KYC / AML Agentic AI — Capability Matrix

Skip to main content Skip to main content
Platform Capability Reference · Restricted Distribution . CAIBots © 2022–2026
Architecture V4.0 · 7 Sub-Agents · 5 Source Tools · Policy-Gated
Platform
Capability
Matrix
KYC / AML Agentic AI
Covered Now — Production Ready Phase 2 Roadmap Out of Scope
Document Class
Platform Capability Reference · Restricted Distribution
Audience
Chief Compliance Officer · BSA Officer · Head of AML Technology · Chief Risk Officer
Regulatory Scope
BSA · FinCEN · FATF · OFAC · FFIEC · SR 11-7 · CTA BOI (as amended 2025) · NIST 800-63-3
Architecture
V4.0 · 7 Sub-Agents Parallel · 5 Source Tools · Policy Gate KYC-2026.07 · HITL v5 (dual sign-off · committee) · 11 Scenarios + Live AI
01
Platform Overview

The Platform at a Glance

Continuously running compliance intelligence across three simultaneous time horizons
7
Sub-Agents
Simultaneous Parallel
5
Source Tools
Equal Authority
52
FATF Typologies
Pattern-Matched
3
Time Horizons
Simultaneous
<10s
Agent Pipeline
Runtime
4.5hr→55m
EDD Review
Time Reduction
90m→12m
CDD Review
Time Reduction
70%+
SDD Auto-Clear
Target Rate

The CAIBots KYC/AML Agentic AI platform is a continuously running compliance intelligence layer that operates across three simultaneous time horizons: new customer onboarding (CDD/EDD), perpetual KYC monitoring, and real-time transaction screening. Seven specialized AI sub-agents run in parallel against five peer data source tools. Agent pipeline runtime: <10 seconds end-to-end.

Mandatory HITL — Non-Negotiable

Every critical action is gated behind a mandatory human-in-the-loop checkpoint. The system cannot auto-file a SAR, auto-exit a customer, or resolve a sanctions hit without an authorized human decision. The agent prepares — the BSA Officer decides.

01 · BSA / AML Compliance Core — US Commercial & Institutional Banking

Capability
Status
Implementation Notes
Retail / Commercial / Institutional CDD/EDD
Covered
All three DD levels. Correct regulatory routing. 25%/10% UBO thresholds enforced. FinCEN CDD Rule 31 C.F.R. §1010.230 compliant. Automatic tier assignment — SDD (<30), CDD (30–65), EDD (>65 or any PEP/sanction adjacency).
Perpetual KYC — Event-Driven Monitoring
Covered
Score drift >15pts triggers automated refresh with no analyst request. Replaces calendar-based periodic review entirely. 60–70% pKYC workload reduction vs. traditional scheduled review programs.
52-Pattern Typology Transaction Monitoring (FATF/FinCEN-curated)
Covered
All 52 curated patterns: structuring, layering, TBML, bulk cash, MSB patterns, hawala (core). Connector architecture for Actimize/Verafin (institution-licensed, bidirectional by design). For structuring: calculates total avoided CTR reporting amount and maps full counterparty network.
OFAC / EU / UN / HMT Sanctions Screening
Covered
Real-time — not batch. Fuzzy name matching handles transliteration variants and name permutations. 2-hop PEP proximity screening. SDN cache TTL 1 hour — system never relies on stale list data for more than 60 minutes. Immediate blocking on any SDN hit.
SAR / CTR Filing Automation
Covered
Mandatory HITL approval gates all filings. Direct FinCEN BSA E-Filing post-approval. Agent never auto-files. BSA Officer is always the filer of record. 30-day SAR clock tracked with T-14 and T-7 escalation alerts. CTR auto-populated on qualifying cash events within 15 calendar days. OFAC blocking report filed within 10 business days of confirmed SDN hit.
FinCEN 314(a) / 314(b)
Covered
14-day window tracking. See Scenario 07 — live 314(a) batch match with BOI non-compliance (foreign reporting company — 2025 rule scope) and circular UBO detection. Automated search executed on every 314(a) batch receipt. 5-year lookback applied automatically. HITL gate for BSA Officer response decision. 314(b) voluntary sharing referral generated when Knowledge Graph identifies cross-institution patterns. See Scenario 09 — registration-first 314(b) response workflow with SAR-status disclosure structurally excluded.
Beneficial Ownership & CTA BOI (2025 rule scope)
Covered
Full UBO traversal to natural persons. Shell company, circular ownership, and nominee structure detection. FinCEN BOI cross-reference via the institution’s authorized access, where in scope (2025 interim final rule: foreign reporting companies registered in the U.S.). CDD-Rule beneficial-ownership certification discrepancy flags triggered automatically when certified BOI diverges from detected ownership.
Correspondent Banking Risk
Covered
See Scenario 06 — live correspondent banking EDD with FFIEC Chapter 13. SWIFT path risk assessment, nested correspondent chain analysis, HIFCA designation checking, de-risking workflow with advisory HITL for significant relationships. See Scenario 10 — De-Risking Committee: three signatures with rationale, individualized fair-access basis, confidentiality-safe exit letter. Multi-hop correspondent risk propagated through Knowledge Graph.
TM Alert Disposition (False Positives)
Covered
Scenario 11: alert investigated, structuring hypothesis tested and REJECTED with counter-evidence; close-with-evidence memo (5 pillars) → KYC Analyst concurrence; rule-tuning queued under SR 11-7 governance — never auto-applied. Answers the industry’s highest-volume pain: 90%+ of TM alerts are false positives.
Sanctions Near-Match Resolution
Covered
Scenario 12: strong-identifier match analysis (DOB, POB/nationality, documentation, aliases) reaches a documented NOT-the-listed-person determination; hold enforced until Sanctions Officer + BSA 4-eyes authorization; list-delta re-screen trigger set. Paired with Scenario 07’s true-match block: the system that proves a block proves a release.
Derived Risk Scoring (Engine of Record)
Covered
KYCRISK v1.0: score = round(Σ weight × dimension), weights published (geo .20 · product .15 · customer .25 · behavioral .25 · network .15); bands from policy thresholds (policy of record KYC-2026.07 #d3083beb84); every score computed at boot, never typed; audit line [risk@1.0] on every run.
Client-Side Session Verification
Covered
⛨ Verifier recomputes the full hash chain (SHA-256(prev | seq | ts | actor | msg)), reproduces all eleven risk derivations from published weights, and re-runs the SAR grounding check on the sealed narrative — in the reviewer’s browser, no server trust required.
Policy Threshold Sandbox
Covered
⚖ Live re-banding of all eleven scenarios against explorable SDD/EDD/SAR thresholds; statutory-class controls (sar_dual_review · ofac_dual_review · committee outcomes) display LOCKED with their dual-role matrices; exploration never persists; SR 11-7 change-management path stated.
Case Binder Export
Covered
⬇ One JSON export: policy-rules snapshot, engine block (weights + formula), per-scenario derivations, session grounding record, hash-chained ledger, and a written reconstruction procedure an examiner can follow offline.
PEP Screening & Adverse Media NLP
Covered
Dow Jones, Refinitiv World-Check, LexisNexis. 300+ monitored sources. 0–100 adverse media score with network propagation — UBO adverse media contaminates related-party network risk scores.
SR 11-7 Model Governance Framework
Covered
Back-testing (Gini >0.65 target), material change taxonomy with 7 defined change types, independent validation requirement framework. Full MRM documentation package included with every production deployment.
GDPR & Data Residency Architecture
Covered
PII boundary design — no PII in LLM prompts. DPIA template provided. EU regional deployment available (Pinecone Frankfurt, Neo4j AuraDB EU). Anthropic EU DPA with Standard Contractual Clauses available.
BSA Lookback Review Automation
Phase 2
Consent order lookback requires a distinct retroactive pipeline with different trigger logic and historical data traversal. Architecturally designed — not yet built. Estimated 8–12 weeks to production.
Real-Time Payments (FedNow / RTP)
Phase 2
Sub-100ms latency screening not yet architected. Agent-only pipeline: <10 seconds. Full CDD case (including HITL): <90 seconds end-to-end. Real-time payment rail screening requires dedicated low-latency inference path. Estimated 10–14 week build.
Capital Markets Surveillance
Out of Scope
Layering, spoofing, wash trading, MiFID II MAR — market abuse monitoring is a separate product domain from BSA/AML. Extension path available: 12–16 weeks as a parallel development track.
02
Competitive Differentiation

CAIBots vs. Traditional AML Platforms

Dimension Traditional AML
(Actimize / NICE / FircoSoft)
CAIBots
Agent Execution Sequential · Rule-based alerts 7 agents parallel · AI-synthesized risk
KYC Refresh Trigger Calendar-driven (annual / quarterly) Event-driven · >15pt risk drift fires instantly
Human Oversight Optional / configurable Mandatory HITL · Non-negotiable by design
Integration Model Replace existing stack Plugs into Actimize / Verafin · Zero rip-and-replace
SAR Narrative Generation Manual analyst draft AI-drafted FinCEN-aligned narrative + HITL approval
SR 11-7 Examination Package Manual documentation burden Auto-generated per case · Examiner-ready
Intelligence Architecture

5 Peer Source Tools

Equal authority · No hierarchy · Parallel execution — every trigger dispatches all five simultaneously
Architecture Principle

No source tool has authority over another. All five execute simultaneously on every trigger event. Conflicting signals are resolved by the orchestrator through weighted synthesis — not by tool priority. Every determination is traceable to a specific source tool, specific data input, and specific regulatory citation.

[VECTOR-RAG] Regulatory Playbooks Pinecone + pgvector DR
Technology: Pinecone (production, namespaced by LOB/regulation/jurisdiction) + pgvector (DR). Chunked 600–900 tokens with text-embedding-3-large, with retrieval indexed over FinCEN SAR narratives and BSA exam findings. MMR hybrid retriever with Cohere reranker · Top-8 chunks · Cross-encoder reranking.
Contents: FinCEN CDD Rule (2016), FFIEC BSA/AML Examination Manual, FATF 40 Recommendations + 52 Typologies, OFAC compliance framework, FinCEN guidance letters 2012–2025, state MSB regulations, institution's internal BSA/AML policy manual (institution-specific namespace), prior SAR filings (PII-stripped). Updated on every FinCEN guidance release and FATF mutual evaluation publication.
[API-CALLS] Live Screening Parallel Real-Time APIs
Vendors: Jumio / Onfido / IDEMIA (identity verification + biometrics) · OFAC SDN + Consolidated, EU/UN/HMT sanctions · Dow Jones + Refinitiv World-Check + LexisNexis Bridger (PEP/adverse media) · D&B ORBIS + FinCEN BOI + OpenCorporates (ownership). All results normalized to CAIBots Evidence Schema v2.
Real-time parallel API calls dispatched on every trigger event. Identity documents authenticated, biometric liveness confirmed (NIST 800-63-3 IAL2/IAL3 assurance level established), synthetic identity scored (Socure/Alloy). Sanctions lists queried with fuzzy name matching, transliteration variants, and 2-hop PEP proximity. All API calls execute in parallel — not sequentially — to meet latency SLA.
[SQL-DB] Customer Intelligence Snowflake Data Warehouse
Technology: Snowflake (read-only connection). CAIBots never writes to core banking data. Velocity aggregation windows pre-computed at 7d/30d/90d/12mo. Query SLA: <5 seconds for 36-month lookback.
Contains: 36-month transaction history with counterparty analysis and velocity windows, current and historical risk scores, prior KYC review outcomes with reviewer decisions, SAR/CTR filing history, FinCEN 314(a) match history, correspondent bank profiles with SWIFT BIC risk tiers, regulatory exam findings and MRA/MRIA tracker, prior bank exit records.
[KNOW-GRAPH] Ownership & Networks Neo4j AuraDB Enterprise
Technology: Neo4j (AuraDB Enterprise or self-hosted). Node types: Customer, Entity, Person, Account, Transaction, NewsArticle, RegulatoryAction. All relationships typed and weighted. Incremental updates from ownership change events and adverse media pipeline.
Models: UBO chains with 25%/10% threshold enforcement enforced at graph level, PEP → family member → associate → business entity networks with cross-jurisdiction linkage, correspondent bank chains with SWIFT routing path risk, SAR subject ↔ account ↔ counterparty linkages, adverse media → allegation type → entity clustering. Enables multi-hop risk propagation across the full relationship network — risk score contamination travels through UBO and network edges automatically.
[MEMORY] Compliance Intelligence 4-Tier Memory Architecture
Technology: Curated Knowledge Base (governed retrieval) + Redis (long-term KV) + In-Context (session state) + KV Cache (TTL-managed list data).
Tier 1 — Curated KB: Governed retrieval over the institution's closed KYC case library, SAR narrative style guide, and risk-appetite calibration — versioned and change-controlled under SR 11-7. This is the institution's "compliance voice" applied at retrieval time; no institution data is trained into model weights. Tier 2 — Long-Term/Redis: Prior case decisions, examiner findings, remediation patterns from MRA/MRIA tracker. Tier 3 — In-Context: Session state carrying risk score evolution across multi-trigger investigations. Tier 4 — KV Cache: OFAC SDN TTL 1hr, PEP list TTL 4hr, adverse media alert cache, 314(a) batch cache.
Book the Architecture Call

In a 30-minute architecture session we map this pipeline to your AML platform, data infrastructure, and compliance team workflow. A scoped 90-day production pilot can begin within two weeks of that session.

contact​@caibots​.com · +1 (609) 721-2815 · caibots.com · Princeton, NJ

Why every estate runs the same model. A SAR narrative is the most consequential artifact this suite generates — it is filed with FinCEN, it cites statute, and it is read by an examiner. Running it on a cheaper model than the credit memo or the research note would be a choice with no defence if asked, and inference cost is under half a percent of solution cost, so there is nothing to save. All four estates call claude-sonnet-4-6.

The SAR token budget is set at 2,400 rather than the 1,500 used elsewhere, because a narrative carrying the full FinCEN field structure will otherwise truncate mid-sentence. A truncated SAR is a worse failure than an expensive one.