A computer helper says how sure it is. We pick a number. If it's surer than that, it can go ahead. If not, it asks a grown-up.
A decision model gives each case a score. The threshold is the score above which it acts without asking. You set it by comparing the cost of a wrong yes with the cost of a wrong no, and by agreeing it with the people affected.
Set the action threshold where the expected cost of a false positive equals that of a false negative, adjusted for review capacity. Publish the threshold, the costs behind it and the review date.
What a threshold does
A decision model gives each case a score, often written as a confidence between 0 and 1. On its own the score does nothing. The threshold turns it into an action: above it the system acts, below it the case goes to a person.
Many systems use two thresholds. Above the top one, act. Below the bottom one, reject or stop. In between, ask a person.
Start from the cost of mistakes
There are two ways to be wrong:
- A wrong yes: the model acts when it shouldn't. An invoice gets paid that should have been queried.
- A wrong no: the model holds back when it could have acted. A clean invoice waits a day for a person.
If a wrong yes costs ten times as much as a wrong no, the model should only act when it's very sure. A common starting point is to act when the score is above cost of a wrong yes ÷ (cost of a wrong yes + cost of a wrong no). With costs of £50 and £5, that's 50 ÷ 55, so act above about 0.91. That only works if the scores are calibrated.
Then check capacity
A threshold that sends 60% of cases to people won't survive a busy week. Look at how many cases fall in the "ask a person" band and compare that with the time people actually have. If it's too many, either accept more risk or improve the model before going live.
Make it defensible
Write down the costs you used, the threshold you picked and who agreed it. Set a review date. When someone asks why the model approved a case, you can show the reasoning rather than just the score.