As financial institutions expand artificial intelligence across operations, oversight, data quality and accountability are becoming central to the business case.
From promise to production
Artificial intelligence has become an increasingly practical subject for banks rather than a purely experimental technology. Institutions are examining applications in fraud detection, service operations, document processing, software development and risk analysis. Yet the transition from pilot programmes to routine use raises a more difficult question: can banks establish reliable controls around tools that may change how information is produced, interpreted and acted upon?
The Basel Committee's assessment of digitalisation describes both benefits and vulnerabilities associated with increasingly technology-dependent banking. Digital tools can improve services and efficiency, but can also expand operational exposure and dependencies on outside providers. That distinction is particularly important for AI, where a successful demonstration is not proof that a system will behave consistently under unusual conditions.
For banks, the commercial attraction is clear. Repetitive customer inquiries can be routed faster; internal research can be searched more efficiently; and analysts may spend less time consolidating large document sets. Those improvements, however, only create durable value when the accuracy of the output, the cost of integration and the accountability for decisions are understood. A model that reduces processing time but generates additional manual checking may offer a more modest net benefit than the initial trial suggests.
Why model governance matters
The risks differ considerably by use case. An internal tool that summarizes a public document has a different consequence profile from a tool contributing to credit decisions or communications affecting customers. The appropriate level of review should therefore depend on the system's purpose, access to sensitive data, potential effect on individuals and ability to trigger downstream action.
The Bank for International Settlements' Financial Stability Institute analysis of AI regulation identifies familiar challenges including data privacy, model risk and outsourcing, while drawing attention to generative AI errors and users' tendency to place excessive confidence in conversational systems. The report finds that many authorities are initially using established financial-sector rules rather than assuming that every AI application requires a wholly new regulatory framework.
A practical governance approach begins with an inventory of use cases, a named business owner and a documented decision about where automation is appropriate. That should be followed by testing on representative data, monitoring after deployment and an escalation process for errors. Human review is not simply a checkbox: staff need sufficient knowledge, time and authority to challenge questionable results.
Data and third-party dependencies
Model quality is partly a data-management question. Information pulled from multiple banking systems may contain inconsistent labels, outdated records or restrictions on how it may be used. When an AI system draws on that material, weak data controls can produce answers that appear coherent without being reliable. Data lineage, permission management and record retention therefore matter as much as the model interface.
The NIST Generative AI Profile provides a voluntary risk-management reference for identifying, measuring and managing concerns particular to generative systems. It can help organizations structure their own controls, but it is not a substitute for local banking obligations or institution-specific risk assessments.
Cloud and model providers add another layer. A bank may control its application while depending on an external provider for infrastructure, training updates or model access. Procurement teams consequently need to understand service continuity, testing rights, information handling, incident response and exit arrangements. Concentration risk can also arise when many institutions depend on a small number of technology providers.
Measuring returns without overlooking costs
A useful evaluation looks beyond the number of AI pilots launched. The central measures should reflect the job the system is expected to do: time saved per reviewed case, error rates, customer resolution quality, employee productivity or measurable reductions in operating friction. Results should be compared with a baseline and adjusted for the costs of data preparation, controls, vendor services and ongoing supervision.
There are limits to what existing evidence can establish. A time saving observed in one business line may not transfer to a different bank with older systems, different customers or more restrictive data policies. AI can also shift work rather than eliminate it, creating new tasks in model validation, monitoring and exception handling. Clear reporting of assumptions reduces the risk of confusing projected savings with realized performance.
Effective oversight also requires coordination between technology, compliance, operations, information security and frontline management. A model may work from a technical standpoint but remain unsuitable if its outputs cannot be explained to customers or audited by internal control teams.
What comes next
The next phase of AI in banking is likely to be defined less by dramatic demonstrations than by sustained operational results. Institutions that separate low-risk assistance from consequential automated decisions, preserve audit trails and test outcomes over time will be better placed to assess where the technology genuinely improves services.
The broader lesson is that innovation and governance are not competing objectives. Responsible deployment can provide the consistency needed to scale useful applications, while weak governance can turn seemingly inexpensive tools into sources of financial, operational and reputational risk. As banks expand adoption, the strongest evidence of progress will be measurable service improvement accompanied by clear accountability.
References
Basel Committee on Banking Supervision, Digitalisation of finance (2024) — https://www.bis.org/publications/digitalisation-finance
BIS Financial Stability Institute, Regulating AI in the financial sector (2024) — https://www.bis.org/publications/fsi-insight-63-regulating-ai-financial-sector-recent-developments-and-main-challenges
NIST, AI Risk Management Framework: Generative AI Profile (2024) — https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence
NIST, AI Risk Management Framework resources — https://www.nist.gov/itl/ai-risk-management-framework
Basel Committee, Digitalisation of finance - release — https://www.bis.org/media-releases/20240516-basel-committee-publishes-report-digitalisation-finance