Vague prompts cost twice

A vague extraction prompt produces invalid JSON, triggers a retry, and the final successful call is what you pay for. Tight prompts are cheaper, not just better.

3 min read

bizSupply meters successful AI calls only. That sounds reassuring until you realise that an LLM giving up after a parse error and succeeding on the second attempt is two AI calls of work but one chargeable call. The cost shows up as a higher credits-per-document figure, and the cause is almost always a prompt the LLM had to interpret rather than execute.

From the billing docs

Failed AI calls do not consume credits. Internal retries are not double-billed — only the final successful call counts. But if your prompt regularly forces a retry, you are doubling the work behind every chargeable credit.

What "vague" looks like

"Extract the relevant information from this contract" — relevant by whose standard?
"Return the dates" — formatted how, in what timezone, named what?
"List the parties" — as strings, objects, with addresses?
No example output, no schema reference, no fallback when a field is genuinely missing.

What tight looks like

State the shape, the format, and what to do when something is missing. Reference the ontology by name. Show one short example. Set temperature: 0.1. The result is fewer retries, fewer parse errors, and a credit-per-document figure that stops drifting upward over time.

Test on five real samples

Before deploying any prompt change, run it against five documents from the actual source. If even one produces invalid JSON or a "best guess" string for a missing field, refine before you scale.

Still need help?

If this article does not solve it, the bizSupply team is one ticket away.

Submit a ticket