Shadow AI in Banking: Detecting Unauthorized Models Before They Become Regulatory Findings

Ask any model risk team how many AI models the bank runs, and you'll get a confident number. Ask them how many AI tools employees and business units are actually using, and the honest answer is "we don't fully know" — and that gap is shadow AI. . The real governance problem isn't the dollar figure anyway — it's that a bank with unsanctioned AI in production cannot produce a complete model inventory on demand, and that is now a named finding examiners write up.

There are two completely different shadow AI problems, and most banks only govern one of them

Employee-level shadow AI — someone pasting customer data into a personal ChatGPT or Gemini account to draft a memo faster — is a data-leakage and DPDP Act 2023 exposure that acceptable-use policy and DLP tooling can mostly catch. Institutional-level shadow AI is the harder problem: a business unit procuring an AI vendor tool outside the model risk committee's intake process, AI embedded inside a software product the bank already licenses, or an analyst's own Python script doing AI-assisted scoring that nobody logged as a model. A bank that has solved the first problem with a good acceptable-use policy and still has no idea how many AI features are embedded in its licensed software stack has not solved shadow AI — it's solved the easy quarter of it.

A complete AI model inventory is now the actual examination bar, not a best practice

Model risk guidance built on SR 11-7's conceptual-soundness and ongoing-monitoring requirements has always implicitly required a complete inventory — you can't validate what you haven't listed. . Whatever the newest circular says, "we can't produce a complete AI model inventory" is already an answer no bank wants to give an examiner, because it implies the bank doesn't actually know its own AI risk surface — that's not a technical gap, it's a governance credibility problem.

Detection has to cover identity, network, cloud billing, and procurement — not just the endpoint

Device and endpoint monitoring catches an employee typing into a browser-based AI tool. Network and API traffic analysis catches calls to known AI vendor endpoints even from sanctioned apps. Identity and access log review catches new SaaS AI tools provisioned through SSO without a formal onboarding ticket. Cloud provider and billing log review catches AI API consumption buried inside an existing cloud bill line item nobody is reading closely. And procurement and contract review — the layer banks skip most often — catches AI features bundled into a vendor product renewal that nobody flagged as "this is now an AI system" because the contract calls it a "smart" or "intelligent" feature instead.

India's shadow AI entry points are specific, not generic

WhatsApp Business API integrations with third-party AI chat layers, regional-language AI tools adopted by relationship managers to handle vernacular customer queries faster, and AI embedded inside fintech-bank co-lending or BNPL partnership platforms are the three pathways where Indian banks consistently find shadow AI they didn't provision. Each of these sits on customer personal data covered by the DPDP Act 2023, and each is exactly the kind of AI the RBI's 2024 circular on AI expects to show up in a bank's own inventory — a fintech partner's AI model making a lending decision on the bank's balance sheet is the bank's model risk, not the fintech's, regardless of whose engineers built it.

What a bank should actually do

1. Build one AI inventory that spans both employee-level tool usage and institutionally procured/embedded AI — stop treating these as separate problems with separate owners.
2. Extend shadow AI detection past endpoint monitoring into cloud billing logs and vendor contract review, where embedded AI features hide longest.
3. Require every business-unit AI procurement to pass through model risk intake before contract signature, not after deployment.
4. Audit fintech and BNPL partnership platforms specifically for embedded AI decisioning on the bank's balance sheet, and bring that AI into the bank's own inventory regardless of who built it.
5. Treat "we cannot produce a complete AI model inventory" as an active examination risk today, independent of which specific circular eventually codifies the requirement.