An end-to-end product case study for an agentic AI onboarding platform — designed with explainability, audit, and human-in-the-loop as first-class requirements, not afterthoughts.
Banks lose customers in onboarding. Manual KYC reviews take days. AML screening produces ~30% false positives. Existing automation stops at OCR — judgment work still routes to humans without context, audit trails, or explainability.
This case study designs an agentic AI onboarding platform that assists humans on judgment work, while keeping every decision auditable, explainable, and reversible.
| KPI | Industry Baseline | Design Target | Driver |
|---|---|---|---|
| Onboarding time | 2 days | < 15 min | Parallel agents, OCR-first capture |
| Manual review rate | 100% | < 20% | Confidence-gated auto-approval |
| AML false positives | ~30% | < 10% | Multi-signal risk scoring |
| Customer drop-off | ~40% | < 15% | Async re-entry, clear UX |
| Audit evidence assembly | hours | < 60 sec | Event-sourced audit log |
Seven bounded-scope agents. Deterministic stage gates. Human-in-the-loop above the risk threshold. Full audit trail per event.
Session context, account type, jurisdiction routing.
OCR + structured extraction with per-field confidence.
Jurisdiction rules, completeness, format checks.
Sanctions, PEP, adverse media via approved providers.
Multi-signal score 0–100 with factor breakdown.
Policy-driven routing to reviewer queue.
Event-sourced log; case evidence in < 60 sec.
| Decision | Alternatives | Why This Choice |
|---|---|---|
| Multi-agent architecture | Monolithic LLM | Per-step auditability; bounded failure |
| Human review for high-risk | Full automation | Regulatory mandate (FATF · EU AI Act Art. 14) |
| Risk score 0–100 | Binary approve/reject | Explainability; reviewer prioritization |
| Event-sourced audit log | Periodic snapshots | Reconstruct any decision on demand |
| Fail-closed on uncertainty | Fail-open with monitoring | Compliance non-negotiable |
Designed against named regulations — not generic "compliance" hand-waving.
Rec. 10 CDD · Rec. 11 record-keeping · Rec. 12 PEP · Rec. 24 beneficial ownership.
High-risk classification (Annex III §5). Art. 14 human oversight is a mandate, not a UX choice.
Art. 22 right to human review · Art. 25 privacy by design · Art. 32 security.
Identity verification within reasonable time · OFAC SDN screening · 5-yr retention.
Beneficial ownership ≥ 25%; ongoing CDD.
Predicate offences; criminal liability for AML failures. Reviewer override always justified.