Banking & Lending

For banks, building societies, credit unions, NBFCs and fintech lenders, card issuers and the vendors whose models they use, and for the sponsor banks, capital providers, boards and counsel who rely on their decisions: an independent finding on one named AI decision, against the rule that binds it.

The decisions we examine

  • Credit – whether a loan or credit limit is approved, how much, and on what terms.
  • Fraud – which payments or applications are blocked, stepped up or declined.
  • Financial crime – which customers or payments are flagged, cleared or escalated in KYC and anti-money-laundering checks.
  • Collections – how customers in arrears are contacted and treated.
  • Customer contact – what customers are told and offered, including by chatbots and other AI agents.

What is changing

  • United States: SR 26-2 (17 April 2026), which replaced SR 11-7, sets model-risk expectations that cover AI models (Calendar). A lender that declines an application must give the applicant a statement of the specific reasons, whatever model made the decision (Regulation B, 12 CFR 1002.9).
  • United Kingdom: the PRA’s SS1/23, Model risk management principles for banks, current version effective 23 April 2026, covers the use of artificial intelligence in modelling (Calendar).
  • European Union: from 2 December 2027, AI used to assess the creditworthiness of individuals or to set their credit score is high-risk under the AI Act (AI Act, Annex III; Calendar).
  • India: the Reserve Bank of India’s draft guidance on model risk management calls for override, suspension or deactivation mechanisms, “including kill-switch arrangements” (paragraph 60(ii); comments closed; no date set; Calendar).
  • Canada: OSFI Guideline E-23 on model risk management, which covers AI models, takes effect on 1 May 2027 for federally regulated financial institutions (Calendar).
  • Singapore: the MAS Guidelines on AI risk management take effect on 7 October 2027 for financial institutions, banks included (Calendar).
  • The Calendar also lists rules and guidance that reach banks in Australia, Hong Kong, Indonesia, Malta, Qatar, South Korea, Switzerland, Taiwan, Thailand and the UAE, and data-protection rules on decisions based solely on automated processing in many more, each with its source.

These are short summaries of each rule, each linked to its source; they are not legal advice. A finding names, at intake, the exact rule it measures against.

Who asks for the evidence, and who pays

  • The bank or lender commissions and pays for a finding on its own AI; the chief risk officer usually signs, and the chief executive in a smaller lender.
  • A sponsor bank may commission a finding on a fintech partner’s system, and a warehouse lender, credit fund or securitisation investor on an originator’s, with that firm’s written agreement. The reviewed firm counts as a Fijishi client (Governance).
  • Its board or audit committee may commission the finding itself.
  • A credit bureau or a decisioning, fraud or AML vendor may commission a certificate on its own product, which its bank customers can check by number.
  • The third-party risk team commissions a Procurement Review before an AI contract is signed or renewed.

What each role finds here

  • Chief risk officer: which AI decisions a binding rule reaches, and whether the human override is real or only on paper.
  • Head of model risk: validation checks the model; a Fijishi finding checks one decision against the rule, end to end, including vendor models.
  • Chief credit officer: whether a declined applicant gets the true reasons, and whether a credit officer’s override is recorded.
  • Compliance and fair lending: whether decline notices are accurate for AI decisions, and what an examiner will ask first.
  • Financial crime and fraud: who is accountable when the AI clears an alert or blocks a customer, and who can stop it, per decision type.
  • Chief executive of an NBFC or fintech lender: evidence a sponsor bank, capital provider or investor can rely on.
  • Internal audit: an independent finding to rely on in its own work, never as its own opinion.

Which product answers which question

The proof you can check

What existing work covers, and what it does not

Management-system certificates such as ISO/IEC 42001 show that a firm manages AI through a documented system. Model validation and testing show whether a model works. Each has real value. Fijishi answers the question that comes after them: for one named decision, does it meet the rule that binds it, and can the firm show who was accountable and who could stop it? What already exists, and what each covers

Start

Write to protocol@fijishi.com with the AI system, the rule it must meet and any deadline. We reply within 5 business days. Intake has no fee; scope and fee are fixed in writing at intake, before any work begins. Both founders examine and sign every finding (Founders). Fijishi sells no insurance or other financial products, builds or sells no AI, and does no consulting work, so it has nothing to gain from what a finding says. Start an engagement

Insurance · Banking & Lending · Capital Markets & Wealth · Who a Fijishi finding is for · Start an engagement