KYC / customer due diligence (agentic)
An AI agent that extracts, verifies, screens, and compiles — autonomously.
The average bank spends $60 million annually on KYC — with large institutions spending up to $30M on new client intake alone. A single corporate KYC review costs $2,211 on average (up 17% since 2022). One manual corporate KYC case takes 40.3 hours of analyst time (PwC). KYC represents 31–50% of total compliance budget for most banks. (Sources: Consult Hyperion; Fenergo; PwC; Statista 2024)
Overview
Agentic KYC manages the entire due diligence workflow autonomously: document intelligence extracts data from unstructured identity documents, registry APIs verify beneficial ownership, sanctions screening checks against live watchlists, adverse media search surfaces reputational risk, and the agent compiles a complete compliance report. The agentic model consumes 5–30× more tokens per task than a standard GenAI interaction — a single KYC workflow may invoke 10–20 model calls across 100,000–500,000 tokens. Goldman Sachs reduced client onboarding time by 30% and saves thousands of manual labor hours weekly through Claude-based agents.
The Penalty Stakes
- A standard chatbot Q&A interaction: 1,500–3,000 tokens. A KYC agentic workflow: 100,000–500,000 tokens per case — covering document extraction, entity resolution, screening API calls, adverse media synthesis, and report generation
- At $0.40–$15/million tokens for frontier cloud models, a 500,000-token KYC case costs $0.20–$7.50 per run in API fees — multiplied across thousands of cases per day, cloud API costs become the dominant KYC cost driver
- EU AI Act Article 12 requires automatic event logging throughout the lifecycle of high-risk AI systems — agentic KYC decisions must maintain complete timestamped audit trails with 6-month minimum retention
- Document intelligence accuracy: modern ML pipelines achieve 98–99%+ MRZ extraction accuracy on high-quality passport images; deepfake detection and liveness verification are mandatory capabilities in 2025 (Onfido/Entrust, Jumio, Trulioo)
- Periodic review burden: high-risk customers require annual KYC refresh; medium-risk every 24 months; low-risk every 36 months — agentic AI automates the refresh trigger and evidence gathering, not just initial onboarding
Business Impact
40.3-hour manual KYC cases become minutes — client onboarding that took 61–150 days compresses to same-day or 3 days. Goldman Sachs reduced onboarding time by 30%, saving thousands of manual labor hours weekly. At $60M average annual KYC spend, AI automation delivers $18–30M in annual savings. Faster onboarding directly enables revenue from new client relationships.
EU AI Act Article 12 audit trail requirements for high-risk AI (including automated KYC decisions) take full effect August 2026 — €15M or 3% revenue penalties for non-compliance. FATF Recommendation 10 and FinCEN CDD Rule requirements for beneficial ownership verification cannot be compromised for speed. Manual KYC operations expose institutions to missed sanctions hits on periodic reviews.
Infrastructure Requirements
Agentic orchestration with persistent memory for multi-step state. Sub-ms state access via Redis. Heterogeneous compute for document intelligence (vision models), screening (fast ML), and report generation (LLM). Every agent decision auditable per EU AI Act Art. 12. Customer identity data never leaves on-premises infrastructure.
- On-premises orchestration and inference: NEXUS OS runs the full agentic orchestration layer and all model inference on-premises — customer identity documents, PEP/sanctions data, and beneficial ownership structures never reach third-party LLM APIs
- EU AI Act Art. 12 compliance by design: Complete timestamped audit trail of every agent decision, retrieved document, screening result, and report generation step — 6-month retention minimum, full reconstructability for regulatory examination
- NEXUS Foundry entity resolution training: Domain-specific entity resolution models trained on your historical CDD data — producing name matching and alias detection calibrated to your customer base's geographic and linguistic distribution
- Agentic token cost management: At 100,000–500,000 tokens per KYC case, self-hosted inference is 3–80× cheaper than cloud API pricing at scale — a $60M/year KYC budget at 30–50% AI reduction produces very different economics on Trinidy vs. frontier cloud APIs
- Perpetual KYC enablement: Trinidy's event-driven architecture supports continuous monitoring triggers (sanctions list updates, adverse media alerts, beneficial ownership changes) without waiting for scheduled review cycles — moving from periodic to perpetual KYC
- Deepfake detection integration: Document intelligence pipeline includes liveness detection and injection attack prevention as standard, keeping pace with 2025 fraud typologies targeting AI-assisted KYC systems