False positives & efficiency

How many ordinary customers does your screening flag, and why?

Every alert on an innocent customer costs analyst time, and it doesn't fall evenly. We measure alert rates by customer corridor and threshold, show which names carry them, and trace each one to the engine, the configuration or the data.

Customers alerting, name only, at threshold 0.70
US1.1%24 alerts per 1,000 customers
UK4.0%292 alerts per 1,000 customers
Nigeria10.1%603 alerts per 1,000 customers
25×the review work per customer in the Nigerian corridor compared with the US, on the same engine and the same lists
Measured · sample, Oct 2026

Sample report · 2,250 synthetic customers · October 2026

Read alerts per 1,000 customers, not just the rate.

Three corridors screened against the UK and OFAC lists on name alone. The rate says how many customers are touched; alerts per 1,000 says how much an analyst has to clear. A flagged Nigerian name brings six alerts on average, because names such as Mohammed and Ibrahim appear in many listed records.

Customers alerting, at every threshold · UK and OFAC lists

0%5%10%15%20%0.600.650.700.750.800.850.900.70 used in the sampleNigeriaUKUSThreshold →

Lowering the threshold from 0.70 to 0.60 nearly doubles alerting Nigerian customers (101 to 196 of 1,000); raising it to 0.80 cuts them to 62. The corridors keep their order at every setting, and a threshold moves them all at once. Thresholds also don't transfer between engines: in your test we run your own setting and one either side.

Who it falls on, and why

The alerts land on a few name origins, and each cause has an owner.

In the Nigerian corridor, 100 of the 101 alerting names were Muslim or Hausa-Fulani names. Most alerts came from the engine treating a customer's names as matching a longer listed name, not from a true namesake.

Share of names alerting, by name origin

NigeriaMuslim / Hausa-Fulani names16.9%100 of 593
NigeriaOther origins0.2%1 of 407
UKEast Asian40.0%2 of 5
UKBangladeshi15.8%3 of 19
UKPakistani9.7%3 of 31
UKOther origins1.0%2 of 195
USHispanic6.3%8 of 126
USAnglo & European0.2%2 of 847

Name origin is assigned from general naming knowledge and is approximate. Small groups move on single names: in the UK set, Mohammad Ahmed alone raised 38 of 73 alerts.

Why they alerted, and who owns the fix · alerts at threshold 0.70

CauseUKUSNigeriaOwner
Same name as a listed personName alone can't clear it; date of birth, nationality or address can.104Data
All the customer's names inside a longer listed nameThe listed person's other names count for nothing. Weighing unmatched words cuts these.111384Engine · matching logic
As above, with at least one name as a near spellingA tolerance such as Mohammed for Muhammad, applied to very common names.4510432Engine · spelling tolerance
Only part of the name matchedA very common word carries as much weight as a rare one. Weighting words by how common they are is the usual fix.16183Engine · scoring
All alerts7324603

Secondary data

What date of birth can clear, and what it can't.

Measured on the list records behind every alert. A full date of birth is the biggest lever, but some alerts sit against records with none, and no amount of customer data will clear those. Your own test measures how many your system actually clears.

Full date of birth82%

of the 603 Nigerian-corridor alerts are against a record with a full date of birth, the biggest lever for clearing them

Year only11%

clear only when the birth years differ

No date of birth6%

39 alerts that date of birth can never clear. 23 of 101 flagged customers have at least one

Run this on your engine

Synthetic customers built for your corridors, run through your own screening.

The figures on this page come from Escalate's in-house test engine, built on open-source matching, screening names alone. Your report runs the same method through your own screening, with date of birth and nationality, so you see how many alerts clear automatically and how many reach an analyst.

  1. 20 minutes to agree corridors, lists, the secondary data you hold, and what you return: alert or no alert, score and matched record.
  2. You receive test IDs and names, plus the secondary fields your system accepts. No expected results.
  3. You run the file and send back the results.
  4. You get the report on your own engine, across every corridor and threshold.