Expected Value Method Proposed to Improve AI Decision Outcomes

A method that replaces raw probability thresholds with expected monetary value calculations is described for predictive AI deployments. Under the current common practice, a model outputs a probability such as a 30 percent chance a transaction is fraudulent, and a fixed cutoff such as 50 percent determines whether the transaction is blocked.

The article states that this approach treats a $100 transaction identically to a $5,000 transaction despite differing financial consequences.

The proposed process multiplies the model's fraud probability by the transaction amount to produce an expected loss figure. For a $100 transaction with a 20 percent fraud probability, the expected value is $20; for a $5,000 transaction with a 5 percent probability, the expected value is $250.

A decision threshold is then applied to the expected value rather than the raw probability. The article states that setting a $15 expected-value threshold would block both example transactions.

The same multiplication step is presented for nonprofit fundraising, where a donor's predicted response rate is multiplied by estimated giving capacity to prioritize outreach spending. The article also lists churn modeling, collections targeting, and credit scoring as additional areas where the method has been applied.

The article notes that the technique requires only basic arithmetic after the model produces its probability output and does not require changes to the underlying model itself.