Step 1 — The problem: manual review doesn't scale

Before building this system, each AWS cost review for a client took 1–2 days of manual work. The process usually went:

• Open AWS's cost analysis tool, filter by each service, take screenshots
• Log into the console and manually inspect every resource type (volumes, snapshots, network endpoints...)
• Compile everything into a spreadsheet and write the report

The issues: reliance on the reviewer's experience, no consistent standard, and hard to scale across clients — every review started from scratch, with no guarantee of full, even coverage.

Case anonymized — all absolute figures and client resource names are hidden.

Step 2 — Building the automated system (Python + AI)

I designed the workflow around 3 principles:

1. Fixed framework, changing data — always review against the same 7 pillars and a consistent way of classifying waste
2. Machines do the repetitive work, humans do the judgment — scripts collect & compute automatically; humans verify and make the call
3. Evidence, not guesswork — every conclusion traces back to a specific data source

The full workflow — Python handles data, AI handles interpretation & drafting, humans keep the decisions.

Step 3 — The 7-pillar review framework

Every engagement is reviewed against the same 7 pillars, so nothing gets missed:

Step 4 — Pilot on a real client

Applied to the first engagement — a production AWS environment in a US region (3 months of billing analyzed):

Cost breakdown by architecture group (3 months combined):

Figures anonymized — only percentages shown, no absolute amounts or client resource names.

Step 5 — Key findings

EBS volumes — $218–300/month in savings

• Finding: 27 problematic volumes — orphaned ones (attached to nothing), and production databases still on the older, pricier volume type (gp2) instead of the newer gp3 at the same performance.
• Recommendation: Delete orphaned volumes, migrate databases to the cheaper newer volume type.

aws ec2 modify-volume --volume-id vol-xxxx \
  --volume-type gp3 --iops 3000 --throughput 125

Long-stopped servers still billing for storage

• Finding: A common cost trap — a stopped server (EC2) no longer bills for compute, but its attached volumes (and static IP, if any) keep billing continuously. Many machines had been off for months with nobody noticing, because the bill never "complains".
• Recommendation: For machines stopped long enough → consider deleting the machine and its volumes. For recently stopped ones → confirm with the owner first, to avoid deleting something intentionally paused.

Network endpoints (VPC Endpoints) — $30–60/month in savings

• Finding: 10 private connections to AWS services costing ~$128/month, some of them barely used.
• Recommendation: Keep the essential endpoints, review and remove the rarely used ones.

Cost governance — indirect impact

• Finding: No budget alerts, no mandatory resource tagging, no one reviewing bills regularly.
• Recommendation: Enable budget alerts and build a tagging policy to catch abnormal cost increases early.

Step 6 — Every recommendation is backed by evidence

The guiding principle: no cost-cutting proposals based on guesswork. Every recommendation is cross-checked across multiple data sources (billing, AWS advisory tools, actual resource configuration) and its real impact verified before it enters the report.

• Check the relationships between resources before recommending deletion — e.g. a snapshot may still be backing a machine image — to avoid breaking a running system.
• Estimate savings conservatively, with specific resource IDs so the engineering team can act immediately.
• Classify clearly: optimizable / accepted (due to business requirements) / needs further monitoring.

snapshot has no source volume?
        │
        ▼
registered by any AMI?  ── YES ─▶ KEEP (not waste)
        │
        NO
        │
        ▼
has a backup marker (AWS Backup / DLM)? ── YES ─▶ KEEP
        │
        NO
        │
        ▼
   deletion candidate (awaits human confirmation)

Results

Lessons learned

1. Automate for speed, keep humans for correctness — the machine handles collection & computation, the expert handles judgment & verification.
2. A fixed framework prevents blind spots — 7 pillars ensure full, consistent coverage across every engagement.
3. Every recommendation is backed by evidence — cross-check multiple data sources before proposing.
4. Governance matters as much as optimization — budget alerts & tagging catch cost spikes early and prevent recurrence.
5. Measurable and actionable results — conservative savings estimates with concrete actions for the engineering team.