An AWS bill rarely has one cause. It accumulates: instances that were sized for a launch and never revisited, storage that grows without a lifecycle rule, accounts with no owner, data transfer nobody predicted and discounts bought before the workload was understood. That is why a single tactic, such as buying a commitment or deleting idle volumes, produces a short-lived improvement and then flatlines.
This comparison sets out five approaches available to a small or mid-sized business. They are not mutually exclusive; the useful question is which one to run first, and which ones need to become routine.
Why the bill resists a single fix
Cloud cost is a product of consumption, price and architecture. Consumption changes with traffic and with code. Price changes with commitments, regions and service tiers. Architecture determines how much work is needed per unit of demand. A programme that only addresses one of those will be overtaken by the other two.
Small teams also have a visibility problem. Without tags, account boundaries and an owner for each workload, cost conversations become guesses. The first intervention is usually organisational, not technical.
Five approaches
1. Visibility and accountability
Tag resources by workload, environment and owner. Split accounts where a shared account hides cost. Build a monthly view by service and by product, then review it with the people who can change it. This approach is slow to show savings but it makes every other approach measurable. It is the correct first move when nobody can explain the top ten cost lines.
2. Rightsizing and waste removal
Match instance families and sizes to observed utilisation, remove unattached storage and idle load balancers, set log retention and lifecycle rules, and schedule non-production environments. The work is mechanical but requires care: CPU averages can hide memory pressure and burst behaviour. Change one workload at a time, keep a rollback path and measure performance after the change.
3. Commitment-based discounts
Savings Plans and Reserved Instances trade flexibility for a lower unit price. They work when usage is stable and well understood. They are risky when a migration, a product change or a seasonal pattern could strand the commitment. Start with a small, short commitment on the most predictable baseline, not on the whole estate. Model the break-even utilisation before signing.
4. Architectural change
Move suitable workloads to managed services, adopt storage tiers and compression, use Graviton-based instances where compatible, cache at the edge and remove chatty service calls. Architecture can deliver the largest durable change, but it carries delivery risk and needs engineering capacity. It should follow visibility, not replace it.
5. A FinOps operating model
Give cost a cadence: a monthly review, a forecast, budgets with alerts, unit economics and a documented decision log. The operating model is what stops savings from eroding. It also changes behaviour: teams start asking for the cost of a design choice before they build it. For a small team, this can be a 90-minute monthly meeting with a fixed agenda and one owner.
Five approaches side by side
| Criterion | Visibility | Rightsizing | Commitments | Architecture | FinOps model |
|---|---|---|---|---|---|
| First result | Slow | Days to weeks | Immediate, if baseline fits | Months | One quarter |
| Effort | Low to medium | Medium | Low | High | Low, ongoing |
| Main risk | No action follows | Performance regression | Stranded spend | Delivery disruption | Meeting without decisions |
| Durability | Enables others | Medium | Fixed term | High | Highest |
| Skills needed | Finance and tagging | Operations | Procurement | Engineering | Finance plus engineering |
How the approaches interact
The sensible sequence is visibility, then waste, then commitments, then architecture, with the operating model running throughout. Commitments bought before rightsizing lock in the wrong baseline. Architecture changes made before tagging cannot be attributed to a product. A FinOps review without the authority to change anything becomes a reporting ritual.
There is one exception to the sequence. If a workload is clearly idle or duplicated, remove it immediately. There is no reason to wait for a full tagging programme before switching off a forgotten test environment.
Unit economics for a small business
Total spend is a poor steering metric because it rises when the business grows. Use a denominator that reflects demand: cost per order, per active user, per thousand API requests, per gigabyte processed or per tenant. Track the denominator and the cost together. If cost per order is stable while volume grows, the platform is scaling rationally. If it rises, find the driver before cutting capacity.
Use amortised cost for commitments so a monthly bill does not swing when an upfront payment lands. Separate production from non-production, and separate fixed platform cost from variable workload cost. That split makes a budget conversation possible without an accounting degree.
What to avoid
- Cutting observability: removing logs and monitoring saves money until the first incident takes longer to diagnose.
- Skipping backups: retention is a cost line, but recovery is a business requirement.
- Buying the maximum discount: a commitment that assumes today's architecture will persist is a bet, not a saving.
- Optimising one account: shared accounts, shared data transfer and central services often hide the real driver.
- Annual clean-ups: costs drift back within weeks without a monthly owner and review.
Choose the first move
- Choose visibility when nobody can explain the top cost lines or name an owner.
- Choose rightsizing when utilisation data is available and the estate has obvious idle or oversized resources.
- Choose commitments when baseline usage has been stable for several months and the architecture is not about to change.
- Choose architecture when the unit cost is structurally uncompetitive and the team can absorb a delivery project.
- Choose a FinOps model when savings keep eroding or when cost is never part of design decisions.
A 90-day cost programme
- Days 1-15: tag the estate, split accounts where necessary and produce a cost view by workload and environment.
- Days 16-30: remove idle resources, set retention and schedules, and fix the ten largest obvious waste items.
- Days 31-50: rightsize two workloads with performance checks before and after; document the rollback.
- Days 51-70: model a small commitment against the stabilised baseline; sign only what the forecast supports.
- Days 71-90: establish the monthly review, unit economics and a pipeline of architectural candidates.
Where an independent review helps
If you need an independent review of the estate before committing to discounts or a migration, start with Managed Technology & Cybersecurity. The pricing page explains how assessment and ongoing support are scoped. For the routine that keeps savings in place, read FinOps for small teams, year by year.
Sources
- AWS Well-Architected Framework, Cost Optimization Pillar, accessed 17 November 2026.
- AWS Savings Plans User Guide, accessed 17 November 2026.
- FinOps Foundation, FinOps Framework, accessed 17 November 2026.
- Flexera, 2026 State of the Cloud Report, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.