Automating AWS Well-Architected Reviews at Scale with GenAI

commbank · Finance · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Cloud architecture review

CommBank and AWS needed to make AWS Well-Architected Reviews feasible across many workloads instead of only the most critical ones, because traditional review meetings consumed hours from many subject-matter experts.

Constraints:

traditional reviews take three to four hours
reviews require ten to fifteen subject-matter experts
review context lives in diagrams, CloudFormation, Terraform, and procedural documents
review coverage should span design, build, migration, and post-migration lifecycle stages
financial-sector workloads need lens-specific and policy-specific context

Architecture

AWS and CommBank built a Well-Architected Infrastructure Analyzer. Users upload architecture diagrams, CloudFormation templates, Terraform files, and supporting procedure documents. The application lets the user choose a full Well-Architected review, selected pillars, or lenses such as the financial-sector lens. Amazon Bedrock maps resources and documentation to Well-Architected best practices and returns a report listing adhered-to practices, gaps, and recommendations.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  workload-artifact-uploader[Workload artifact uploader] --> review-scope-selector[Review scope selector]
  workload-artifact-uploader --> bedrock-analysis-engine[Bedrock analysis engine]
  review-scope-selector --> bedrock-analysis-engine
  bedrock-analysis-engine --> review-report-generator[Review report generator]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Workload artifact uploader`	Accepts diagrams, CloudFormation templates, Terraform files, and procedural documentation.	`Review scope selector`, `Bedrock analysis engine`
`Review scope selector`	Lets users choose full reviews, pillars, or Well-Architected lenses.	`Workload artifact uploader`, `Bedrock analysis engine`
`Bedrock analysis engine`	Maps uploaded resources and documentation to Well-Architected best practices.	`Workload artifact uploader`, `Review report generator`
`Review report generator`	Reports adhered-to practices, missing practices, and recommendations.	`Bedrock analysis engine`

Data flow

The review starts with workload artifacts: diagrams, CloudFormation/Terraform templates, and organizational procedure documents. The analyzer collects those inputs and the selected review scope, sends the combined context to Bedrock, maps resources and documents to Well-Architected best practices, and returns a report of adherence and recommendations.

4 supporting quotes

“you probably have a diagram or even if you are in a build phase already build a workload you probably have a cloud for template or terraform file”transcript, stated

“provided the functionality where you can upload any of these documentation about your standard procedures that you cannot”transcript, stated

“what happens in the background the application take all of the documentation you uploaded and the diagram or the clformation template and”transcript, stated

“bedrock will map all of these resources and documentation to the well architected best practices and then it will come back with a report.”transcript, stated

Why this architecture

The architecture works because Well-Architected reviews require both resource-level cloud configuration and organization-specific procedures. Uploading templates/diagrams gives Bedrock infrastructure facts, while procedure documents supply context not present in IaC. Selecting pillars or lenses lets the same analyzer run targeted reviews instead of always running the full framework.

3 supporting quotes

“the cloud for template or the diagram have the services have the resources the configuration but it won't really have the context about your procedures”transcript, stated

“what do you do for example from cloud financial management practices what do you do from an operational practices”transcript, stated

“you go and then choose are you want do you want to run full well architected or certain pillars of the well architected”transcript, stated

Models used: Amazon Bedrock

Scale

Traditional reviews averaged three to four hours.
“Normally well architected would take three four hours and usually probably 10 to 15 subject matter experts both from AWS and from”transcript, stated
Traditional reviews usually required ten to fifteen subject-matter experts.
“and usually probably 10 to 15 subject matter experts both from AWS and from the customer. So in this case from us”transcript, stated
The intended coverage is all workloads, not just critical workloads.
“other customers really to scale well architected reviews across all of their workload not just the most critical ones”transcript, stated
The workflow targets multiple lifecycle stages: pre-migration, post-migration, design, and build.
“embed well architected review in all the life cycle of the workload pre-migration post migration when you design when you build”transcript, stated

Alternatives considered (roads not taken)

Traditional SME meeting reviews

Rejected because: Three-to-four-hour meetings with 10-15 busy experts were a logistics bottleneck.

“any of you who tried to create a 4 hours meeting with 10 to 15 very busy individuals know how much of a logistic nightmare it is.”transcript, stated

Run reviews only for critical workloads

Rejected because: The objective was to streamline reviews across all workloads.

“streamline it so instead of running it in just the most critical workload run it across all of the workloads. So we”transcript, stated

Analyze only infrastructure templates

Rejected because: Templates and diagrams do not include operational, financial-management, or cyber checklist procedures.

“the cloud for template or the diagram have the services have the resources the configuration but it won't really have the context about your procedures”transcript, stated

Patterns

`tool-use-orchestration` (workflow)

The analyzer combines uploaded IaC artifacts, diagrams, procedure documents, and framework/lens selection into a Bedrock-backed review workflow.

“you upload this uh cloud for template or the diagram to the well architected infrastructure analyzer applications”transcript, stated

`rag-retrieval` (retrieval)

Uploaded standard-procedure documents provide contextual knowledge not present in the infrastructure artifacts.

“provided the functionality where you can upload any of these documentation about your standard procedures”transcript, stated

`structured-output-contract` (reliability)

The analyzer returns a structured review report of adhered-to practices, missing practices, and recommendations.

“The report will go and tell you these are the best practices that you adhere to and these are the best practices you haven't adhere to”transcript, stated

`batch-offline-processing` (operations)

The solution is intended to run reviews across all workloads instead of relying on scheduled expert meetings for selected critical systems.

“streamline the well architected review across all of your workload”transcript, stated

Tradeoffs

Automate first-pass framework mapping

Chose Bedrock-backed analyzer reports over expert-only review meetings for every workload.

Rationale: Automated mapping made it plausible to run reviews across all workloads, while experts could focus on interpreting and acting on findings.

“to scale well architected reviews across all of their workload not just the most critical ones we need to do something different.”transcript, stated

Accept user-uploaded procedure documents

Chose templates/diagrams plus operational documents over templates/diagrams alone.

Rationale: Review quality depends on organizational practices that are absent from resource templates.

“you can upload any of these documentation about your standard procedures that you cannot find in a diagram or cloudformation template”transcript, stated

Gotchas (what bit them)

Infrastructure artifacts lack procedural context.

Root cause: CloudFormation, Terraform, and diagrams show resources and configuration but not operating practices.

Resolution: The analyzer supports uploads of standard-procedure documentation.

“it won't really have the context about your procedures meaning what do you do for example from cloud financial management practices”transcript, stated

Traditional Well-Architected meetings are hard to schedule at scale.

Root cause: Each review requires several hours with many experts.

Resolution: The team built a GenAI analyzer to automate mapping and reporting.

“create a 4 hours meeting with 10 to 15 very busy individuals know how much of a logistic nightmare it is.”transcript, stated

“is there any way we can use AI or and geni to optimize time of our and your SMS.”transcript, stated

Evaluation

The talk demonstrates the analyzer workflow and frames success as reducing time/resources while expanding reviews across workloads; it does not provide quantitative accuracy metrics.

Metrics: review duration · SME time required · workload review coverage · recommendations generated

“How long did it take you? But average 3 four hours.”transcript, stated

“run well architected review at a scale on all of your workload and you don't need that much time or resources anymore.”transcript, stated

Outcomes

AWS published the solution to GitHub for customers to download.
The analyzer maps workload artifacts and documentation to Well-Architected best practices.
The intended outcome is broader review coverage across all workloads.

The transcript emphasizes workload-wide review coverage and three-to-four-hour, 10-15-person baseline meetings, but does not report measured post-automation duration.