AI Governance in KYC Banking: Deepfakes, Perpetual KYC, and SR 26-2's Scope Gap

Every bank that has run an RBI KYC inspection in the last two years has heard some version of the same finding: the model works, the file doesn't. Examiners are not asking whether your face-match or document-OCR tool is accurate. They are asking who validated it, what it was tested against, and who signed off on the exception queue. AI-assisted KYC has made onboarding faster — most large private banks now clear a clean digital KYC case in well under a day, down from the multi-day paper-and-branch cycle — but speed is not the thing regulators examine. Documentation is.

Deepfakes have moved from novelty fraud to a KYC control-design problem

Synthetic face and voice generation is now good enough to beat a static liveness check, and law-enforcement and industry fraud reports through 2025–2026 describe sustained, triple-digit year-on-year growth in deepfake-enabled onboarding fraud rather than a one-off spike. . The practical response is not "add liveness detection" — most banks already have that — it is layering active (challenge-response, randomized prompts) liveness on top of passive video analysis, because passive-only checks are exactly what pre-recorded and real-time-rendered deepfakes are built to pass. A bank that still relies on a single static selfie-match vendor for Video-based Customer Identification Process (V-CIP) under the RBI Master Direction on KYC is running a single point of failure against an attack class that is improving faster than most vendor contracts get renegotiated.

Perpetual KYC turns a periodic control into a continuously running model

Risk-based periodic KYC updation (every 2 years for high-risk, 8 for medium, 10 for low, under the RBI KYC Master Direction) is being replaced at several large banks by continuous, event-triggered re-verification models that watch transaction behavior and external data for material change. That is a genuine improvement over calendar-driven review, but it converts a periodic compliance task into a live model that needs SR 11-7-grade validation: documented sensitivity analysis on what counts as a "material change" trigger, back-testing against cases where a human reviewer would have caught something the model missed, and drift monitoring once the model starts inferring risk from data patterns nobody explicitly coded for. A perpetual-KYC model that silently stops flagging a customer segment is not a performance issue — it is an unreported change in the bank's KYC risk posture, and PMLA Rule 9 still holds the Principal Officer accountable for it.

Generative AI writing your KYC case notes is a documentation-integrity risk, not a productivity feature

Banks increasingly use generative AI to draft case summaries, enhanced due diligence narratives, and risk rationale text that sits in the customer file. . Whatever the formal scope turns out to be, an examiner reading a KYC file does not care whether the drafting tool falls inside a named circular; they care whether the narrative is accurate. A hallucinated employer name or transaction description in a case note is a documentation-integrity failure the moment it is filed, independent of which guidance technically governs the model that wrote it.

Risk-scoring opacity fails the Fair Practices Code test before it fails any AI-specific test

If a customer is declined onboarding or downgraded to enhanced due diligence by an AI risk score, the bank still owes that customer — and a DPDP Act 2023 data-rights request, and an RBI examiner — an explanation in terms a human can act on, not a confidence percentile. Banks that built their KYC risk engine around a black-box gradient-boosted model without a reason-code layer on top are retrofitting explainability under deadline pressure that could have been designed in at build time.

What a bank should actually do

1. Map every KYC AI touchpoint (liveness, document OCR, perpetual-KYC trigger model, case-note generation, risk scoring) against SR 11-7 validation requirements regardless of what any newer circular's scope turns out to be — don't wait for scope clarity to start documenting.
2. Add active, challenge-response liveness as a second layer on every V-CIP flow; stop relying on passive video analysis alone against deepfake-capable fraud.
3. Treat perpetual-KYC trigger models as live, monitored models with documented drift thresholds and a defined escalation path when trigger rates shift — not as a one-time build.
4. Require human review and sign-off on every AI-drafted EDD narrative or case note before it enters the permanent file, with the AI tool's role documented in the file itself.
5. Build the reason-code layer for KYC risk scoring now — before the first DPDP Act human-review request forces it under deadline.