Key Takeaways
- Castellum.AI’s Arbiter AI agent platform, launched this week, integrates with LexisNexis Bridger Insight XG to resolve up to 95% of AML alerts across sanctions, PEP and adverse media checks.
- Arbiter cuts alert review time by 83%, according to the company, letting compliance teams concentrate human analysts on the cases that actually need judgment rather than routine triage.
- The platform produces audit-ready decisions calibrated to an institution’s risk thresholds, a design choice that targets the EU AI Act’s explainability requirements for high-risk AI systems, which take effect December 2, 2027. Castellum.AI says its Arbiter agent can close roughly 95% of sanctions and AML alerts without human intervention, and it is now doing that inside LexisNexis Bridger Insight XG, one of the screening environments compliance teams are already running. The integration, announced this week, is a concrete production deployment in a space where most AI announcements are still proofs of concept.
The Alert Backlog Problem
U.S. financial institutions filed roughly 4.7 million Suspicious Activity Reports in 2024, averaging around 12,870 per day. The manual workload behind those filings is the core problem Arbiter is targeting. Traditional rules-based screening generates enormous volumes of false positives, and compliance teams spend most of their time closing alerts that turn out to be nothing. Average annual AML/KYC spend sits at approximately $72.9 million per firm, according to reports, with technology adoption rising but uneven across institution size.
Global money laundering estimates range from $800 billion to $2 trillion annually, though figures at that scale are inherently difficult to verify. What is measurable is the operational drag: compliance headcount grows, costs compound, and the alerts requiring genuine analyst judgment get buried under routine casework. That is the bottleneck Arbiter is designed to clear.
How Agentic Triage Works
Arbiter sits at the first-level triage stage. When an alert fires, the agent gathers supporting data, applies the institution’s configured risk logic, and produces a decision with a full audit trail, without routing the case to a human analyst first. Castellum.AI reports the platform resolves around 95% of alerts this way, cutting review time by 83%, though those figures come from the company’s own testing rather than independent benchmarking.
The audit trail is the detail that matters most for regulated institutions. Regulators are not just asking whether AI can close alerts faster; they are asking whether the decision logic is traceable and reviewable. Arbiter’s design, decisions configured to institution-specific risk thresholds, with documented adjudication steps, is a direct response to that pressure. FIS built a comparable agent with Anthropic targeting AML investigation time, though its architecture routes differently through the analyst workflow.
Beyond Alert Triage
Castellum.AI is not the only team working this problem. On the transaction monitoring side, AI models score risk dynamically rather than applying fixed thresholds, which lets them catch structuring activity and mule account networks that rules-based systems routinely miss. Napier AI markets an AI-native compliance platform built around real-time monitoring and risk scoring. Alessa offers AI-driven sanctions, watchlist and PEP screening designed to cut false positive rates. Unit21 focuses on risk and compliance workflow automation, including alert investigation.
Sanctions and PEP screening carry a specific false positive problem: transliterated names across languages produce high alert volumes in cross-border operations. Fuzzy-matching algorithms paired with contextual entity profiles, pulling from global registries, internal notes and media, improve precision here, though the accuracy gains vary considerably by data quality and configuration.
What Regulators Are Requiring
The EU AI Act classifies most financial risk-scoring applications as high-risk, with full requirements taking effect December 2, 2027. That means documented validation, defined controls, independent oversight and explainability by design, AI systems must produce traceable decision logic that auditors can follow. The EU Anti-Money Laundering Authority is expected to issue guidance on how AI-generated risk scores must be documented and reviewed, though that guidance is not yet finalised.
The Financial Action Task Force encourages AI use for AML on a risk-based approach, with data privacy and security compliance as baseline requirements. The regulatory direction is consistent: AI is acceptable, but the decision trail must be human-readable. That requirement shapes which architectures compliance teams can actually deploy rather than just pilot. For a broader picture of where agentic AI sits in the regulatory outlook for retail finance the FCA’s Mills Review is worth reading alongside these AML-specific rules.
Deployment Realities
The Arbiter-Bridger integration matters partly because it avoids ripping out existing infrastructure. Compliance teams do not need to rebuild their screening environment, the agent runs inside the workflow they already operate. That lowers the switching cost substantially, which is usually where AI deployments stall in regulated industries.
The harder constraints are data quality and integration depth. An agent that resolves 95% of alerts in a clean, well-labelled dataset may perform differently against a legacy data environment with inconsistent entity records. Institutions evaluating these platforms should pressure-test the resolution rate against their own data before treating vendor figures as operational targets. Some banks piloting AI-driven AML tools have reported that integrating agents into existing review workflows exposes gaps in process design that predate the AI deployment entirely.
Tier-1 banks were widely reported to be planning increased AI budgets for AML modernisation through 2026, though the specific figures behind those projections vary by source and are difficult to verify independently. The operational case for automation at scale is clear enough without the numbers: at 12,870 SARs per day, any reduction in per-alert handling time compounds quickly.
Originally published at https://autonainews.com/castellumais-new-agent-platform-cuts-aml-alert-review-83/
Top comments (0)