-
Notifications
You must be signed in to change notification settings - Fork 0
Business Case
Status: v1.0 · Last updated: 2026-07-23
Companion to: governance.md, adr/0001…0009, diagrams/nwaf-agent-hld.png
All figures below are illustrative and parameterised — they state their assumptions so an organisation can substitute its own rates. The case does not rest on precise numbers; it rests on the structural shift from point-in-time to continuous review.
Manual AWS Well-Architected (WA) reviews are infrequent, expensive, inconsistent, and stale the moment they finish. The NWAF Agent replaces the mechanical detection in a WA review with a deterministic nightly pipeline, reserving human architects for judgment. It shrinks the risk-exposure window from months to ~24 hours, produces evidence-backed, reproducible findings, and scales across a fleet at near-zero marginal cost — for an illustrative ~$1–2K per workload per year versus ~$6–20K for a far less frequent manual review.
It does not replace the architect; it removes the config-checking toil so the architect's time goes to design and the judgment pillars a scanner cannot assess.
A conventional WA review is a point-in-time, manual engagement:
- Infrequent. Typically quarterly at best, often annual or one-off. Between reviews, posture is unknown.
- Stale on delivery. Infrastructure drifts the day after the review; the report describes a state that no longer exists.
- Labour-intensive & senior. Prep, evidence-gathering, the review session, and write-up consume scarce senior-architect / Solutions-Architect time.
- Inconsistent & unreproducible. Different reviewers, different depth, no machine-checkable evidence, no run-to-run comparability.
- Drift-blind. Out-of-band changes (someone toggles off encryption in the console) go undetected until the next review — if ever.
The real cost is not the review fee; it is the exposure window — the months a High-Risk Issue sits undetected.
Per manual review, one workload
| Driver | Assumption | Cost |
|---|---|---|
| Senior architect / SA effort | 20–40 h @ $150–250/h loaded | $3,000–$10,000 |
| Frequency | 1–2× per year (optimistic) | — |
| Annual, per workload | ~$6,000–$20,000 |
…and that spend still leaves 6–12 month blind windows and no continuous evidence. A paid partner WA engagement typically lands in the same range or higher, per workload.
The unpriced cost: a single misconfiguration-driven incident (a public bucket, an unencrypted database, a missing backup) can exceed the entire annual review budget many times over. Continuous detection is primarily an exposure-window reduction play.
A deterministic nightly pipeline (see governance.md §4 — it passes its own WA
review) that:
- Runs every night, cutting the exposure window from months to ~24 h.
- Detects deterministically (Prowler, Checkov, Steampipe, Infracost, Compute Optimizer) and reconciles live-vs-code to classify drift — catching out-of-band changes manual reviews miss.
- Reserves the LLM for judgment — reconciliation, WA-pillar mapping, prioritisation, remediation narrative — with findings evidence-linked and scanner-fact vs judgment clearly labelled.
- Keeps the WA Tool posture current automatically (scanner-backed answers, to a dedicated automation workload; §ADR-0006).
- Frees senior talent from config-checking toil for design and the soft pillars.
- Scales to a fleet at near-zero marginal cost.
Per nightly run, one workload (illustrative)
| Component | Basis | Cost / run |
|---|---|---|
| CodeBuild compute | ~30 min, general1 medium/large | $0.10–$0.50 |
| Bedrock inference | reconcile over structured findings (compact), Sonnet-tier | $1–$3 |
| Scanners | run inside CodeBuild; API calls free | included |
| RAG retrieval | bundled reference — no always-on cost (ADR-0005) | $0 |
| S3 / SNS / adapters | logs + one thin publish | <$0.10 |
| Per run | ~$2–$4 | |
| Per workload / year | ~365 runs | ~$0.7K–$1.8K |
One-time build: a few engineering-weeks for the pipeline, adapters, and IaC. Amortised across the fleet — the **marginal cost of the Nth workload is ~$0 build
- ~$1–2K/yr run.** Fleet economics strongly favour the agent: manual reviewing a large fleet is prohibitive; the agent's cost is linear and small.
| Manual review | NWAF Agent | |
|---|---|---|
| Cost / workload / yr | ~$6K–$20K | ~$1K–$2K |
| Frequency | 1–2× / yr | ~365× / yr |
| Exposure window | months | ~24 h |
| Evidence / reproducibility | low | high |
| Drift detection | none | yes |
| Scales to fleet | poorly | linearly |
Even on cost alone the agent is ~3–10× cheaper while delivering ~200× more frequent coverage. The dominant return, however, is the shortened exposure window — one prevented incident typically repays years of run cost. Add the recovered senior-architect hours (redirected to design) and the case compounds.
| Risk | Mitigation | ADR |
|---|---|---|
| False positives / inaccuracy | Deterministic scanners + evidence-linkage; human disposition | 0004, 0008 |
| LLM hallucination | LLM reconciles, never invents findings; no auto-remediation; no fabricated "fixed" status | 0003, 0008 |
| Trust in automated WA writes | Dedicated automation workload; human record untouched; reversible milestones | 0006 |
| Security of the reviewing tool itself | Read-only, write structurally impossible in target; per-stage least-privilege | 0007 |
| Over-claiming coverage | Performance / process-Ops / Sustainability labelled as judgment, not scanner fact | 0004 |
| Adoption friction |
report_only default; opt-in advisory PRs; accepted-risk register |
0008 |
| Alert fatigue | Idempotent one-ticket-per-finding; accepted-risk suppression | 0008, 0009 |
| Silent run failure | Open gap — run-failure alerting not yet designed (see governance.md §4) | follow-on |
| Maintenance drift | Pinned scanner versions; corpus refresh per WA revision | 0004, 0005 |
Coverage & timeliness
- Workloads/accounts under nightly review (target: full fleet).
- % of WA questions answered with scanner-backed evidence.
- Mean time-to-detect a new HRI (target: < 24 h, vs months).
Risk reduction
- Open-HRI trend (should decline over time).
- Out-of-band drift findings caught per month.
- Mean time-to-remediate HRIs.
Efficiency
- Senior-architect hours saved vs the manual-review baseline.
- Cost per workload-review (agent vs manual).
Quality & adoption
- % findings actioned (ticket resolved / PR merged).
- False-positive rate (findings dismissed).
- Run success rate (% nights with a completed review).
- Teams opted into
advisory_pr; count of accepted-risk exceptions (a healthy governance signal, not a failure).
- Continuous, audit-ready compliance evidence rather than periodic snapshots.
- Consistency & reproducibility — the same inputs yield the same review.
- Fleet scale — a per-workload practice that was economically impossible manually.
- Talent leverage — senior architects move from checking to designing.
- Cultural shift — Well-Architected becomes a continuous practice, not an event.
Adopt the NWAF Agent, rolled out in report_only mode first to establish trust,
expanding to advisory_pr per team as confidence grows. Prioritise closing the
run-failure alerting gap before fleet-wide rollout. Track the KPIs in §8 from
day one to make the exposure-window and efficiency gains visible.
-
governance.md(framework mapping, trade-off register, AI controls) -
adr/0001…0009(decision records) -
diagrams/nwaf-agent-hld.png(v2.3)
NWAF Agent repository · generated from docs/ — do not edit wiki pages directly.
Design
Decision Records