The Core Mental Model
The CAIBots KYC/AML Agentic AI is not a checklist tool or a rules engine with an AI wrapper. It is a continuously running compliance intelligence layer that sits between your customer data infrastructure and your BSA officer's desk. The fundamental shift: the analyst's job changes from researcher to decision-maker. The agent does all investigation, evidence assembly, regulatory mapping, and documentation. The human approves or rejects the conclusion. Every critical action is gated behind a mandatory human-in-the-loop checkpoint.
§1.1 · Visual Reference
Full System Architecture Diagram
Interactive diagram: all 8 layers from AML platform through BSA Officer outputs. Trigger horizons, orchestrator, 7 agents, 5 source tools, HITL gates, and compliance outputs — all in one view.
View Architecture →
§1.2 · Three Trigger Horizons Running Simultaneously
Onboarding
New customer application · Beneficial owner added · Authorized signatory change · Product upgrade (consumer → business banking) · New correspondent relationship
Perpetual KYC
Risk score drift >15 points · Adverse media NLP alert · Sanctions / PEP watchlist hit · Significant transaction velocity change · Ownership structure change · Source of funds anomaly · Scheduled periodic review due (12 / 24 / 36 months by risk tier)
Transaction Monitoring
AML rule engine alert (Actimize / Verafin / your rule engine) · CTR threshold event >$10K cash · Structuring velocity flag · High-risk jurisdiction wire · Correspondent banking anomaly · 52 FATF typology pattern match · FinCEN 314(a) batch match
Key Principle — Event-Driven, Not Calendar-Driven
pKYC is event-driven, not calendar-driven. A genuinely low-risk customer who never changes never consumes analyst time. Only customers whose risk profile actually moves generate work. This is the source of the 60–70% reduction in pKYC workload.
§1.3 · The 7 Sub-Agents — What Each Does in Real Time
Authenticates identity documents, validates biometric liveness, computes synthetic identity score (Socure/Alloy), and establishes NIST 800-63-3 IAL2/IAL3 assurance level. Detects synthetic identities, document forgery, and identity mismatch across channels. Runs in seconds against live identity APIs.
Traverses the full UBO chain to natural persons in real time. Enforces 25% threshold for standard CDD and 10% for EDD-elevated cases. Detects circular ownership, shell company layering depth, nominee structures, and trust/bearer share arrangements. Cross-references the FinCEN BOI registry under CTA 2024.
Continuous NLP across 300+ monitored sources. Classifies sentiment, allegation type, jurisdiction, and severity on a 0–100 adverse media score. Network propagation: UBO adverse media contaminates the customer's network risk score — guilt-by-association risk travels through ownership edges automatically.
Synthesizes inputs from all other agents into a single 0–100 score across five dimensions: (1) Geographic risk — FATF blacklist/greylist, HIFCA designations; (2) Product/channel risk — private banking, correspondent, wire-heavy; (3) Customer type risk — PEP proximity, cash-intensive, NBFI; (4) Behavioral risk — transaction velocity, source of funds; (5) Network risk — related-party SAR history, adverse ownership. Routes to SDD (<30), CDD (30–65), or mandatory EDD (>65 or any PEP/sanctions adjacency).
Real-time screening — not batch — against all major sanctions lists. Fuzzy name matching handles transliteration variants. Screens 2-hop PEP network proximity. OFAC SDN cache TTL is 1 hour — system never relies on stale list data for more than 60 minutes. Immediate blocking on any SDN hit.
Pattern-matches against all 52 FATF money laundering typologies in real time. For structuring cases, calculates total avoided CTR reporting amount, maps counterparty network, and identifies the precise period of suspicious activity. Evaluates SAR filing obligation under 31 U.S.C. §5318(g). Integrates bidirectionally with your existing AML rule engine.
Assembles the complete examination-ready compliance file: CDD narrative memo, EDD memorandum, SAR narrative (FinCEN-compliant format with activity description, subject identification, transaction schedule, and statutory citations), denial/exit letter, and the mandatory action checklist. Every finding cited to a specific regulatory source. Output structured for direct ECM archival.
§1.5 · What the BSA Officer Receives
When the pipeline completes, the BSA officer's workbench is pre-populated with a complete, examination-ready evidence packet. The officer's job is to evaluate and decide — not to research.
Complete identity confirmation with IAL assurance level and synthetic identity score
UBO structure diagram with traversal path, ownership percentages, and jurisdiction flags
Risk score (0–100) with five-dimension breakdown and historical drift chart
Adverse media score with source citations and allegation classification by type
Sanctions screening result (clear or hit) with list source and fuzzy match confidence score
Transaction pattern analysis with specific dates, amounts, and FATF typology mapping
Draft CDD or EDD memorandum — examination-ready format, with regulatory citations
Draft SAR narrative (if applicable) — FinCEN-compliant with activity description, subject identification, transaction schedule, and statutory citations (31 U.S.C. §5318(g), 31 C.F.R. §1020.320)
Mandatory action checklist with all applicable regulatory deadlines and escalation timers
HITL gate buttons: Approve · Escalate · Deny — every action timestamped and immutably logged
BSA Officer Efficiency — What Changes
EDD review time: 4.5 hours → 55 minutes (80% reduction). CDD review time: 90 minutes → 12 minutes (87% reduction). The analyst is no longer building the case file — they are evaluating a completed one. The 70%+ SDD auto-clear rate for low-risk retail means the majority of onboarding volume never reaches an analyst at all.
§5.1 · Risk Score Dimension Weights
Geographic
20%
10–30%
Increase for institutions with significant cross-border volume or HIFCA-designated market footprints. Decrease for purely domestic retail banks.
Product/Channel
20%
15–30%
Increase for institutions with private banking, correspondent banking, or crypto-adjacent product lines. Standard weight for basic retail/commercial.
Customer Type
25%
20–35%
Increase for institutions with high PEP exposure (international private banking), NBFI relationships, or MSB customer segments.
Behavioral
20%
15–25%
Increase for transaction-intensive businesses (MSBs, cash-heavy merchants). The primary signal for structuring detection.
Network
15%
10–20%
Increase for institutions with complex correspondent banking networks. Most computationally intensive dimension — increase only where network is a primary risk vector.
Mandatory BSA Officer Action
Risk score thresholds and dimension weights must be reviewed and signed off by the designated BSA Officer before any production case is processed. This sign-off constitutes the institution's documented risk appetite calibration under SR 11-7 model governance. Document the rationale for any deviation from default weights — examiners will ask.
Review the complete evidence packet. Choose: (A) Approve onboarding with enhanced monitoring, (B) Request additional documentation — specify what, (C) Decline and initiate exit. Document rationale in the case file. Cannot proceed without this decision.
Review the complete SAR narrative draft. Verify: activity description is accurate and complete, subject identification is correct, transaction schedule matches raw data, statutory citation is appropriate. Sign and submit via FinCEN BSA E-Filing. The agent never submits a SAR — the BSA Officer is always the filer of record.
Authorize the exit letter and confirm the documented rationale. Ensure exit does not constitute a tipping-off violation under 31 U.S.C. §5318(g)(2). Coordinate with relationship manager if required. Log the decision with full rationale in the case file.
On any OFAC SDN hit: immediately block the transaction/account, notify CISO and Legal within 1 business hour. Initiate OFAC blocking order. File OFAC report within 10 business days. Document all actions with timestamps. The agent executes the block — the officer manages the regulatory reporting.
Review the correspondent bank risk assessment and de-risking recommendation. Consider nested correspondent chain risk, SWIFT routing path exposure, and HIFCA designation status. Decision is advisory — the BSA Officer's judgment governs. Recommended: consult with Legal for significant relationships.
PII Boundary Architecture — The Core Design Principle
The CAIBots platform is architected as a PII-minimizing system. No customer PII is stored by CAIBots and no PII is transmitted in LLM inference prompts. SAR prompts contain transaction patterns (amounts, dates, counts), entity type classification, jurisdiction, FATF typology, and statutory references — customer names and account numbers are NOT included in prompts.
Customer Database (Snowflake)
PII Present
Source of truth. CAIBots has read-only access with column-level masking on direct identifiers (SSN, DOB, full account number). Only non-PII tokens (customer_id, amount, jurisdiction_code, risk_tier) transmitted to the agent pipeline.
Agent Orchestration Layer
Pseudonymised
All customer references use internal customer_id tokens. No name, address, SSN, or direct identifier in orchestration messages. Entity resolution happens at output rendering only, within the institution network boundary.
OFAC / Sanctions API Calls
Name Only
Only entity name and jurisdiction transmitted for screening. No account numbers or SSNs. Equivalent to a standard database search — covered by existing vendor DPAs.
LLM Inference (Anthropic API)
No Direct PII
SAR prompts contain: transaction patterns (amounts, dates, counts), entity type classification (MSB, LLC), jurisdiction, FATF typology, statutory references. Customer names and account numbers are NOT included in prompts. SAR narrative generated from structural case data, not raw customer PII.
Vector RAG (Pinecone)
No PII
Regulatory documents only: FinCEN guidance, FATF reports, FFIEC manual, internal policy. No customer data stored. Closed cases used for fine-tuning are PII-stripped before ingestion (§5.2).
Knowledge Graph (Neo4j)
Pseudonymised
Nodes use customer_id and entity_id tokens. Entity names stored only for external companies and public officials — already public information.
ECM Archive (OpenText)
Full Case PII
Final CDD/EDD/SAR file contains full PII as required by BSA record retention. Within institution ECM infrastructure. CAIBots writes the file; institution owns and retains it per 31 C.F.R. §1010.430 (5-year minimum).
Audit Log
Metadata Only
Contains: event_id, timestamp, agent_name, action_type, output_hash. No customer data. Tamper-evident and append-only.
SDD Auto-Clear Rate
>70% low-risk retail
If <65%: review risk score thresholds. If <60%: mandatory calibration review before next monthly cycle.
EDD False Positive Rate
<15%
If >20%: recalibrate risk dimension weights. Excessive false positives indicate over-sensitivity in scoring model.
SAR Draft Quality (A+B)
>85%
If <80%: trigger re-tuning cycle on SAR narrative model. Pause SAR draft use until threshold restored.
OFAC False Negative Rate
0% — Always
Any OFAC false negative is a Severity 1 incident. Immediate escalation to CCO, CISO, and Legal. Root cause analysis within 24 hours.
Pipeline Latency (CDD)
<90 seconds
If >120s average: investigate API bottlenecks. Scale orchestration layer if sustained above threshold.
HITL Decision Time (EDD)
<55 min mean
If >90 minutes: investigate workbench bottleneck or analyst capacity. Target is agent-enabled — delays indicate workflow issue.
Regulatory Deadline Breaches
0 — Always
Any missed SAR/CTR/OFAC/314(a) deadline is a Severity 1 incident. Immediate CCO and Legal notification.
Infrastructure
All external API credentials provisioned and tested (Jumio/Onfido, OFAC, Dow Jones, D&B ORBIS)
AML platform bidirectional connector live and validated with 10 synthetic test cases
Snowflake transaction history query performance validated (<5s for 36-month lookback)
Neo4j graph loaded with existing beneficial ownership data for all commercial accounts
Pinecone regulatory corpus loaded, embedded, and retrieval accuracy validated
Redis cache operational with correct TTL configuration (OFAC 1hr, PEP 4hr, 314a 24hr)
ECM archive (OpenText/Documentum) write connection tested with immutability confirmed
API key vault operational, keys rotated to production values, sandbox keys revoked
Audit logging confirmed: all agent actions and HITL decisions writing to immutable log
Network segmentation confirmed: api.anthropic.com accessible only from orchestration layer
Calibration
In-weights fine-tuning completed on closed case library (minimum 500 cases)
SAR narrative quality validated: >85% of BSA Officer ratings at B or higher
Risk score thresholds reviewed and signed off by BSA Officer (document reference: ____)
Risk dimension weights documented with rationale for any deviation from defaults
Parallel run completed: minimum 4 weeks, material disagreement rate <10%
False negative rate confirmed: <2% on all case types during parallel run
All 10 UAT scenarios passed with BSA Officer evaluation
Governance
AI Governance Policy adopted and approved by board or senior management
Model validation plan executed or formally scheduled within 90 days of go-live
HITL policy written and approved by BSA Officer and CCO
BSA Officer team fully certified (all passed 10-case certification test with >90% accuracy)
Third-party vendor due diligence completed for CAIBots, Anthropic, and all API vendors
Data Processing Agreements executed with all API vendors
FinCEN 314(a) response procedure updated to reflect agent-assisted search workflow
OFAC compliance policy updated to reflect 1-hour TTL cache protocol
SAR filing procedure updated to reflect agent-drafted narrative with mandatory BSA Officer approval
Sign-Off — All Required Before Production Traffic
BSA Officer signature obtained — Date: _____________
Chief Compliance Officer signature obtained — Date: _____________
AML Technology Lead signature obtained — Date: _____________
CISO signature obtained — Date: _____________
Data Governance Officer signature obtained — Date: _____________
Legal / Regulatory Counsel acknowledgment obtained — Date: _____________
Ready to Begin Implementation?
CAIBots implementation engineers are available for a 30-minute architecture scoping session. We will map this pipeline to your specific AML platforms, core banking systems, and compliance team workflow. A scoped 90-day production pilot typically begins within two weeks of the architecture call.
contact@caibots.com · +1 (609) 721-2815 · caibots.com · Princeton, NJ
