Home › Savings Methodology

How We Measure Savings

Our 20% guarantee is only as good as the way we measure it. This is exactly how it works, including its limits.

1. The baseline

Before go-live we agree a baseline for each workload: your cost per equivalent successful request today. We record the usage mix, quality, latency and error rate alongside it.

2. What counts as an equivalent request

The same task, at the same quality bar. A request that fails, retries or produces worse output doesn't count as a saving.

3. The measurement window

After go-live we measure the same thing on your real traffic, typically over 14 days.

4. Growth and price changes

Savings are measured per request, so more or less traffic doesn't hide or inflate them. If a provider changes its list prices during the window, the baseline is repriced too.

5. Reconciled with your invoices

Dashboard numbers are reconciled against your provider invoices, not just token counts.

6. Quality, latency and reliability guardrails

Every change is tested against your evals, or a representative evaluation set we build with you first. A change that makes quality, latency or error rate worse doesn't ship, and every change can be rolled back.

7. The result

If your all-in cost per equivalent request, meaning provider charges plus our fee, isn't down at least 20%, you pay us nothing. That covers comparable work that completes successfully, at the quality and latency thresholds agreed before go-live. Charges you've already paid to model providers aren't refunded. The full rules are in our guarantee terms.

A worked example

Illustrative numbers. Your teardown uses your own.

What we don't do

• Publish average savings figures we can't back with your own numbers.
• Count savings from cutting features or traffic.
• Switch you to a model you haven't approved.

The evidence you keep

Versioned changes, test outputs, the baseline, your approvals and a monthly CFO-ready savings report.