Direct avoided cost
Used when an exact provider call is eliminated, such as a verified cache hit.
How we measure savings
The calculation
We separate what the workload would have cost from what it actually cost.
Baseline costWithout optimization
Actual costProvider spend + overhead
Net savingsCustomer benefit
01 · Evidence
Use the evidence that fits the lever.
There is no single convenient formula applied to every optimization.
Holdback A/B
A controlled share keeps the original path so optimized and baseline behavior can be compared.
Replay evidence
An approved request set estimates the counterfactual when a live holdback is not appropriate.
02 · Process
From request to report.
Tag
Attach safe workload labels such as team, feature, provider, and model.
Measure
Record actual usage, provider pricing, and the optimization applied.
Compare
Use the evidence method approved for that lever and workload.
Reconcile
Publish baseline, actual spend, overhead, and net savings separately.
03 · Integrity
Make the result reviewable.
Pricing coverage
Provider catalog or accepted contracted rates support the cost calculation.
Attribution coverage
Savings remain connected to the team or feature responsible for the workload.
Quality acceptance
Each lever is evaluated against the quality threshold appropriate to its risk.
Next step
Verify one lever before expanding.
Start with a narrow workload where the baseline, fallback path, and savings ledger are easy to audit.