How a global bank cut false positives in fraud detection by 62%.
They had a high-accuracy model. They also had alert fatigue. The breakthrough wasn't a better algorithm — it was a better decisioning layer.
SAME MODEL · NEW DECISIONING LAYER
POLICY + CONTEXT + HUMAN JUDGMENT, COMBINED
The bank's fraud model was accurate by every offline measure. In production, it buried investigators in alerts they had stopped trusting. High-value fraud slipped through not because the model missed it, but because the humans downstream had learned to tune the system out.
We left the model largely untouched and rebuilt everything around it: a decisioning layer that combines model output with policy rules, account context, and explicit human-judgment checkpoints. Every alert now arrives with a reason, a confidence, and a recommended action — and the ambiguous ones route to people by design.
False positives fell 62%. More importantly, investigators started trusting the queue again — the exact adoption outcome we measure every engagement against. The same governance layer now extends to two adjacent decision systems at the bank.
Have a model people
stopped trusting?
About this engagement
If the model was already accurate, what actually got fixed?
The decisioning layer around the model, not the model itself. It was accurate by every offline measure but buried investigators in alerts they had stopped trusting. We left the model largely untouched and rebuilt how its output turns into an action.
What is the decisioning layer you built?
It combines model output with policy rules, account context, and explicit human-judgment checkpoints. Every alert now arrives with a reason, a confidence, and a recommended action, and the ambiguous ones route to people by design. Clear cases auto-clear; genuine risk escalates.
How was a 62% cut in false positives achieved without retraining?
By changing what happens after scoring rather than the score. Policy, context, and thresholds decide which alerts reach a human, so noise no longer drowns the queue. The same model produced far fewer false positives because the routing did the filtering.
Why was investigator trust the real outcome?
High-value fraud was slipping through because humans downstream had learned to tune the system out, not because the model missed it. Once alerts became explainable and fewer, investigators trusted the queue again. Adoption is the outcome we measure every engagement against.
Does this approach transfer beyond fraud?
Yes. Every route in the decisioning layer is logged, explainable, and auditable, so the same governance pattern already extends to two adjacent decision systems at the bank. If you have a model people stopped trusting, the layer, not the algorithm, is usually where the fix lives.