SUNANDO ROY – On Banking, Finance and Society

Technology has always been central to banking, but the present phase of change differs in scale and character. Earlier waves of automation generally transferred established processes from paper to electronic systems. Artificial intelligence can go further: it can classify documents, identify patterns in large datasets, generate written analysis and support decisions that previously depended heavily on human interpretation.

The attraction is clear. Banks face high compliance costs, ageing technology infrastructure, pressure on margins and growing expectations for fast and personalised service. At the same time, they must respond to increasingly sophisticated fraud, cyberattacks and operational threats. AI is therefore being considered not only as a means of reducing staff time, but also as a way of strengthening monitoring, improving the consistency of decisions and using institutional data more effectively.

These expectations require some caution. Estimates of AI’s economic potential are not the same as evidence of benefits already achieved. McKinsey & Company (2023a) estimates that generative AI could create annual value of approximately US$200 billion to US$340 billion for the global banking industry, equivalent to around 9–15 per cent of operating profits. Its wider research estimates potential annual value of US$2.6 trillion to US$4.4 trillion across the use cases studied in multiple industries (Chui et al., 2023). These figures describe an opportunity under assumptions about adoption and implementation; they should not be interpreted as guaranteed savings.

For banks and supervisors, the more useful question is therefore not whether AI has potential. It is whether particular applications produce reliable, measurable and properly controlled outcomes.

Where AI Can Add Value in Banking

Process automation

A substantial share of banking operations consists of repetitive, rules-based work. Examples include document classification, account reconciliation, data extraction, customer-file reviews, report preparation and the initial screening of transactions or applications.

Traditional automation remains suitable where rules are stable and inputs are structured. AI becomes useful when the work also involves unstructured documents, variations in language or large volumes of information that cannot easily be captured through fixed rules. Optical character recognition, natural-language processing and machine-learning tools can, for example, extract information from contracts, invoices, financial statements and customer communications.

The strongest business case usually arises where the automated process has a high volume, a clearly defined output and a measurable error rate. Automation is less convincing where the underlying process is poorly designed or depends on unresolved policy judgments. Applying AI to a weak process may make the weakness faster rather than remove it.

 Credit underwriting and risk monitoring

Machine-learning models can analyse a wider range of variables and identify relationships that may not be apparent in conventional scorecards. They can assist with application screening, behavioural scoring, early-warning indicators, portfolio segmentation and the identification of borrowers whose risk profile is deteriorating.

These capabilities do not eliminate established credit-risk disciplines. Model performance remains dependent on the quality, relevance and representativeness of the data. Historical datasets may reflect earlier lending practices, economic conditions or biases that should not be carried forward. Models must also be monitored for deterioration when customer behaviour, products or economic conditions change.

Where AI materially influences credit approval, pricing, limits or collections, the bank should be able to explain how the output was produced and demonstrate that the model is operating within approved parameters. The use of a more complex model does not reduce the institution’s responsibility for the decision.

 Fraud detection and financial crime controls

Fraud detection is one of the more established uses of machine learning in banking. Models can analyse transaction velocity, customer behaviour, device information, location patterns and relationships across accounts. This may allow institutions to identify suspicious activity more quickly and reduce the number of legitimate transactions incorrectly flagged by static rules.

The technology also benefits criminals. Generative AI can make phishing messages more convincing, automate social-engineering attempts and support the creation of synthetic identities, false documents and deepfake audio or video. Shpachuk, Markova and Adamyk (2026) describe the resulting challenge as a dual one: financial institutions must use AI to improve detection while protecting themselves and their customers against AI-enabled fraud.

Banks should consequently avoid treating AI fraud controls as self-correcting systems. Investigators must be able to challenge alerts, identify changes in fraud typologies and recognise when a model is being manipulated or is no longer performing as expected.

Customer service

Conversational systems can answer routine questions, help customers navigate products and support service staff by retrieving relevant information. Used carefully, they may shorten response times and allow employees to focus on cases requiring judgment, empathy or specialist knowledge.

The risks increase when a system moves from providing general information to giving personalised financial guidance, interpreting contractual rights or making representations about product suitability. Generative systems can produce fluent answers that appear authoritative even when they are incomplete or wrong. Banks must therefore define the questions a system may answer, identify circumstances requiring transfer to a human employee and retain records sufficient to investigate customer complaints.

A customer-facing system should not be assessed only by its ability to reduce call volumes. Error rates, complaint outcomes, accessibility, conduct risk and the treatment of vulnerable customers are equally important.

 Regulatory compliance and reporting

AI can support regulatory-change monitoring, obligation mapping, document review, transaction surveillance and the drafting of routine reports. It may also help identify inconsistencies across policies, procedures and regulatory returns.

However, regulatory interpretation often depends on context, legal definitions, national discretions and supervisory expectations. A generative system may summarise a rule without recognising its exceptions, effective date or relationship with another provision. Compliance applications should therefore preserve links to authoritative source material and clearly distinguish machine-generated analysis from approved legal or regulatory interpretation.

 From Estimated Potential to Realised Benefits

The economic case for AI rests on several possible benefits: fewer hours spent on repetitive work, faster access to information, more consistent processing, reduced rework and earlier identification of risk. Yet institutions frequently find it easier to demonstrate an attractive pilot than to produce sustainable enterprise-wide benefits.

Scaling requires reliable data, integration with existing systems, clearly assigned ownership and changes to the way work is performed. A model that saves time in one part of a process may merely transfer additional checking or remediation work to another part. Similarly, a tool may appear efficient because the costs of validation, cybersecurity, vendor oversight and human review have not been fully included.

Benefits should therefore be tested through measures such as:

• processing time before and after implementation;
• error and exception rates;
• reduction in false positives;
• customer-resolution rates;
• number of cases requiring manual correction;
• operational losses or complaints associated with the system;
• cost of model validation, monitoring and remediation; and
• resilience during periods of stress or unusually high demand.

Institutions should also distinguish between productivity gains and head-count reductions. In many cases, AI changes the composition of work rather than removing the need for staff. Employees may spend less time collecting and formatting information but more time reviewing exceptions, challenging outputs and dealing with complex cases.

A SupTech Tool for LCR Analysis

A browser-based supervisory technology application for reviewing the Liquidity Coverage Ratio provides a practical illustration of controlled automation.

The Basel III LCR is intended to ensure that a bank has an adequate stock of unencumbered high-quality liquid assets to withstand a significant liquidity stress lasting 30 calendar days. The ratio compares the stock of HQLA with total net cash outflows over the stress period. In normal circumstances, the ratio should be at least 100 per cent. The Basel framework nevertheless recognises that banks may need to draw down their liquidity buffers during genuine stress and may temporarily fall below the minimum, subject to supervisory assessment (Basel Committee on Banking Supervision, 2013).

The application reconstructs the relevant return using a deterministic calculation engine. Level 1 assets are generally included without a haircut. Level 2A assets are generally subject to a 15 per cent haircut. Level 2B assets are subject to prescribed haircuts of 25 or 50 per cent, depending on the asset category.

The tool also applies the composition limits within the HQLA stock. After the relevant haircuts and prescribed adjustments, Level 2 assets may not constitute more than 40 per cent of total HQLA, while Level 2B assets are subject to a 15 per cent cap.

For the denominator, the application applies the required run-off and inflow rates to the relevant balance-sheet and off-balance-sheet items. Recognised cash inflows are generally capped at 75 per cent of total expected cash outflows. Subject to the applicable framework and any exceptions, this means that net cash outflows ordinarily cannot be reduced below 25 per cent of gross outflows merely through the recognition of inflows.

Once the return is loaded, the tool can:

• recalculate the reported LCR;
• identify inconsistencies in regulatory classifications or weights;
• flag breaches of the HQLA composition caps;
• compare the result with selected peer institutions;
• apply supervisory stress overlays to haircuts or run-off assumptions;
• identify the principal drivers of the ratio; and
• prepare a preliminary supervisory narrative.

The calculation engine and the AI-assisted commentary function should be treated as separate components. The regulatory calculation should remain deterministic and independently testable. Generative AI may assist in explaining the results, but it should not determine regulatory weights, alter the underlying calculation or substitute for the applicable rulebook.

A privacy-conscious design can further limit the information made available to the AI component. The application may process the return locally in the browser and provide the language model only with qualitative indicators—such as “below minimum”, “moderate buffer” or “high concentration”—rather than bank names, account-level data or precise figures. Where greater detail is needed, the model may be hosted within an approved internal environment.

This architecture reduces unnecessary data transmission, but it does not by itself make the application secure or reliable. Local processing does not address all risks arising from compromised devices, unauthorised access, malicious browser extensions, weak authentication or insecure local storage. The tool also remains dependent on:

• the accuracy and completeness of the uploaded return;
• correct mapping between spreadsheet lines and regulatory definitions;
• proper treatment of national discretions;
• controls over software versions and regulatory-rule changes;
• access and user-entitlement management;
• audit logging;
• independent testing of calculation logic; and
• human review of generated commentary.

The output should therefore be regarded as an analytical aid. Any supervisory conclusion must be checked against the underlying return, the applicable national rules and other information concerning the institution’s liquidity-risk profile.

The example reflects the broader development of SupTech and RegTech described by the Financial Stability Board (2020), particularly the use of more granular data and improved analytical infrastructure to support supervision and compliance. Its value lies not in removing supervisory judgment, but in allowing that judgment to be applied more quickly and consistently.

Cybersecurity, Model and Operational Risks

AI changes the threat environment in two directions. It can strengthen detection and response, but it can also increase the speed, scale and sophistication of attacks.

The U.S. Department of the Treasury (2024a) identifies several AI-related cybersecurity concerns for financial institutions, including more effective phishing and social engineering, data poisoning, weaknesses in third-party tools and the possibility that threat actors may automate parts of the attack process. Deepfakes can be used to impersonate customers, senior employees or service providers and may undermine controls that rely excessively on voice or video identification.

AI systems themselves may disclose confidential information through prompts, logs, outputs or vendor-controlled infrastructure. Employees may copy customer data, internal documents or supervisory information into publicly available services without appreciating how the information is stored or processed. Institutions should therefore control which tools may be used, what information may be entered and whether outputs can be incorporated into official work.

Third-party dependence is another concern. A limited number of cloud providers, model developers and specialist technology firms may support many institutions. A common vulnerability, service interruption or model defect could consequently affect multiple banks at the same time.

The IMF (2024) also points to broader financial-stability concerns. If institutions use similar models, datasets and trading signals, their behaviour may become more correlated. During stress, common model outputs could contribute to crowded positions, rapid repricing or herding. The precise effect remains uncertain, but the concern is sufficiently material to warrant monitoring by institutions and authorities.

Operational risk may also arise without a cyberattack. Models can produce inaccurate outputs because the input data have changed, an interface has failed, assumptions are no longer valid or users have applied the system outside its intended purpose. Generative models add the possibility of plausible but unsupported statements. Controls must therefore address routine failure as well as deliberate attack.

Governance for Responsible Adoption

AI should be governed through the institution’s established risk-management arrangements, with additional controls where its characteristics create new risks. Creating a separate AI committee or policy is not enough if responsibilities remain disconnected from information security, model risk, compliance, operational resilience and business ownership.

The NIST AI Risk Management Framework identifies four broad functions—govern, map, measure and manage—that can be used to structure the lifecycle of an AI application (National Institute of Standards and Technology, 2023). In a banking context, the following elements are particularly important.

Defined purpose and ownership

Each application should have an approved purpose, identified owner and documented boundaries. The institution should specify the decisions the system may support, the data it may process and the circumstances in which its use is prohibited.

Risk classification

Controls should reflect the significance of the application. A tool used to summarise internal meeting notes does not present the same risk as a system that influences lending, customer pricing, sanctions screening or regulatory reporting.

Data governance

The institution should establish the source, ownership, permitted use, quality and retention requirements for the data. Sensitive information should not be transferred to an external model merely because the interface is convenient.

Testing and independent challenge

Testing should cover accuracy, robustness, bias, security, explainability and performance under unusual conditions. Material systems should be subject to review by persons independent of their development.

Human intervention

Human review must be substantive rather than ceremonial. Staff should have sufficient knowledge, authority and time to challenge the output. Escalation and override procedures should be clear, and the organisation should monitor whether employees are becoming excessively dependent on automated recommendations.

Third-party oversight

Vendor assessment should cover data location, subcontractors, security arrangements, incident notification, model changes, service continuity, intellectual-property rights and exit planning. Contractual protections do not remove the bank’s responsibility for the service.

Records and accountability

The institution should be able to reconstruct how a material output was produced, which version of the model was used, what data were provided and who approved the resulting action. Accountability for customer, prudential and regulatory outcomes must remain with the bank.

Sum Up

Artificial intelligence can reduce repetitive work, improve the use of data and strengthen the speed of risk detection. It can also produce new sources of error, dependency and concentration. The outcome depends less on the label attached to the technology than on the design of the process in which it is used.

Banks are most likely to obtain lasting benefits where the application addresses a clearly defined problem, the underlying data are reliable, outputs can be independently checked and the institution is prepared to redesign the surrounding workflow. Supervisors should be cautious when benefits are supported mainly by vendor claims, pilot results or estimates that exclude the cost of controls and remediation.

Ref

Basel Committee on Banking Supervision. (2013). Basel III: The liquidity coverage ratio and liquidity risk monitoring tools. Bank for International Settlements.

Chui, M., Hazan, E., Roberts, R., Singla, A., Smaje, K., Sukharevsky, A., Yee, L., & Zemmel, R. (2023). The economic potential of generative AI: The next productivity frontier. McKinsey & Company.

Financial Stability Board. (2020). The use of supervisory and regulatory technology by authorities and regulated institutions: Market developments and financial stability implications.

International Monetary Fund. (2024). Global financial stability report, October 2024: Steadying the course—Uncertainty, artificial intelligence, and financial stability.

McKinsey & Company. (2023a). Capturing the full value of generative AI in banking.

National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1.

Shpachuk, V., Markova, O., & Adamyk, B. (2026). AI-driven financial fraud: Key risks and legal protections for financial institutions. Journal of Banking Regulation, 27(1), Article 6.

U.S. Department of the Treasury. (2024a). Managing artificial intelligence-specific cybersecurity risks in the financial services sector.

U.S. Department of the Treasury. (2024b). Artificial intelligence in financial services.