drawer front.](https://bigchange.ai/api/media/file/aws-lease-review-archive-hero-v1.png)
AI-generated conceptual editorial illustration by BIG CHANGE.

AWS has published an Amazon Quick reference design for asking compliance questions across a large lease collection. Its useful idea is a strict handoff: the chat model selects a fixed tool and explains its response; a separate rules engine defines the population and makes each determination. The [October 2 post](https://aws.amazon.com/blogs/machine-learning/sweep-thousands-of-leases-for-compliance-using-amazon-quick-and-the-adjudicated-query-pattern/) and [sample repository](https://github.com/aws-samples/sample-quick-adjudicated-query) describe an educational proof of concept, not a validated legal compliance service.

For an engineer or compliance lead, the question is whether that boundary fits the decision being reviewed. The sample uses synthetic leases and invented rules and citations. Its example totals show the intended output shape; they say nothing about accuracy on real contracts or law.

## The big change

- **What changed:** AWS's sample makes AI a controlled interface to fixed review tools. A rules engine determines findings for an enumerated lease population.
- **Why it matters:** Versioned rules and an evidence receipt give reviewers a way to inspect how the selected records were assessed outside the chat.
- **What to watch:** The receipt cannot validate the inventory, extraction or legal rules. Adopters must check those inputs, user attribution and performance on their own data.

## The population comes before the prompt

A user can ask Quick which Texas leases breach a late fee rule on a stated date. Quick routes the request to `sweep_compliance`, one of six named MCP operations. The operation uses the specified jurisdiction and date to select the population and applicable versioned rules. The model does not write the SQL or determine whether a clause passes. The rule engine applies fixed comparison operators, with rule values passed as parameters. AWS says an official sweep does not consult a model. [AWS explains the operation contract here](https://aws.amazon.com/blogs/machine-learning/sweep-thousands-of-leases-for-compliance-using-amazon-quick-and-the-adjudicated-query-pattern/); the [repository describes the implementation](https://github.com/aws-samples/sample-quick-adjudicated-query).

The other tools have narrower meanings. `simulate_rule_change` gives exploratory counts for a proposed value without recording findings. `explore_clauses` ranks a filtered sample by semantic similarity and cannot answer “how many?” `get_finding` retrieves one evidence chain; `list_rules` shows rules in force on a date; `check_connection` checks transport. The distinction matters because a relevant clause sample is not a census.

The sweep writes a receipt that accounts for every scanned record in one of four categories: compliant, in breach, ambiguous, or unreadable. The engine asserts that those counts add up to the scanned total before committing. Quick can show the counts and a small sample, while a Quick Sight dashboard reads the same Aurora data store for the full findings. A finding includes the clause text, extracted and expected values, rule version and citation. These are properties of the AWS sample's design, not a claim that BIG CHANGE ran or independently verified its results.

The receipt accounts for the selected population. It cannot show that the source inventory contains every lease, that extraction captured every relevant clause correctly, or that a rule reflects current law. Those require separate reconciliation, extraction review and legal approval. If “all Texas leases” is an uncertain denominator, an exact count can still describe the wrong set.

## What you would have to build

The [sample architecture](https://github.com/aws-samples/sample-quick-adjudicated-query) puts a Quick chat agent in front of an MCP server on AWS Lambda. Amazon Cognito issues a service token; API Gateway checks it. Lambda reads and writes Aurora Serverless v2 through the RDS Data API. Quick Sight reaches the same database through a VPC connection. AWS reserves Bedrock embeddings and a language model for the exploratory clause-search tool; the official sweep remains deterministic.

The published example has a synthetic 50,000-lease corpus, a versioned rulebook and an acceptance script that AWS says runs 28 checks against a deployed stack. Its repository explicitly warns that the code is not production ready, the legal content is invented, and real tenant data needs further security testing and independent legal validation. We inspected the documentation and repository description; we did not deploy the stack, run those checks, or test the chat agent's tool routing.

To adapt it, establish the authoritative record inventory and an exact membership rule first. Then decide which fields can be extracted reliably, which rule comparisons are genuinely mechanical, and who approves each rule version. Preserve source text, extraction status, rule version, operator, compared values, date and finding ID so a reviewer can reconstruct a result. Reconcile the receipt to the inventory outside the chat response. These are design checks drawn from the sample's stated guarantees and limits, not steps we have tested.

AWS says its Cognito client-credentials token identifies the Quick application, not the person asking in chat. The sample relies on correlating a sweep ID and time with Quick's audit layer to identify the user; AWS suggests passing and storing an end-user ID if the compliance store itself must record that identity. A team that needs a finding to stand alone as an audit record should settle that design before deployment.

## Access, limits and cost

AWS's walkthrough assumes an AWS account, configured AWS CLI v2 credentials, Python and Node 24 for CDK, model access in `us-east-1`, and an Amazon Quick environment with an MCP connector and Quick Sight. The post says Python 3.12; the linked README says Python 3.9 or newer. Check the repository's current requirements when choosing a local environment. The sample pins CDK CLI 2.261.0 in its instructions, but that pin is a sample dependency, not a general AWS requirement.

The current [Quick MCP guide](https://docs.aws.amazon.com/quick/latest/userguide/mcp-integration.html) gives each operation a fixed 60-second timeout, allows at most 100 tools per server connection, and does not transmit custom HTTP headers. That timeout deserves a workload test for a large sweep; a correct long-running database job still cannot complete as a synchronous Quick operation if it exceeds the connector limit. The guide says a custom connector's tool list can be updated with **Sync**. The AWS blog and sample README instead instruct users to delete and recreate the integration after tool changes. Follow the live Quick documentation for the current connector and verify the registered tool list and routing in your own environment.

This is a multi-service build, so there is no defensible single “price per sweep” in the published material. AWS's [Quick pricing page](https://aws.amazon.com/quick/pricing/pricing/) separates subscription and agent-hour terms and lists additional Quick Sight charges for some capabilities. [Aurora pricing](https://aws.amazon.com/rds/aurora/pricing/) depends on capacity, storage and I/O configuration; the sample leaves a 0.5 ACU minimum active instead of pausing at zero. API Gateway, Lambda and any exploratory Bedrock calls also require a workload estimate. The repository suggests destroying the stack after evaluation to avoid ongoing charges. No AWS resources were created for this article.

## A decision checklist

Use the pattern only after the team can answer these questions with its own data and controls:

- Can you enumerate the full population with a reviewed, stable predicate and reconcile it to an authoritative inventory?
- Are the rules mechanical comparisons, with approved versions, effective dates and citations? Which cases must remain ambiguous for human review?
- Can extraction failures and unreadable documents be counted rather than silently excluded?
- Does each finding retain the source clause, compared values and rule version, and can you retrieve the complete evidence set outside chat?
- Can the sweep finish within Quick's operation timeout, and how will you trace a result to the person who requested it?
- Has the team estimated subscription, database and service charges for its expected volume, then tested performance and result quality on permitted data?

If the task needs legal interpretation or judgment about an open-ended standard, a deterministic pass/fail label may hide the very decision that needs a human. If the team wants representative examples, semantic retrieval is simpler. If a reviewed rules engine and dashboard already serve accountable users, the conversational layer is optional. Those alternatives follow from the AWS post's own “wrong choice” discussion and the limits of its synthetic sample.

## Sources & further reading

- [AWS Machine Learning Blog, “Sweep thousands of leases for compliance using Amazon Quick and the Adjudicated Query pattern”](https://aws.amazon.com/blogs/machine-learning/sweep-thousands-of-leases-for-compliance-using-amazon-quick-and-the-adjudicated-query-pattern/) (October 2, 2026). The vendor's architecture, operation semantics, sample walkthrough and stated failure boundaries. Example counts and behavior are AWS's descriptions, not independent measurements.
- [AWS sample repository](https://github.com/aws-samples/sample-quick-adjudicated-query). README, file map, setup assumptions and explicit synthetic-data, invented-law and production-readiness warnings. The code was not deployed or tested for this article.
- [Amazon Quick MCP integration guide](https://docs.aws.amazon.com/quick/latest/userguide/mcp-integration.html). Current connector setup, Sync behavior and operation limits. Its update path differs from the post and README.
- [Amazon Quick pricing](https://aws.amazon.com/quick/pricing/pricing/) and [Amazon Aurora pricing](https://aws.amazon.com/rds/aurora/pricing/). Current pricing structures and variables for estimating an actual workload; neither supplies a complete price for this sample.

## Sources

- [Sweep thousands of leases for compliance using Amazon Quick and the Adjudicated Query pattern](https://aws.amazon.com/blogs/machine-learning/sweep-thousands-of-leases-for-compliance-using-amazon-quick-and-the-adjudicated-query-pattern/) — Official vendor post describing architecture, six MCP tools, sample outputs, setup and limits. Its example is synthetic and is not BIG CHANGE testing or independent validation.
- [aws-samples/sample-quick-adjudicated-query](https://github.com/aws-samples/sample-quick-adjudicated-query) — Official sample README, file map and prerequisites. Explicitly warns that the data and legal rules are synthetic or invented, code is not production ready, and legal/security validation is needed. Not deployed or tested by BIG CHANGE.
- [Model Context Protocol (MCP) integration - Amazon Quick](https://docs.aws.amazon.com/quick/latest/userguide/mcp-integration.html) — Current connector documentation: 60-second operation timeout, 100-tool maximum, no custom headers, and Sync for custom connector tool updates. The Sync instruction differs from the sample's delete/recreate instruction.
- [Amazon Quick pricing](https://aws.amazon.com/quick/pricing/pricing/) — Official plan, agent-hour and Quick Sight pricing structure; does not price the entire reference stack or a sweep.
- [Amazon Aurora pricing](https://aws.amazon.com/rds/aurora/pricing/) — Official variable Aurora capacity, storage and I/O cost structure. Sample README specifies a 0.5 ACU active minimum. No all-in workload estimate was made.