Skip to content

Introducing 04-10-26: the dedicated sort code for legal transactions.   Find out more

Reducing false positives in AML screening with AI


Kord Team

Kord Team

AML and financial crime •

Most alerts a sanctions or PEP screening system generates aren’t real. Industry data consistently puts false positive rates in AML screening somewhere between 85% and 95%, meaning the large majority of compliance effort goes into ruling out matches that were never genuine to begin with.

Here’s why traditional screening generates so many false positives, and how better matching is starting to change the numbers.

Why false positives are so common

Sanctions and PEP screening relies heavily on name matching. To avoid missing a genuine match, systems use fuzzy matching logic that tolerates spelling variations, transliteration differences, and partial matches, which is necessary given how many ways a name can legitimately be rendered, but it comes at a cost.

A common surname, a name that shares characters with a sanctioned entity, or a business name containing generic terms can all trigger an alert with no real connection to the person or entity actually being screened.

Compliance teams, understandably, typically set matching thresholds conservatively, since the cost of a missed genuine match is far higher than the cost of an unnecessary alert. That trade-off is rational for any individual decision, and it’s exactly what produces alert volumes far beyond what most teams can meaningfully investigate.

What traditional matching gets wrong

Classic fuzzy matching compares names as strings of characters, largely in isolation from other context. It doesn’t naturally account for entity type, jurisdiction, or the other identifying details, date of birth, nationality, registration number, that would help distinguish a genuine match from a coincidental one. Where that additional context isn’t built into the matching logic, name similarity alone is left to carry the entire decision, and name similarity alone is a weak signal at the volumes modern screening operates at.

How AI-based matching changes the picture

More recent approaches move beyond pure string comparison, weighing names alongside other available context to judge whether a potential match is plausible, not just textually similar.

A 2025 US Federal Reserve study comparing large language models against standard fuzzy matching algorithms found the language model approach reduced false positives by 92%, while simultaneously increasing genuine detection rates, across a range of realistic matching thresholds.

That’s a meaningful reduction, and demonstrates how better matching doesn’t have to mean trading detection for efficiency. Done properly, it improves both at once.

What this means

One benefit is fewer alerts. That has a demonstrable effect on a compliance team’s workload, though it’s only part of the story.

The real benefit is a screening process that produces more defensible outcomes, fewer legitimate clients wrongly flagged, and a higher proportion of the alerts that do get raised representing genuine threats or inconsistencies. Less noise, more signal.

This, in turn, improves a firm’s risk posture, rather than simply reducing the number of alerts a team has to close.

Building this into onboarding

Reducing false positives matters most when it’s built into the screening process itself, rather than added as a second pass after alerts have already piled up.

Kord Data applies contextual matching as part of standard AML screening, using more than a name string to judge whether a potential match is genuine, so that fewer legitimate clients are caught in the noise a cruder system would generate.