1 · Assumed fraud exposure by coverage band
Share of fraud exposure in each successive 10% coverage band.
2 · Incremental fraud prevention
Additional fraud dollars prevented relative to current coverage.
3 · Alerts and false positives at the configured fixed rate
Added alert volume, and the estimated false-positive share, at the same model and alert rate.
4 · Incremental benefit, cost and net value
Additional prevention less additional alert-handling cost.
5 · Savings, cost and net waterfall
Fraud savings, false-positive impact, alert / triage cost (true + false positive) and resulting net value at the proposed coverage.
What changes between the assumptions?
Same gross fraud exposure, model effectiveness, alert rate and cost. Only the distribution changes.
Assumptions, evidence and limits
What the curves mean
The horizontal axis follows the order in which a selection policy adds transactions to scoring. It is not chronological order. A policy may select remote payments or transactions above a risk threshold first, wherever they occur in the day.
For coverage fraction c, assumed exposure in scope is ck. k < 1 concentrates exposure in initially selected transactions; k = 1 makes exposure proportional to transaction coverage; k > 1 concentrates exposure in transactions added later. These smooth curves are illustrative assumptions. They are not fitted to bank data or to the evidence below.
How value is calculated
Additional prevention = exposure after other controls × incremental prevention effectiveness × (targetk − currentk). Effectiveness includes successful intervention. The exposure input is measured before this modeled AI intervention, including at current coverage.
Additional alerts = annual transaction count × (target − current) × fixed additional alert rate. A fixed share of those alerts is treated as false positives; the remainder are treated as true positives. Alert / triage cost = additional alerts × average triage cost, applied to every alert regardless of outcome. False-positive impact = false positives × an additional false-positive impact cost, on top of triage. Net = additional prevention − alert / triage cost − false-positive impact.
Scoring, alerting and intervention are separate decisions. In practice, thresholds, population mix, automation and case consolidation can change the alert rate, the false-positive share, the triage cost or the false-positive impact cost. This demonstration holds them constant to isolate selection effects. Fraud prevention is measured in dollars; alerts are counts. It does not infer the false-positive share from prevented dollars; that share is a separate fixed assumption.
What the evidence supports
The joint ECB–EBA 2025 Report on Payment Fraud, pages 18–19 reports that remote payments represented around 18% of card transaction counts and 28% of value, but around 83% of card fraud by count and value in 2024. This establishes unequal fraud exposure across those channels. It does not establish the curve for any bank's scoring selection policy, or the incremental effectiveness of AI.
Celent's 2022 Operationalizing Fraud Prevention on IBM z16, pages 6 and 8 links limited advanced-model coverage to missed preventable fraud while also describing high-risk selection and other fraud controls. Coverage percentage alone cannot establish the percentage of fraud missed. This explorer tests that inference; it does not disprove Celent's savings estimates.
What must be validated
Measure fraud counts and dollars inside and outside current AI coverage, outcomes of existing controls, additional successful prevention, alert rates, the true false-positive rate, actual triage cost per alert and the true cost of false-positive impact (customer friction, support contact, lost legitimate-transaction value). A representative selection can produce proportional expected exposure; a risk-prioritised selection can produce diminishing incremental exposure; a policy missing risky populations can produce increasing exposure.
The best modeled expansion is searched in 0.1 percentage-point steps. Technology, implementation and recurring platform costs are excluded. Case consolidation, segment overlap, intervention timing, capacity constraints and changing effectiveness are excluded. These results are conditional estimates, not a complete investment case.