Early accessNow onboarding design partners.Request access
AutoCompliance
← All articles

16 November 2026

How to reduce false positives in sanctions screening

False positives cost analyst time and customer conversion. Practical levers to cut alert volume without increasing the risk of missing a true match.

Every sanctions alert needs a human decision. When most alerts turn out to be false positives, the cost shows up in three places: analyst headcount, onboarding delays and customer drop-off. Here are the levers that reduce noise without lowering your guard.

1. Fix the input before you match

Many false positives start with messy input: titles in the name field, company suffixes, double spaces, all-caps strings or a name and surname pasted into one field twice. Normalising input (removing honorifics, punctuation and duplicates) removes a surprising share of noise before matching even starts.

2. Use secondary identifiers

Name alone is weak evidence. A date of birth is the most widely available secondary identifier in sanctions lists, even when it is only a year or a range. A screening engine should:

  • boost confidence when the date of birth matches,
  • give partial credit for a matching year or a date inside a listed range,
  • tolerate day and month swaps and off-by-one years,
  • reduce confidence when the dates clearly conflict,
  • never penalise a missing date, because many list entries do not have one.

3. Treat single-word and common names carefully

A list entry that is just one given name, or a customer known by a mononym, can match thousands of people. Rather than ignoring these cases, apply a score penalty that a matching date of birth can lift. Very common names deserve a stronger penalty than rare ones.

4. Verify candidates with a smarter model

String similarity cannot tell Maria from Marina. A second-stage model trained or prompted to compare names can reject look-alikes while keeping genuine variants such as transliterations and diminutives. This is where most of the false-positive reduction in modern systems comes from.

5. Remember your decisions

Once an analyst confirms that customer X is not list entry Y, you should not be asked again tomorrow. A whitelist keyed on your customer ID and the list entry stops repeat alerts, and it should expire automatically when the list entry changes.

6. Set thresholds with data, not intuition

Pick a threshold on a labelled benchmark that contains both perturbed list entries (which must be caught) and control names (which must not). Re-run the benchmark whenever the model, the threshold or the lists change.

7. Measure the queue

Track alert volume, hit rate, time to decision and the share of alerts closed as false positives. If one rule or list produces most of your noise, you will see it.

AutoCompliance applies input normalisation, fuzzy date-of-birth scoring, single-word penalties and AI verification by default, and returns a score breakdown with every match so your team can see exactly why an alert was raised.

Screen smarter from day one

We are onboarding a small group of design partners. Tell us about your screening volume and lists, and we will set up sandbox access.