Fraud decisioning: when the alert is not the unit of work

The shape of the problem

A retail bank's fraud operation typically runs a rules engine written over a decade or more, one or two machine-learning scores bolted on at different times by different vendors, and several human review queues that each believe they are the queue of record.

Alert volume outruns clearing capacity. The backlog ages. When it gets bad enough, thresholds go up — which is another way of saying the bank stops looking at a category of risk until the queue feels manageable again.

Nobody in the building thinks this is good. The reason it persists is structural: no single team owns the decision. The rules belong to financial crime, the scores to the data science group, the queues to operations, and the audit obligation to a fourth team that can only inspect the outcome after the fact.

Why the obvious fix does not work

The obvious fix is a better model. It rarely is the problem.

Both scores in a setup like this are usually reasonably calibrated on the data they have seen. What is broken is that nothing in the estate can say why a case was flagged in terms a reviewer can act on. An analyst opening an alert sees a score, a rule id and a transaction. To decide, they open the core banking record, the card platform, the KYC file, and the shared inbox where the branch network reports anything unusual.

Most of the handling time is retrieval, not judgement. The operation is not short of intelligence. It is short of assembled evidence.

The unit of work was never the alert. It is the decision — and in most estates no system holds one.

The design

1. A governed record, before any model runs

The Cognitive Data Hub ingests the systems that matter — core banking, cards, KYC, the branch reporting inbox, device telemetry, and the sanctions and watchlist bureau. Entity resolution collapses the customer across all of them into one governed record with lineage back to every source row.

This is the unglamorous half and it takes the longest. It is also the half that makes everything above it possible. Expect a few per cent of records to fail the quality bar and go to quarantine rather than into the decision path. The instinct is to treat that as a defect. It is the opposite: a funnel with no rejects is one nobody has run.

2. Three evidence lanes, running concurrently

Every case opens three lanes at once:

• Ground — Cognitive RAG retrieves the customer's history, prior dispositions, related parties and any narrative the branch network filed.
• Verify — deterministic models check the case against real constraints: velocity limits, geography, device reputation, sanctions exposure. These are not learned, and they are not negotiable.
• Explain — a reasoning model weighs the assembled evidence and drafts the rationale a reviewer will read.

No lane decides alone. A case only leaves the board when the evidence quorum is satisfied, and confidence is computed from lane agreement rather than asserted by a single score.

3. Human approval where policy demands it, and nowhere else

The reflexive design puts a human in front of everything, which is how the queue got there. Map the actual regulatory obligation instead — it is usually narrower than the operation assumes, and the remainder can be auto-dispositioned with full trace, subject to a blind sample.

The cases that do go to a human arrive with the evidence already assembled.

What to measure afterwards

Agree these before go-live, and measure them against the manual baseline rather than against each other:

• Time from trigger to decision, at the median and at P95.
• False positive rate — the number that decides whether a legitimate customer gets declined at a supermarket till.
• Analyst handling time, split into retrieval and judgement. The split is the interesting part.
• Agreement between automated dispositions and a blind human re-work sample.
• Time to produce an audit pack for a regulator's sample.

Where it runs

Inside the bank's own tenancy, in-region, with the bank holding the keys. The models are trained on the bank's governed data, evaluated against its own baselines, and portable out if it ever wants to take them elsewhere.

We build it, run it and watch it. The bank owns it.