From scoring models to fraud detection systems, artificial intelligence is already present in many banking processes. However, its rise does not turn an algorithmic recommendation into an infallible decision. For bankers, the challenge is to understand what the systems are actually measuring, to assess their limitations, and to maintain human accountability when decisions directly affect customers.
From Traditional Banking Expertise to AI-Assisted Decision-Making
Historically, the banking profession has relied on a combination of financial expertise, customer knowledge, risk analysis, and compliance with prudential regulations and customer protection standards. Depending on the specific role, this may involve evaluating a financing request, supporting a business, managing a wealth management relationship, monitoring transactions, or contributing to compliance efforts. Digital tools did not wait for artificial intelligence to transform banking: payment systems, rules engines, document automation, and statistical models have been in use for a long time.
AI now adds a new layer to this infrastructure. In this article, the term refers to several distinct families of systems: predictive models, used, for example, to estimate risk or detect anomalies; natural language processing systems, capable of extracting or classifying information; generative models, which produce text or summaries; and more agent-based systems, designed to perform a sequence of tasks using tools. These technologies do not share the same functions, levels of maturity, or risks.
Nevertheless, AI adoption has become widespread in the European banking sector. In September 2025, the European Banking Authority (EBA) reported that 92% of the European Union banks surveyed were already implementing AI use cases, while 8% were still in the pilot, testing, or discussion phase. The EBA specifically cites customer and transaction profiling, customer support, fraud detection, anti-money laundering, credit scoring, and creditworthiness assessment.[1] This figure reflects the prevalence of these applications, not their effectiveness or level of autonomy.
The European Central Bank has observed the same trend. Its supervisory work has noted an increase in the use of AI, particularly for credit scoring and fraud detection, while emphasizing that the potential of generative AI is still relatively new and requires specific oversight.[2] The paradigm shift, therefore, is not one in which a bank abandons human judgment, but rather one in which an increasing portion of the preparation, sorting, and analysis of information can be supported by algorithmic systems.
This development requires us to distinguish between three levels that are often conflated: generating a score, recommending an action, and making a decision. A model can estimate the probability of default. A system can use that estimate to suggest a course of action. But the legal, commercial, or financial impact of the decision depends on the process into which that output is integrated, the applicable rules, and the actual degree of human intervention.
How AI Is Transforming Banking Practices
AI does not transform all banking activities in the same way. Some applications are now relatively well-established, while others remain in the experimental phase or depend heavily on the quality of the institution’s data and information system. The real innovation, therefore, lies not in automation itself, but in the ability of certain models to learn from data, to detect patterns that are difficult to formalize using simple rules, or to generate content based on instructions.
Scoring and Credit Risk Assessment
Statistical and machine learning models can use a wide range of variables to estimate the probability of default or segment risk profiles. The ECB identifies credit as one of the AI use cases closely monitored by banking supervision.[2] However, a score remains an estimate generated within a specific context, not an absolute measure of a borrower’s future ability to repay.
Fraud Detection and Transaction Monitoring
Systems can detect unusual patterns in large volumes of transaction data and prioritize alerts for review. The EBA cites fraud and anti-money laundering as examples of how these systems are used by European banks.[1] Their usefulness depends on the trade-off between sensitivity and false positives, as well as on the teams’ ability to handle alerts without becoming overwhelmed.
Customer Insight and Personalization
Models can group customers based on their behavior, history, or preferences, and then support business segmentation. The EBA notes that profiling and clustering are among the common uses.[1] However, this personalization must remain consistent with rules regarding personal data, non-discrimination, and consumer protection.
Literature Review and Compliance
Natural language processing can help extract information from documents, prepare checks, or identify inconsistencies. In KYC and anti-money laundering processes, AI can assist with prioritization, but compliance is not limited to generating a score: it requires rules, audit trails, and escalation procedures.
Support for Advisors Through Generative AI
Language models can summarize a case file, rephrase an explanation, or search for information in a database. Their main advantage is the speed with which they provide access to information. Their limitation is just as significant: a smoothly phrased response may contain inaccurate or fabricated information. Outputs must therefore be linked to verifiable sources and proofread before being presented to a client.
Toward More Agent-Based Task Chains
Systems can be designed to perform a sequence of actions, such as searching for information, preparing a file, and forwarding a task to a business application. In a banking environment, the key question is not only whether an agent can take action, but what permissions they have, which operations require approval, and how each action is logged.
The transformation of banking work therefore involves less of a blanket replacement of activities and more of a redistribution of tasks. Highly structured operations can be further automated or pre-processed; ambiguous, sensitive, or exceptional situations, on the other hand, require greater contextualization, explanation, and accountability.
A New Role for the Banker
The integration of AI does not turn a banker into a data scientist. It does, however, reinforce several responsibilities: understanding where a result comes from, identifying situations in which a model might be wrong, explaining a decision, and knowing when to take back control. Professional value is therefore shifting in part from the execution of a procedure to the interpretation and supervision of systems that perform part of the analysis.
Interpreting without confusing the score and the decision
A risk score summarizes certain information according to a defined model. Bankers must be able to understand what the score measures, what factors influence it, and under what circumstances it becomes unreliable. Human intervention is meaningful only if it is genuine, competent, and capable of altering the outcome when the context warrants it.
Explain decisions to the client
When a decision significantly affects an individual, the issue of explainability goes beyond technical considerations. It also involves being able to present the relevant factors, avenues for appeal, and—when the framework provides for it—opportunities for human intervention.
Supervise generative assistants
An advisor may use an assistant to prepare a report or a response, but remains responsible for verifying the content submitted. An error in a summary of rates, fees, risks, or contractual terms can have real-world consequences for the client.
Participate in the governance of usage
Field professionals can identify recurring errors, biases, procedural workarounds, or situations that the designers did not anticipate. This feedback is essential for adjusting models, thresholds, and escalation rules.
Maintaining a Relationship of Trust
Algorithmic personalization is no substitute for an understanding of the economic or personal context. In sensitive situations—such as debt restructuring, project financing, cash flow difficulties, or wealth management—the quality of the interaction remains a key component of banking services.
The banker’s new role is therefore not simply that of an algorithm operator. It involves integrating these tools into a practice in which decisions remain justifiable, proportionate to the risk, and consistent with the institution’s obligations to its clients.
What Skills Do Bankers Need in the Age of Generative AI?
The fundamentals remain a priority: financial analysis, credit, customer knowledge, regulations, banking products, and risk management. AI does not reduce the need for expertise; on the contrary, it makes it more difficult to evaluate an algorithmic output when the professional does not have a firm grasp of the field to which it applies.
Technical and digital skills
- Distinguishing between different types of systems: a scoring model, a rule engine, an anomaly detection system, and a large language model do not serve the same purpose. Understanding this difference helps prevent us from attributing capabilities to generative AI that actually belong to other tools.
- Interpret the metrics with caution: accuracy, recall, false positive rate, calibration, and stability over time must be interpreted in light of the specific use case. A good average performance may mask significant variations across customer segments.
- Verify data and sources: The quality of a result depends on the data used. Outdated data, data entry errors, indirectly discriminatory variables, or changes in economic behavior can reduce a model’s relevance.
- Use generative AI without blindly trusting it: Generated content should be treated as a suggestion that needs to be verified. Sensitive information should only be shared with tools authorized by the institution and compatible with its security and privacy policies.
Analytical and decision-making skills
- Putting a recommendation into context: A model can detect a correlation without providing an economic or personal explanation for the situation. The banker must compare the model’s output with the relevant information in the client’s file and the client’s objectives.
- Acknowledge uncertainty: forecasts and scores are probabilistic. Knowing how to communicate a margin of uncertainty and identify borderline cases is more useful than presenting a result as a certainty.
- Knowing when to escalate: When a case is atypical, when the consequences are significant, or when the data is inconsistent, the right course of action is to pause the automation process and bring in the appropriate expertise.
Ethical, Legal, and Organizational Skills
- Understanding automated decisions under the GDPR: Article 22 of the GDPR provides a framework for decisions based solely on automated processing that produce legal effects or significantly affect an individual. It provides for specific exceptions and, in certain cases, safeguards, including human intervention and the right to challenge the decision.[5]
- Understanding the Status of Credit Scoring Under the AI Act: Assessing creditworthiness or assigning a credit score to individuals is classified as a high-risk use under the AI Act, with the exception of certain uses related to fraud detection. This classification results in a more stringent regulatory framework for the systems in question.[4]
- Integrating Operational Resilience: DORA, effective as of January 17, 2025, requires financial institutions to adopt a framework for managing risks related to information and communication technologies. Dependencies on models, infrastructure, or third-party providers must therefore also be analyzed from the perspective of continuity and resilience.[6]
- Participating in governance: Banks must be able to catalog their use cases, identify responsibilities, test systems, and track incidents. The front-line banker is not solely responsible for this governance, but becomes a participant in it when he or she uses or challenges a system’s results.
The new skill, therefore, is not just knowing how to use an AI interface. It involves understanding when the output is relevant, when it needs to be verified, and when it should not be used to make decisions.
Can artificial intelligence make banking decisions more reliable?
The answer depends on what we mean by “reliable.” A system can improve the detection of an event in a given context without improving all aspects of the quality of care. Reliability must therefore be examined from several perspectives: measurement accuracy, predictive performance, quality of integration into the department, and the teams’ ability to act appropriately on the information generated.
What the evidence allows us to conclude
Supervisory data show that European banks are already using AI for credit scoring and fraud detection. The ECB confirmed an increase in these uses between 2023 and 2024 and conducted workshops with thirteen banks in 2025 to examine the relevant practices in greater detail.[2] This indicates actual adoption, but does not prove that an AI model is consistently superior to a traditional method in all contexts.
The EBA also describes the widespread adoption of AI in the European banking sector and highlights both potential benefits in terms of efficiency and risk management, as well as challenges related to privacy, cybersecurity, compliance, and technological dependencies.[1] The results must therefore be evaluated on a per-use-case basis, using representative data and under conditions that closely resemble actual operations.
Credit is a particularly good example of this requirement. In France, the CNIL published a recommendation in May 2026 on credit granting, noting that Article 22 of the GDPR applies when automated creditworthiness assessments result in approval or denial without significant human intervention.[7] The question, therefore, is not merely whether the score is accurate, but how it is used in the decision-making process.
Limits to Watch For
- False positives and false negatives: In fraud detection, a system that is too sensitive can generate an excessive number of legitimate alerts; a system that is too permissive may allow fraudulent transactions to slip through. The choice of threshold reflects an operational trade-off that must be explicitly governed.
- Bias and discrimination: seemingly neutral variables may be correlated with protected characteristics or perpetuate historical inequalities. A generally acceptable overall performance is not enough: disparities between groups and segments must be identified and documented.
- Model drift: Economic behaviors, fraud patterns, and customer profiles evolve. A model that performs well at the time of deployment may lose its relevance if the data changes. Monitoring over time therefore becomes just as important as the initial validation.
- Lack of transparency and contestability: When an outcome significantly affects a client, a decision that is difficult to explain undermines the ability to challenge it and erodes trust in the process. The AI Act specifically requires a level of transparency that enables implementers to understand and correctly use the outputs of high-risk systems.4
- Hallucinations in generative AI: AnAI assistant may invent a figure, a rule, or a justification. This limitation makes it particularly risky to use generated text as the sole basis for financial advice, a compliance decision, or a contractual communication.
- Operational risk and supplier dependency: an outage, a change in the model, or an incident at a service provider can affect banking processes. DORA specifically strengthens the management of risks related to ICT and third parties within financial institutions.6
AI can therefore improve certain banking processes when it is properly designed, validated, and integrated. It does not, by itself, make a decision reliable. Reliability stems from the entire socio-technical system: data, models, business rules, human oversight, security, customer recourse, and governance.
What will the banking profession look like in the future with AI?
It would be premature to describe a single model of the “banker of the future.” Practices vary across retail banking, corporate banking, wealth management, risk management, compliance, and operations. Financial institutions also differ in terms of infrastructure, data, and levels of maturity. However, several trends are sufficiently evident to warrant examination, without presenting them as certainties.
- More automated case preparation: systems will be able to aggregate information, flag inconsistencies, and prepare summaries. The time saved will only translate into better advice if the organization actually reallocates that time to analysis and communication with the client.
- More continuous monitoring of risks: fraud, transaction anomalies, and changes in certain indicators can be tracked more frequently. This monitoring also increases the volume of alerts and therefore requires prioritization and control mechanisms.
- Generative assistants will play a greater role in customer service: they will be able to facilitate document searches or draft responses. Their deployment will require well-managed knowledge bases, confidentiality controls, and a precise definition of the situations in which a human must take over.
- Agent-based systems in controlled environments: some assistants will be able to perform multiple tasks in sequence. In sensitive processes, the key issue will be the permissions granted to the system: reading, suggesting, transmitting, modifying, or executing are not equivalent levels of autonomy.
- More structured AI governance: The ECB has included AI strategy, governance, and risk management among its supervisory priorities. In particular, it plans to continue targeted monitoring of banks’ generative AI applications.8
- Increased regulatory compliance requirements: Bankers will have to operate in an environment where banking regulations, data protection, digital resilience, and the AI Act all intersect. Not all AI systems will be subject to the same requirements; the use case, purpose, and level of impact will remain key factors.
We will also need to monitor the effects of automation on on-the-job training. If systems take over the analysis of simple cases too early, young professionals may have fewer opportunities to understand how credit assessments, risk analyses, or relationships of trust are built. Automating a task therefore does not eliminate the need to organize the transfer of the skills that the task helped develop.
The future of the profession will depend less on a scenario of replacement than on how institutions allocate responsibilities between tools and professionals. The more the system prepares and recommends, the more necessary it becomes to define who verifies, who explains, who can make exceptions, and who makes the final decision.
A Bank Enhanced by AI, at the Heart of Trust That Remains Human
Artificial intelligence can speed up data analysis, identify unusual patterns, and generate summaries. However, it cannot bear the economic consequences of a credit denial, an institution’s regulatory liability, or the relationship of trust with a customer. These responsibilities remain with people and organizations.
Responsible use therefore requires a clear purpose, appropriate data, performance metrics measured in a relevant context, sufficient traceability, and avenues for human intervention when the consequences so require. This requirement is particularly stringent when technology is used in the context of access to credit, an area that the AI Act considers sensitive with regard to fundamental rights.[4]
For bankers, the most significant transformation may be less dramatic than complete automation. It involves the ability to work with systems that provide more information and recommendations, while retaining enough critical thinking to recognize their limitations. The profession can thus shift toward providing more explanation, context, and judgment.
Technology can generate a score in a matter of seconds. It can also create an illusion of certainty. The banker’s role remains to understand what that score means, what it does not mean, and what consequences may result from its use.
The question to consider, therefore, is not so much “Will AI make decisions in place of bankers?” but rather: Will banks be able to use automation to improve the quality of their decisions without compromising explainability, customer recourse, and human accountability?
Learn more
To further this discussion on AI in finance, decision-making governance, and the evolution of skills, here are four additional posts from the aivancity blog.
Sources
[1] European Banking Authority. (2025). The Growing Use of AI in the EU Banking and Payments Sector. View the report
[2] European Central Bank, Banking Supervision. (2025). AI’s Impact on Banking: Use Cases for Credit Scoring and Fraud Detection. View source
[3] European Banking Authority. (2025). Special Topic: Artificial Intelligence. View the publication
[4] European Union. (2024). Regulation (EU) 2024/1689 establishing harmonized rules on artificial intelligence (AI Act), including Article 6 and Annex III. View the regulation
[5] European Union. (2016). Regulation (EU) 2016/679, Article 22: Automated individual decision-making, including profiling. View the regulation
[6] European Union. (2022). Regulation (EU) 2022/2554 on Digital Operational Resilience in the Financial Sector (DORA). View the regulation
[7] CNIL. (2026). Recommendation on the processing of personal data for the purpose of granting credit. View the recommendation
[8] European Central Bank, Banking Supervision. (2025). Supervisory Priorities 2026–2028. View the priorities