AI in Bank Operations: The Human Factor

Banking operations have historically involved processes that relied on manual review. In the world of AI the same operations become dependent on a model, a prompt chain, a data feed and an automated workflow.

The difficulties arise in the event of a failure: the risk is that when setbacks occur, staff may no longer know how to perform the same processes manually.

Consider a few banking operations and AI related resilience risk.

• Take a contact centre, where AI routes complaints or vulnerability cases incorrectly.
• Loan or mortgage processing, where document extraction fails silently, delaying approvals or causing wrong decisions
• In payments operations, AI flags unusual transactions at scale but in the process overwhelms investigators.
• For regulatory reporting, AI-generated commentary conceals defects in data quality.

And AI can generate complacency around automation, where staff fail to check because the system is usually right. This is compounded where domain knowledge is lacking. The consequences are that errors become harder to detect.

To mitigate that risk, humans must be able to identify failure scenarios that models overlook and that requires intimate domain knowledge. In the case of operations, detailed knowledge of critical processes, manual workarounds, and business continuity is critical.

AI and Third-Party risk

Banking is a business that is more dependent on outsourcing than ever. But while the modern retail bank, for example, outsources large parts of its operations, the bank nonetheless remains responsible for operational risk, and so keen vigilance and due diligence is required. Aside from IT and cloud computing, significant outsourced operations include:

• Payments: Many retail banks outsource this function to specialist payments services providers (PSPs).
• Contact centers
• Printing, mailing and customer communications.
• Collections
• Mortgage servicing

Nowhere is third-party risk more apparent than the dependency of banks on external AI providers. And, as with all instances of systems outsourcing, human intervention is required to evaluate suppliers beyond their technical ability.

Naturally, evaluation of technical architecture will require deep understanding of systems dependencies and integration risk, but vendor management also requires operational domain knowledge to mitigate outsourcing risk, concentration risk, and to provide for systems exit planning.

When AI is technically correct but economically wrong

Historically, credit officers challenged credit models. Today, the same principle applies to AI and humans provide reasonableness testing to mitigate AI model risk.

So, in a credit context for example, that requires expertise in concepts such as leverage, cash flow, collateral, covenant structures and industry cycles.

So, if AI concludes that a borrower presents low credit risk, intervention by an experienced credit officer may notice that the customer operates in a weakening sector or that a major customer concentration exists.

Hence, the officer overrides the AI-based recommendation on the grounds that the AI may be statistically correct based on historical data but economically wrong.

Last, in an bank operations context, there is conduct risk. This is perhaps the area where human judgement is most important. AI remains relatively weak at understanding nuanced human circumstances. Humans are better at recognizing states such as vulnerability, distress, coercion or hardship.

So, when a customer defaults on a mortgage following a spouse’s death, AI will provide a technically correct approach to the issue, but a trained customer service officer will immediately recognize that the circumstances call for an entirely different response.

By Divya Pariti of Intuition Finance Digest.