Skip to content
Understand
Tool

Cost per successful outcome calculator

Which setup delivers finished work at the lowest full cost? Runs in your browser. Nothing is stored or transmitted.

Compare up to four setups for the same workload on what each successful outcome costs, including the people who fix what fails. Showing the illustrative document-extraction example from The Economics of Model Selection.

Workload

Documents, cases or tasks.

Higher when errors are found late, by a customer or a regulator.

Sampling or review applied to every item, pass or fail. Zero if none.

Setups to compare
Small model, routed to mid-tier when unsure is cheapest per successful outcome, at £0.114 per item or £11,400 a month. Small model, no check is cheapest per attempt but costs £0.334 per outcome once failures are fixed.

Cost per successful outcome

  • Small model, no check£0.334 · £33,400 a month

    £0.004 attempt, £0.000 check, £0.330 fixing failures.

  • Mid-tier model£0.140 · £14,000 a month

    £0.020 attempt, £0.000 check, £0.120 fixing failures.

  • Small model, routed to mid-tier when unsure (lowest)£0.114 · £11,400 a month

    £0.009 attempt, £0.000 check, £0.105 fixing failures.

  • Frontier model£0.150 · £15,000 a month

    £0.090 attempt, £0.000 check, £0.060 fixing failures.

Each item costs one attempt and any routine check, plus a person's fix for the share that fails. Clearance rates should come from a test set of real work, including the hard cases, judged against a quality floor set with the business owner before testing. Try raising the fix cost: when errors are found late, the more capable setups often become the cheaper choice.

The method is set out in The Economics of Model Selection, which also explains how to set the quality floor before testing. Once a setup is chosen, the inference cost estimator turns its rate card into a monthly forecast.