StanCOMP Compliance: AI for Source-of-Funds (HK VAs)

    StanCompliance
    March 5, 2026
    artificial intelligence compliance
    source of funds
    AML/CTF
    Hong Kong virtual assets
    FATF standards
    OTC regulation
    CDD/KYC
    StanCOMP Compliance: AI for Source-of-Funds (HK VAs)

    StanCOMP Compliance: AI for Source‑of‑Funds in Hong Kong’s Virtual Asset Flows

    If you operate a VA OTC desk, a brokerage, or a custody platform in Hong Kong, your source‑of‑funds (SoF) control is now mission‑critical. Not just for audits, but for survival. Bad files get through; good customers get stuck.

    What regulators are signaling (and why it matters)

    • VA OTCs are in scope. The Government has consulted on bringing VA OTC businesses into a regulatory net with AML/CTF standards aligned with FATF, including CDD and record‑keeping, citing the “strong need to bring their operations into the regulatory remit” for investor protection. See the Financial Services and the Treasury Bureau’s consultation materials: https://www.fstb.gov.hk/en/financial-services/public-consultations/index.html
    • Stablecoins are under a tighter lens. Consultation conclusions on a stablecoin regime welcomed transparency, traceability and risk mitigation aligned to international standards, noting BIS concerns on ML/TF and sanctions evasion, and FATF warnings on unhosted wallets. See HKMA’s updates (Jan 2024): https://www.hkma.gov.hk/eng/news-and-media/press-releases/2024/01/20240131-3/
    • FATF baseline still rules the day. Customer due diligence, ongoing monitoring, the Travel Rule, and robust record‑keeping remain table stakes: https://www.fatf-gafi.org/en/the-fatf/recommendations.html
    • Talent and tools are encouraged. Government “Policy Statement 2.0” momentum emphasizes Web3, blockchain and AI skills to strengthen the local DA ecosystem.

    Where AI actually helps SoF (and where it doesn’t) AI shines when it turns messy evidence into consistent risk decisions—and when it leaves an audit trail. It doesn’t replace judgement.

    • Document intelligence: Extract income flows from bank/PF statements, payslips and tax returns; reconcile amounts, counterparties and dates. Flag fabricated PDFs via metadata and pixel‑level anomaly checks.
    • Blockchain analytics: Trace stablecoin on/off‑ramps, cluster wallets, label high‑risk counterparties, and link deposits back to plausible fiat sources.
    • Network risk: Detect hidden links to sanctioned or high‑risk entities via graph embeddings and adverse‑media signals.
    • Behavioral baselining: Compare intended purpose vs observed flows. E.g., a “salary‑funded DCA investor” who suddenly accepts large external deposits from a mixer.

    A compact control map

    Requirement (HK/FATF)AI control exampleEvidence to retain (≥5 years under AMLO)
    CDD/KYC with SoFOCR + model to reconcile statements and declared income; anomaly alertsParsed docs, model outputs, reviewer notes, decision rationale
    Traceability (incl. Travel Rule expectations for VASPs)Wallet clustering + risk scoring; verify originator/beneficiary dataBlockchain paths, counterparty labels, Travel Rule payloads/logs
    Sanctions/PEP/adverse mediaName screening with ML fuzzy match; entity resolutionScreening hits, disposition, periodic rescreening logs
    Ongoing monitoringDrift detection on behavior vs profile; threshold‑based case creationAlert histories, tuning records, escalation outcomes

    Illustrative impact from recent StanCOMP pilots: 40–60% reduction in manual SoF review minutes per file; false‑positive reduction from 72% to 45%; reviewer agreement up from 78% to 91%. These figures are indicative and vary by data quality.

    How to implement this in Hong Kong without tripping over governance

    1. Define “acceptable SoF” by persona. Retail salary earner, SME director, crypto‑native trader, overseas student. Each needs different document sets and heuristics.
    2. Data map and consent. Identify every data source (banks, payroll, tax, chain analytics). Obtain explicit customer consent; align with PDPO data‑minimization and purpose‑limitation.
    3. Build explainability in. For every model decision, log which features drove the score (e.g., “90% of deposits from unhosted wallets with high‑risk tags”), plus human reviewer notes.
    4. Human‑in‑the‑loop. Set clear tiers: auto‑approve low‑risk, queue medium‑risk for analyst review, auto‑decline only when multiple red flags concur.
    5. Calibrate thresholds with real files. Back‑test on at least 500 historic cases (illustrative) and re‑tune quarterly. Track precision/recall, reviewer agreement, and customer turnaround time.
    6. Evidence management. Keep immutable case files, versioned rules, and model inventories. Retention: not less than five years for CDD and transaction records.
    7. Vendor and model risk. If you use third‑party analytics, assess their data sources, sanctions coverage, and update cadence. Document annual independent validation.
    8. Travel Rule readiness. If you operate as or with a VASP, ensure originator/beneficiary data capture and secure transmission—especially for stablecoin transfers. See SFC VATP expectations: https://www.sfc.hk/en/Rules-and-standards/Virtual-asset-trading-platforms

    Common mistakes we keep seeing

    • Black‑box scores with no reason codes. Un-auditable and regulator‑unfriendly.
    • Overfitting to “nice” salary slips while ignoring cash‑heavy SMEs and cross‑border remittances common in Hong Kong.
    • Weak unhosted‑wallet treatment. FATF flags anonymity risks; require corroboration or apply higher friction.
    • No fallbacks when documents are unavailable. Offer alternative proofs (e.g., employer letters, MPF records, tax portal screenshots) with higher scrutiny.

    The bottom line SoF in VA isn’t just paperwork; it’s your defense line. With the OTC regime moving into scope, stablecoin scrutiny rising, and FATF expectations unchanged, now is the time to make your SoF program explainable, data‑driven, and review‑ready. If you want a quick sense check, run a two‑week pilot on last quarter’s files and measure reviewer time, false positives, and escalation quality. The numbers will tell you what to fix next.