Well-Architected reviews die of clerical load: dozens of lens questions, each needing evidence nobody has gathered. This playbook has AI walk the question set, pull evidence from IaC and docs, and draft findings — while the architect validates every piece of evidence and personally owns every risk rating.
Tier: Core · Hands-on time: ~35 minutes
Lens question sets are structured, exhaustive, and evidence-hungry — the ideal shape for AI walkthrough. Since AWS shipped the Generative AI Lens in 2025 and added agentic and responsible-AI lenses at re:Invent 2025, an AI workload review means working through pillar questions most teams have never seen before. The division of labor that keeps it honest: the model reads the artifacts and drafts "here is what the evidence suggests"; the human confirms the evidence is real and decides what is actually a risk. AI-drafted findings carry a meaningful false-positive rate — rejecting bad findings is as much the job as accepting good ones.
Skip AI when the workload has no reviewable artifacts — a review of a system that exists only in someone's head needs interviews first, not a model.
Scope the review to one workload and one lens pass. "Review everything" produces mush. Anita scopes: orbit-docs-copilot, GenAI lens plus the reliability and cost pillars of the base framework, against the infra/ repo, the HLD, and the eval documentation.
Walk the questions with an evidence-first prompt.
You are assisting a Well-Architected review of orbit-docs-copilot
(customer-facing RAG copilot). Artifacts provided: Terraform in infra/,
HLD.md, eval-plan.md, runbook.md.
For each lens question I paste, produce:
1. EVIDENCE — verbatim quotes/snippets from the artifacts bearing on
the question, each with file and line/section. If nothing bears on
it, write "NO EVIDENCE FOUND" — that is itself a finding.
2. DRAFT FINDING — what the evidence suggests (met / risk / gap),
in two sentences.
3. PROPOSED SEVERITY — with your reasoning, clearly marked PROPOSED.
4. WHAT WOULD CHANGE THIS — the single artifact or console check that
would confirm or refute the draft finding.
Do not mark anything as resolved based on intent stated in docs alone;
distinguish "documented" from "implemented in IaC".
###
[paste 3–5 lens questions at a time]
Validate evidence before reading findings. For each drafted item, check the citation first: does the quoted Terraform exist, and does it mean what the model says? The "documented vs implemented" distinction earns its place here — models routinely credit an HLD sentence ("all data encrypted at rest") as if it were a deployed control. Evidence that fails validation kills its finding regardless of how plausible the finding sounds.
Run the WHAT WOULD CHANGE THIS checks. These console/CLI checks are where the review touches reality: the drafted finding says throttling is unconfigured — the service console says otherwise, or confirms it. Ten minutes of console time converts a draft from prose into a review. NO EVIDENCE FOUND items get a human decision: genuinely unaddressed (a gap finding), or addressed somewhere the model could not see (attach the evidence yourself).
Rate risks yourself, and reject freely. Fill the DISPOSITION column: accept (with your severity — the model's PROPOSED is anchoring, not authority), reject (record why: false positive, out of scope, compensating control), or investigate (owner + date). Expect to reject a real fraction of drafted findings — a review where every AI finding survives means the human pass did not happen. The rejection log is itself evidence of review rigor, and it is what you will defend at the ARB, so write the reasons as if Vikram will read them.
File the review in the tool of record. Transfer accepted findings into the WA Tool (or your framework tracker) so remediation tracking, milestones, and the improvement plan live in the system your org already audits. The AI conversation is scaffolding; the tool entry is the review.
Review artifacts are the client's crown jewels — IaC, HLDs, runbooks. Run the walkthrough inside the approved boundary: Claude on the client's Bedrock/Azure OpenAI tenancy, or a local model over a cloned repo for prohibited-egress engagements. The lens PDFs are public and travel freely; console verification happens in the client's accounts under their access controls regardless. Where no AI may read the artifacts at all, the playbook degrades gracefully: use the model only to plan the review (question sequencing, evidence checklists per question — no client content in the prompt) and gather evidence by hand.
AI turns the lens walkthrough from a day of clerking into an hour of judging. The judging — evidence validation, risk ratings, rejections — is the review, and it never delegates.