Opinion · Banking · The first AI auditAfter the underwriter
Banking has been automating customer-facing decisions for two decades.
It is about to be audited as if the underwriter were still signing.
A small-business credit application is declined inside a regional bank. The applicant calls the branch. The relationship manager apologises but cannot explain. The branch escalates. The escalation, three days later, produces a model name and a segment score. It does not produce a rationale a customer can read. The applicant goes to social media. The regulator goes to the supervisor's file. A version of this scene is happening every week, in every regulated bank, on every continent. Six months from now, the regulator will require an answer the architecture, as currently built, cannot give.
Why banking goes first
Banking is where the architectural argument about AI accountability stops being a regulator's question and becomes a procurement decision. Three reasons.
It has automated customer-facing decisions earliest, and at scale. Credit scoring, fraud detection, KYC and next-best-offer all run at machine speed at every tier-one and most tier-two banks. The underwriter at the desk has been the model in the rack for some time.
Its regulators have moved first, and concretely. The RBI in August 2025, SAMA in 2026, the CBUAE in February 2026, the EU AI Act from August 2026, and the Federal Reserve and OCC in April 2026 have each issued a rule, framework or guidance that requires the bank's architecture to do something the typical stack cannot do today.
Its P&L impact is measurable. Capital adequacy reflects model risk. Non-performing assets reflect credit-decision quality. Conduct fines reflect the accountability gap. The bank that builds accountability into the architecture earns measurable advantage on each. The bank that does not gets cited.
The five drafts on the same calendar
RBI's FREE-AI Framework, published in August 2025, lays down seven principles for AI in Indian financial services. The principle every Indian audit committee will recognise: an institution deploying an AI system is accountable for the decisions of that system, regardless of the level of autonomy. Larger NBFCs are now expected to maintain formal model inventories covering every model used in credit and collections. AI failures must be reported promptly. Consumers have the right to challenge AI-driven decisions.
The Saudi Central Bank has folded cloud and AI obligations into its 2026 annual revision cycle for licensed institutions, aligned with SDAIA's ethics principles. An AI architecture that cannot account for where the decision was made will fail the residency-driven adequacy assessment the Personal Data Protection Law imposes.
The Central Bank of the UAE, in February 2026, issued its Guidance Note on Consumer Protection and Responsible Adoption of AI and Machine Learning by Licensed Financial Institutions. It distinguishes human-in-the-loop, human-on-the-loop, and human-out-of-the-loop AI systems. The line every GCC compliance officer should commit to memory: fully automated credit or insurance decisions without the possibility of human intervention are unlikely to meet supervisory expectations. That is the GCC's first direct accountability rule for AI in banking.
The EU AI Act enters its high-risk phase on 2 August 2026. Annex III explicitly names credit scoring, and the pricing of life and health insurance, as high-risk AI systems. Conformity assessment, immutable audit logs, bias-assessed training data and human override capability are required. Deployer accountability does not transfer to the vendor. Fines run to 3 per cent of global annual turnover.
The Federal Reserve and the OCC, in April 2026, rescinded SR 11-7, the 2011 model-risk guidance that has shaped American bank model risk for fifteen years, and replaced it with a more explicitly risk-based, principles-driven framework. Generative and agentic AI are formally out of scope of the new framework. Supervisors are already applying its principles by analogy to AI-based underwriting assistants and customer-facing copilots.
Each draft is a different shape. All five are on the same compliance officer's calendar.
The three decisions, by stakes
The three decisions in Exhibit 1 share a shape but differ in stakes.
Credit is the most-regulated of the three. The EU AI Act names it in Annex III. RBI's right of challenge applies. SAMA and CBUAE consumer protection rules apply. The bank that cannot produce a rationale at the moment of the decline pays in NPA cycles when the decisions turn out poorly, and in customer churn when complaints accumulate. The Indian small-business segment, the Saudi retail credit segment and the Gulf SME book are all where the next eighteen months of audit attention will land hardest.
Fraud has lower individual stakes but is the largest single source of conduct provisioning at most banks. The architecture must produce not only the rule that fired but the basis for it and the customer's path to challenge. The false positive at the till, at the airport, at the worst moment of the customer's day is the most-shared frustration on social media. It is also the auditor's most-quoted example when the supervisor opens the file.
Customer service, at the individual level, has the lowest stakes of the three. At the population level, it has the highest audit risk in the bank. Every interaction is logged. Every decision is in the sample the regulator will inspect. The chatbot, the next-best-offer engine, the routing system are each producing decisions at machine speed, and very little of it is currently audit-grade. The bank that cannot defend the rationale on demand finds, when the inspection comes, that the rationale does not exist.
Banking's first AI audit is not coming. It is on the calendar.
What the bank that gets this right looks like
The architecture that survives this redrafting is built for it from the start, not retrofit after the regulator's letter arrives. Reason-coded decision logs produced as a by-product of every credit, fraud and customer-service decision, in a format the bank's existing model-risk management discipline can already ingest. Model risk documentation generated as a build artefact, ready for the audit committee on demand. Override authority encoded in the platform itself, with the escalation path visible to the supervisor and the customer's right of challenge attached to the decision record. An audit trail exportable in the regulator's preferred format on the day the supervisor asks for it.
The architecture that fails it tends to be the accreted one. A campaign tool that does not see the model. A model that does not know the segment. A segment that was last validated two model versions ago. A line-of-business owner who is named in the org chart and unreachable in the architecture. This is the buying discipline a CFO would call theatre. It will not produce an architecture that survives the regulator's draft.
| Accountability dimension | What it costs when the architecture cannot answer | What it earns when the architecture can |
|---|---|---|
| Model risk (provenance) | Higher capital add-ons as supervisors mark the model book as elevated risk. | Lower add-ons; faster supervisory cycle. |
| Credit decision (basis) | Higher NPA cycles when the bank cannot defend the rationale; provisioning revisits. | Tighter NPA performance and faster portfolio remediation. |
| Conduct (override) | Conduct fines and remediation costs when customer harm cannot be promptly remediated. | Reduced provisioning for conduct events; faster regulator engagement. |
| Customer trust (audit) | Complaint volume, social escalation, churn on the high-value segment. | Lower complaint volume; measurable churn reduction on high-value customers. |
| Operational | Cost of audit-led remediation: model retrofits, evidence-gathering exercises. | Audit completed without remediation; supervisor moves on. |
After the bank
The architectures that pass have four properties: an open architecture the bank can read and audit on its own terms, reason-coded decision logs as a by-product of every decision, outcomes-aligned pricing against decisions made rather than seats licensed, and operator-controlled deployment so neither the data nor the audit trail crosses a vendor's boundary. What this looks like when the customer is a policyholder or a patient, and the regulator is IRDAI, the FDA, the MHRA or the EMA rather than the RBI or SAMA, is the next question worth asking. The shape of the accountability question changes by sector. The architecture's answer does not.