Cutting AML False Positives Without Missing Real Matches

Every compliance team knows the number, even if they don't say it out loud: the share of screening alerts that turn out to be real. In most queues it is a small share. The rest are namesakes, partial matches, and old entries for people who share a common name with your customer.

High false positives are not a nuisance. They are a control failure. When almost every alert is noise, analysts learn to clear fast, and the two that matter get cleared with the rest. Fixing it is not about lowering the alarm. It is about giving the engine and the analyst better information.

1. Send everything you know on the screening call

The single biggest lever is the request itself. A call with only a name forces the engine to decide on the name. Add date of birth, nationality, and entity type, and most namesakes rule themselves out before they reach anyone.

If your KYC provider gives you these fields, pass them through. If your onboarding flow screens before KYC completes, consider moving the screening call a few seconds later, to the point where the verified fields exist.

2. Screen the full name

Truncated names create noise. "Abdullah Saad" matches far more records than "Abdullah Saad Mohammed Ibrahim". Screen the full name chain your KYC provider returns, and store it so re-screening uses the same input.

3. Treat individuals and companies differently

A person named "Gulf Trading" does not exist, and a company named "Mohammed Ali" is rare. Passing the entity type stops individual records from matching against company entries and the other way round. It sounds obvious, and it is missing from a surprising number of integrations.

4. Tune on evidence, monthly

Look at last month's cleared alerts and ask why each one was cleared. You will find patterns: a specific list producing most of the noise, a common family name, alerts where the birth year was off by decades. Each pattern is a tuning decision you can make on evidence rather than gut feel.

Track three numbers every month:

• REVIEW rate: the share of screenings that need a person.
• True match rate: the share of REVIEW results that were real.
• Time to close: how long a REVIEW sits before someone decides.

Before you approve a tuning change, run it against a test set: confirmed matches from past cases, known namesakes, and the Arabic spellings and transliterations your customers actually use. The change is good when every confirmed match still reaches review and the noise drops.

5. Make the evidence easy to read

Half of the cost of a false positive is the time it takes to rule out. Put the customer's details and the list entry side by side: names, aliases, date of birth, nationality, and the source. An analyst who can see that the list entry was born in 1952 and the customer in 1994 clears it in ten seconds. One who has to open three tabs takes ten minutes.

What not to do

• Don't raise the threshold until the alerts go away. You will miss real matches on transliterated names.
• Don't exclude whole lists because they are noisy. Fix the matching or add identifiers instead.
• Don't auto-clear by rule without recording why. Someone will ask later.

A realistic target

Measure confirmed-match detection, review volume, and resolution time. A healthy queue is one your team clears the same day, with the evidence and outcome recorded for every alert.

Naiza scores candidates on the name and the supporting fields together and shows the customer and the list entry side by side in the case. The AML screening API covers the request fields, and the false positive entry has the definitions.