Shift-left FinOps: putting cost into architecture decisions before anything is committed
In brief
Where cost belongs in architecture review: design docs, ADRs, ranged estimates and IaC guardrails, what accuracy to expect, and a short checklist to use.
Steve Elliott · · 7 min read
On this page
- What does "shift left" mean in FinOps Foundation terms?
- Where does cost belong in architecture review?
- What should a pre-deployment estimate contain?
- What accuracy should you expect from a pre-deployment estimate?
- Which cost guardrails belong in infrastructure as code?
- A lightweight cost checklist for architecture review
- How do you show that shift-left worked?
- Sources
In our view, putting cost into architecture decisions takes four artifacts engineers already produce: a cost section in the design document, an estimate stated as a range, a cost line in the architecture decision record, and guardrails in infrastructure as code. That is what shift-left FinOps looks like before resources are provisioned or a commitment is bought.
Practitioners want this. In the FinOps Foundation's State of FinOps 2026, pre-deployment architecture costing ranked second among the tooling features respondents asked for. Below: what the FinOps Framework and AWS and Azure guidance ask for at design time, an illustrative estimate, and a review checklist.
What does "shift left" mean in FinOps Foundation terms?
The FinOps Foundation describes shift-left FinOps as moving cost decisions earlier in the engineering and product lifecycle, so teams can weigh cost alongside performance before anything is provisioned. Its FinOps Framework 2026 announcement uses similar words: shift left is "embedding FinOps earlier in the delivery lifecycle in design, architecture, and procurement."
Two Framework capabilities carry most of the detail, as of November 2026:
- Architecting & Workload Placement (formerly Architecting for Cloud). It says value and cost efficiency are best achieved by architecting them into a system's design. At Walk, a documented model weighing cost, value, sovereignty, resilience and sustainability guides placement; at Run, what was built is tracked against what was modeled. Its shift-left note puts it plainly: "the trade-off model used at design time follows the workload into build."
- Planning & Estimating. It defines estimating as exploring potential cost and value, separate from forecasting, which estimates later feed. Its shift-left note is about continuity: "the same estimate travels forward with the decision" instead of being reinvented at each gate. The note places a shared estimating method and a lightweight register at Walk; the maturity levels put them at Run, with an explicit range and confidence level. At Run, each estimate also carries a documented counterfactual that is checked against what actually happened.
The State of FinOps 2026 report (1,192 respondents, more than $83 billion in annual cloud spend) says FinOps teams are already building pricing calculators and giving pre-deployment guidance alongside enterprise architecture and engineering teams.
Where does cost belong in architecture review?
In our view, cost belongs in the artifacts engineers already produce, not in a separate FinOps form.
| Stage | Artifact | What the cost content should be | Who leads |
|---|---|---|---|
| Concept | Design document | Cost drivers, a ranged estimate, the business metric the cost scales with | Engineering, with FinOps support |
| Decision | Architecture decision record | The cost trade-off between options, and confidence in the estimate | Architect |
| Build | Infrastructure as code | Tags, budgets, allowed resource types, a cost check in the pipeline | Engineering |
| Run | Estimate register | Estimate against actuals after the first full quarter | FinOps |
The table is our proposal. Its first row follows Planning & Estimating, which says estimating is primarily supported by Engineering personas, supported by FinOps teams.
In our view, the decision record is the weak link. Microsoft's Well-Architected guidance asks each architecture decision record to state the context, the options considered, the important trade-offs and the confidence level of the decision. The log is append-only, and the page never mentions cost. We think the cost of each option, and the confidence in that number, should be a required field in the template. An append-only log also preserves the design-time estimate for later comparison with actuals.
What should a pre-deployment estimate contain?
Microsoft's Well-Architected checklist item CO:02 asks for a cost model covering initial cost, run rates and ongoing costs, and a budget that covers it with a buffer for unplanned spending. Its cost-model guidance calls observability, security and governance often-forgotten costs, adds testing and backups to the operating estimate, and suggests deriving hard-to-know inputs like data transfer from a business target.
AWS makes the same point in its Well-Architected question COST05, on evaluating cost when selecting services. Total cost of ownership includes operational overhead, licensing, data transfer and the effort to maintain a component, and "a comparison based on listed price alone misleads."
Check the price basis. AWS Pricing Calculator distinguishes before-discount rates, which are public on-demand prices, from after-discount rates, which reflect a customer's own discounts. Only bill estimates, available to management and standalone accounts, can model commitments such as Savings Plans; workload estimates can't. Planning & Estimating at Run maturity uses discount-adjusted, amortized pricing. Comparing options priced on different bases distorts the trade-off.
What accuracy should you expect from a pre-deployment estimate?
None of the sources we read sets a standard accuracy band for cloud estimates. The FinOps Framework's estimating KPI is estimate-versus-actual variance within an established threshold, and its Forecasting capability says each organization defines its acceptable variance. Nor do the calculators. AWS's Pricing Calculator FAQ says actual cost depends on actual use, excludes taxes, and will differ if you use services the estimate left out. HCP Terraform's cost estimation calls its output just an estimate and says some resources have "unpredictable usage-based pricing."
There is also an arithmetic reason to expect a calculator total to run low. The GAO Cost Estimating and Assessment Guide shows why adding up the most likely cost of each element gives an optimistic total. In its worked example, once risk was modeled, an unmanned aerial vehicle program had only about a 15% chance of costing its point estimate or less. We think a pricing calculator total, built from one input per line, is this kind of sum.
Take an illustrative new service. The numbers are made up.
| Line (monthly) | Low | Likely | High |
|---|---|---|---|
| Compute | $16,000 | $18,000 | $22,000 |
| Managed database | $9,000 | $9,000 | $12,000 |
| Storage | $3,000 | $4,000 | $7,000 |
| Data transfer | $1,500 | $3,000 | $9,000 |
| Logging and monitoring | $2,000 | $2,500 | $6,000 |
| Shared services and support | $3,500 | $3,500 | $4,500 |
| Total | $35,000 | $40,000 | $60,500 |
The calculator view says $40,000 a month. The three lines hardest to forecast (storage, data transfer and logging) make up about a quarter of the likely total but about 61% of the $20,500 gap between likely and high. We'd spend the review there, tying each line to a volume driver the product owner can forecast. In this example, data transfer, where egress fees sit, has the widest single range.
In our view, a design-stage estimate should be a likely figure, a high case and the lines that drive the spread, not a single number. Agree a variance threshold before go-live, record it in the register, and compare actuals after the first full quarter.
Which cost guardrails belong in infrastructure as code?
Azure's checklist item CO:04 lists release gates, governance policies, resource limits and access controls, and says to favor automation over manual processes.
AWS Well-Architected best practice COST02-BP05 suggests a sequence we think is right: start with notifications when cost or usage falls outside policy, then enforce once you understand the workload's limits. Enforcement options include IAM policies on which resource types can be created and where, service control policies, service quotas, and AWS Budgets actions, which can apply policies or stop EC2 and RDS instances, automatically or after approval.
Cost checks can also run in the pipeline. HCP Terraform, when cost estimation is turned on, adds an estimate phase between plan and apply that shows each resource's monthly cost and the change from the current state, for AWS, Azure and Google Cloud. Its documentation shows a Sentinel policy failing any run that adds more than $100 a month; as of November 2026, this works only in the legacy policy-checks mode.
We'd put four guardrails in code from day one: required tags matching the allocation model, a cloud budget per workload sized to the high case, allowed resource types and regions, and a pipeline check that flags increases beyond an agreed amount.
A lightweight cost checklist for architecture review
In our view, a design passes cost review when it answers these eight questions:
- Driver. Which business volume does cost scale with, and who forecasts it?
- Range. What are the likely and high monthly cost, and which three lines drive the spread?
- Price basis. Is the estimate at list or after-discount rates, and does it assume existing commitments?
- Full cost. Does it include data transfer, observability, security, backups, support and shared services?
- Options. What did the rejected options cost, and is that comparison in the decision record?
- Placement. Was a different technology category (public cloud, private cloud, data center) considered, as Architecting & Workload Placement now asks?
- Guardrails. Are tags, a budget, allowed resource types and a pipeline cost check in the code?
- Check-back. Who compares estimate to actuals, when, and against what variance threshold?
How do you show that shift-left worked?
The State of FinOps 2026 report counts measuring shift-left work among the discipline's most important open problems: avoided cost never appears on a bill. It also quotes a practitioner asking "how do we give developers credit for shift-left activities?" and notes that incentives have not caught up. Planning & Estimating's answer at Run maturity is to record the counterfactual, the design that would have been built, and compare projected avoidance with what happened.
We think the estimate register is the practical way to do that: one row per significant decision, with the options, the chosen estimate, the counterfactual and the actuals. Over a year, we think it would show which cost drivers are consistently underestimated. The report also lists team-level unit cost tracking as a way to recognize the work; in our view it also gives a slower second signal. A design-time estimate is only useful if someone can later trace it to the decision it informed and the money it committed, which is the case for keeping architecture decisions inside the same operating model as funding and portfolio choices.
Sources
- FinOps Foundation, State of FinOps 2026, released February 19, 2026
- Linux Foundation, State of FinOps Survey: AI, Value and Skills Top Priorities, February 19, 2026
- FinOps Foundation, FinOps Framework 2026: Executive Strategy, Technology Categories, and Converging Disciplines, March 19, 2026
- FinOps Foundation, Shift Left FinOps topic page
- FinOps Foundation, FinOps Framework capability: Architecting & Workload Placement
- FinOps Foundation, FinOps Framework capability: Planning & Estimating
- FinOps Foundation, FinOps Framework capability: Forecasting
- AWS, Well-Architected Framework, COST05: How do you evaluate cost when you select services?
- AWS, Well-Architected Cost Optimization Pillar, COST02-BP05 Implement cost controls
- AWS, Understanding AWS Pricing Calculator concepts
- AWS, AWS Pricing Calculator FAQs
- Microsoft, Azure Well-Architected: Design review checklist for Cost Optimization, updated November 15, 2023
- Microsoft, Azure Well-Architected: Architecture strategies for creating a cost model, updated November 15, 2023
- Microsoft, Azure Well-Architected: Architecture decision record, updated April 13, 2026
- HashiCorp, HCP Terraform: Cost estimation overview
- U.S. Government Accountability Office, Cost Estimating and Assessment Guide (GAO-20-195G), March 12, 2020; Cost risk analysis data and methods
Frequently asked questions
What is the difference between shift-left FinOps and cost optimization?
Optimization changes resources that already run: rightsizing, commitments, deleting waste. Shift-left FinOps puts cost into the design before anything is provisioned, so fewer costly decisions need undoing later. The FinOps Foundation describes shift left as moving cost decisions earlier in the engineering and product lifecycle, and links it to its Architecting & Workload Placement capability.
How accurate are cloud pricing calculator estimates?
Neither AWS nor HashiCorp publishes an accuracy figure for its estimates. AWS says actual cost depends on actual use, excludes taxes and differs if services are left out of the estimate. HCP Terraform notes that some resources have usage-based pricing it can't predict. In our view, treat a calculator total as a likely case and add a high case driven by data transfer, storage and logging.
Who should own the cost estimate for a new architecture?
The FinOps Framework says estimating is primarily supported by engineering, with help from the FinOps team; product, finance or leadership join for important or complex estimates. In our view the architect should own the figure in the decision record, while FinOps owns the method, the register and the later comparison of estimates against actual spend.
Should cost checks block deployments?
Not at first. AWS Well-Architected guidance describes notifications as a common first step, with enforcement once you know the workload's limits. We'd begin with pipeline checks that flag cost increases above an agreed amount, then add blocking rules for clear cases such as disallowed regions or resource types.
Steve Elliott
Published November 22, 2026
More on FinOps & Cloud
FOCUS, commitments, shared cost.
Keep reading
What the FinOps Framework 2026 changes for practitioners
A close reading of the FinOps Framework 2026: the new Executive Strategy Alignment capability, rewritten Scopes guidance, technology categories and renames.
Should FinOps report to the CTO, CIO or CFO?
Most FinOps teams report to the CTO or CIO; 8% report to the CFO. When each line fits, what it costs, and how to make a dotted line to finance work.
What AI agents should and shouldn't do in FinOps
Agents suit FinOps work that is read-only or easy to undo. Commitments, policy and chargeback disputes need a person. Where the line sits, with guardrails.
What State of FinOps 2026 shows, and what it can't
State of FinOps 2026 shows FinOps spreading to AI, SaaS and data centers. What its 1,192 responses support, where the method is thin, and what to tell a CIO.