Can an AI agent reduce false positives in AML screening?

Yes, if you measure the right thing. An AI agent can cut the time false positives consume by resolving clear mismatches with evidence, fixing the data gaps that cause them, and proposing tuning backed by tests. Changing thresholds, lists or matching rules stays with a compliance officer, because fewer alerts can also mean more missed matches.

What causes false positives in screening?

A false positive in AML is an alert that, on review, turns out not to involve the risk it flagged. In sanctions and PEP screening, most false positives come from four sources:

• Common names: a customer shares a name, or part of one, with a listed person.
• Missing data: no date of birth or nationality on the customer record, so the tool cannot rule a match out.
• Loose settings: fuzzy matching thresholds set wide to avoid missing transliterations.
• Poor list data: list entries with few identifiers.

Each source needs a different fix, and only some of them are safe to fix by reducing alerts.

Which approaches actually reduce false positives?

There are two ways to reduce the burden, and they carry very different risk.

Faster resolution, same alerts. The tool raises the same number of alerts, but the agent resolves the clear ones with evidence: different date of birth, different nationality, a company registered in another country. A person reviews by sample. This keeps detection as it was. It reduces cost, not alerts.

Fewer alerts. The agent analyses which settings, list fields or customer data gaps produce noise and proposes changes. Filling missing dates of birth is low-risk and often the largest single improvement. Tightening a fuzzy match threshold is not: it can stop the tool matching a real target spelled differently.

Firms often start with the first and move carefully into the second. The mechanics for each screening type are in can an AI agent do sanctions screening? and can an AI agent do PEP screening?

What may the agent prepare, and what is the officer's call?

The test against true matches is the heart of it. A tuning change that would have suppressed any historical true match, or any planted test name, is not a false positive reduction. It is a gap.

What does a tuning proposal look like?

Take a remittance firm screening senders and recipients. Its analysts clear thousands of alerts a month, almost all false.

The agent analyses three months of alerts. It finds that a large share come from recipients with no date of birth recorded, and that one list contains many entries with only a single name. It proposes two changes: collect date of birth for recipients above a set transfer value, and treat single-name entries differently, requiring an additional identifier before alerting.

It tests both against every true match the firm has recorded and a set of planted test names, including transliterations. The first change misses nothing. The second would have missed one historical true match, a listed person known only by a single name. The officer approves the first change and rejects the second. The agent records both decisions with the test results.

Which controls stop false positive reduction becoming a blind spot?

• Every tuning change is tested against historical true matches and planted test names before approval.
• Below-the-line samples continue. A person reviews alerts that just missed thresholds, to see what tuning excludes.
• Discount reasons are recorded for every agent-resolved alert.
• The agent cannot change settings. Proposals go to the officer; configuration changes are made by a person.
• The regulatory angle is checked. In the EU, consider how the AI Act applies to agents in AML work; see EU AI Act and AI agents in compliance and AML teams.

The general principle is in AI compliance agents: second line, never the last word.