13 credits to pick the model for a 13,000 credit job

The job was to find which companies actually sell B2B. The budget was 13,000 credits. The obvious plan classified every row and spent 10,432 of them, which left almost nothing to find people.

There was a free signal available. Nobody had tested it.

We normalized every domain with a credit free action first. Zero duplicates came out of it.

Then we picked rows we knew were hard. We rewrote the classifier prompt three times against real failures, and ran a model bake off on those anchor rows before touching the full table.

• The whole validation cost 13 credits.
• The cheaper model failed the same anchor three times in a row, once with a confident false positive on a personal use page. The expensive model found the real trade portal on the first try. That's a model ceiling, not a prompt problem.
• The free signal filled 809 rows at zero credits. It also missed a known B2B company and flagged consumer store locators. We kept it as a backstop, not a source.
• The plan: the expensive model on the larger segment, 3,216 rows at 6,432 credits. The other 2,000 rows go through app detection at zero classification spend. That leaves 6,568 credits for people finding, 4,000 more than the obvious plan.

Here's the awkward part. The full run never happened. The term ended before it was approved. What's left is the method and the price, and both held on every row we tested.

Before you buy an AI column at scale, send us five rows you know are hard. That's where we'd start.