Credit-based enrichment tools are easy to overspend in. A column runs on the wrong range, a waterfall calls five providers instead of two, or a test on 50 rows quietly runs on 5,000. By the time you notice, the credits are gone.
The estimate formula
For a waterfall, each provider only sees the rows the previous ones missed. So the expected cost of a run is:
- rows × (chance a row reaches this provider) × (this provider's hit rate) × (price per hit), summed over the providers in the chain.
- The chance of reaching a provider is the product of the miss rates of everyone before it.
- The most a run can cost is one hit per row at the most expensive step, because a hit ends the row.
Example: 1,000 rows through three providers at 72%, 64% and 62% hit rates and 3 credits each. Expected matches are about 962, so the expected cost is about 2,885 credits, and the most it can cost is 3,000.
Controls worth having
- A pre-run estimate on every run, not just the first.
- A hard cap that stops a run, not an alert after the fact.
- A per-run ceiling for one-off experiments.
- Free misses and free repeats, so a bad list and a re-run cost nothing extra.
- A ledger that explains every charge, cell by cell.
In Kiln
Every run shows its estimate and worst case first. Each workspace has a spend cap on by default at its plan allowance, checked on the server before any provider is called, with an alert at 75% of it. Misses, errors, failed verifications and cached repeats are never billed, and every cell records what it cost and why.