98% to 1%: How A Global Bank Cut AML False Positives with Digital Workers

A global financial services firm was processing ~105,000 AML (anti-money laundering) screening alerts a month. 95–98% were noise. Here's what happened when a Digital Worker took over the investigation.

TL;DR

Over 50 FTEs. ~105,000 alerts per month. A 95–98% false positive rate. The vast majority of analyst time was spent confirming non-matches — not investigating real risk.

causaLens deployed an AML Name Screening Digital Worker: a coordinated multi-agent system that ingests, investigates, and classifies alerts end-to-end.

In the MVP Phase: 0% false negative rate. 1% false positive rate. Case resolution dropped to roughly five minutes per alert. This was achieved by handling false positives autonomously.

The problem wasn't the alerts. It was everything that happened after.

Each alert required an analyst to manually cross-reference watchlists, search Google, Companies House, LinkedIn, and Bloomberg, apply their own judgment, and write it up. No systematic evidence capture. No audit trail that persisted across cases. No learning from prior decisions. Every alert started from scratch.

At ~105,000 alerts per month, that's over 1,500 analyst hours every month — the vast majority spent confirming false alarms. The operational overhead was substantial. And it did not scale.

What a Digital Worker actually does differently?

The causaLens Digital Worker orchestrates the full AML investigation lifecycle through a coordinated multi-agent system.

A Triage Agent monitors the incoming alert queue continuously, classifying by risk tier and structuring for downstream processing. What used to take 1–2 hours of analyst time happens instantly.
An Entity Resolution Agent handles the genuinely hard part: aliases, transliterations, abbreviations, partial dates of birth, inconsistent address formats. It produces a ranked list of match candidates by confidence — in 1–2 minutes.
An OSINT Research Agent then searches Companies House, Google, LinkedIn, and Bloomberg against those candidates automatically, re-triggering when confidence is low. Per-candidate evidence records with full source citations — generated in 5–10 minutes, not 2–3 hours.
A Classification Agent renders the match/no-match decision with full causal reasoning and a confidence score. Not a black box — a structured explanation, every time.
For cases that don't resolve cleanly, an Escalation Agent routes to human analysts via structured human-in-the-loop, with the full agent reasoning already surfaced before any decision is locked. Analysts review genuine ambiguity. Not noise.
A Report Agent closes the loop: structured investigation report, timestamped audit trail, full provenance — in 1–2 minutes.

A unified operator interface gives compliance leads real-time visibility across every step. Individual alert decisions can be queried and re-evaluated directly from the interface. Nothing disappears into a black box.

The results

On a 300-customer test dataset:

0% false negative rate — no genuine matches missed
1% false positive rate — down from a 95–98% industry baseline
~5 minutes end-to-end case resolution per alert
50+ FTEs freed from low-value triage to focus on genuine risk investigation

What this actually changes

Agents handle 95%+ of false positives autonomously — escalating to humans only where genuine judgment is required. When analysts do step in, the full evidence chain is already in front of them.

The system maintains long-term memory of prior decisions, so accuracy improves over time rather than resetting with every new case. The audit trail deepens. The cost per case falls.

A compliance capability that scales without scaling headcount — and gets measurably better the longer it runs.