Contribution · Application — Finance

AI for Sanctions and AML Screening

Sanctions and AML screening suffers from brutal false-positive rates — 95%+ of alerts turn out to be nothing, buried under name fuzz and dual-use ambiguity. LLMs can dramatically reduce triage time by reading news, adverse-media hits, and transaction context to decide whether 'Ramesh Kumar' is the Ramesh Kumar on the list. But the model is a decision-support aid, not a decider — every true-positive still needs a compliance officer and a SAR/STR filing decision.

Application facts

Domain
Finance
Subdomain
Compliance
Example stack
Claude Opus 4.7 for narrative reasoning and news summarization · Sanctions list feed (OFAC, UN, EU, MHA, SEBI debarment) · Fuzzy matching (NameAPI, Jellyfish, Elasticsearch) · LangGraph workflow with compliance-officer approval · Audit log with rationale and citations

Data & infrastructure needs

  • Live sanctions and PEP lists — updated daily
  • Customer KYC and beneficial ownership data
  • Transaction monitoring rules and alerts
  • Adverse media sources (vetted)

Risks & considerations

  • False negatives — missing a real sanctioned entity is a serious violation
  • Bias in name matching — Indian, Chinese, Arabic names generate more fuzz
  • Prompt injection via transaction memo fields
  • Regulatory — RBI Master Direction on KYC, PMLA, FATF, OFAC
  • Tipping off — accidentally warning a subject during investigation

Frequently asked questions

Is AI for sanctions screening safe?

As an alert-triage copilot, yes — it saves compliance officers hours and surfaces the genuine hits. As an autonomous decider, no. Every escalation-worthy alert must be reviewed by a human compliance officer, and SAR/STR filing decisions remain with the BSA/AML officer.

What LLM is best for sanctions screening?

Claude Opus 4.7 for careful, cited reasoning. Deploy in-VPC or private cloud with a DPA that forbids training on customer data. Consider fine-tuned smaller models for the high-volume name-match reasoning step.

Regulatory concerns?

India: RBI Master Direction KYC, PMLA, FIU-IND reporting. US: BSA, OFAC, FinCEN. EU: AMLD6, EBA guidelines. All require documented risk-based approach and model risk management (SR 11-7 in US).

Sources

  1. FATF — accessed 2026-04-20
  2. RBI — Master Direction KYC — accessed 2026-04-20
  3. OFAC — US Treasury — accessed 2026-04-20

Related

  • AI for Credit Scoring Explainability — LLMs translate complex ML credit model decisions into plain-language adverse-action notices and turn internal model explanations into customer-facing reasons consistent with fair lending laws.
  • AI for Invoice Processing Automation — Accounts-payable teams use document AI plus LLMs to extract invoice fields, match to POs, code to GL accounts, and route approvals — achieving straight-through processing with audit trails.
  • AI for Public Records and FOIA/RTI Triage — Agencies use LLMs to triage public records requests, identify responsive documents, suggest redactions, and draft response letters — reducing backlog while respecting access-to-information laws.
  • All Applications — Browse the complete applications library
  • VSET — Programmes offered — B.Tech programs (CSE, AI&ML, AI&DS, Electronics, IIoT)