AI in banking calls for a decision for each use case. Between an assistant that helps an adviser find a procedure and a model that weighs on a credit decision, neither the data, nor the control, nor the regulatory framework are the same.
We propose to sort these use cases into four families: customer relations, fraud prevention and compliance, credit decisions, internal operations. Each is examined against three requirements: the data it consumes, the control exercised over its outputs, the texts to be examined. The order we propose starts from uses where a person reviews every output and moves towards those that affect a customer’s situation: internal operations, adviser assistance, triage of fraud alerts, credit last. This classification is a proposed method, with no standing as a reference framework.
The families of AI use cases in banking
- Customer relations: an assistant helps the adviser find a procedure or a pricing condition, prepares the minutes of a meeting, or answers the customer online itself.
- Fraud prevention and compliance: a model scores payment transactions, prioritises monitoring alerts or helps an analyst investigate a case.
- Credit decisions: a score estimates a borrower’s risk; other systems read the supporting documents and prepare the file.
- Internal operations: reading and classifying documents, reconciliations, searching documentation, support for software development.
These families mix predictive models, which score or classify, and generative models, which write: our comparison between AI and generative AI helps to place each case.
Data that the families do not obtain at the same price
An assistant, whether it serves an adviser or an internal team, answers from documents. If it works by RAG, it returns what the base contains, outdated versions and contradictions included: keeping the document base in order is its first data workstream.
Fraud detection has transactions in volume, but its labels arrive late: a case may be confirmed only after an investigation or a customer dispute. The model learns yesterday’s fraud, and its monitoring in production counts as much as its training.
A credit score is trained on accepted applications and their repayment. Rejected applications have no observed outcome: the training population has been filtered by past decisions. The choice of variables is examined with the lawyers, because a seemingly neutral variable can be correlated with a characteristic whose use would raise a question of discrimination.
In all four families, one check comes before the technology: the right to use this data for this purpose, the subject of our article on data governance for AI. The owners who decide on it are the ones put in place by data governance.
Control set according to what an error does to the customer
The criterion is the moment at which a person can set aside the system’s output. The adviser reviews the assistant’s answer before using it, and the error stops with the adviser if the review is genuine. The assistant that answers the customer has no one between it and the recipient: its scope of answers is bounded, a request that goes beyond it passes to an adviser, and a sample of the exchanges is reviewed.
A fraud alert lands in front of an analyst; a blocked payment affects the customer immediately. False alerts and the time taken to unblock are measured in the same way as the frauds detected.
For credit, control requires a person authorised to depart from the score, who records the reason. One measure helps to check this in each family: the share of outputs modified or set aside by those who review them. A share of zero may indicate a reliable system, or a review that no longer takes place.
The texts to examine with the lawyers and compliance
Regulation (EU) 2024/1689, known as the AI Act, refers in Annex III, point 5(b), to “AI systems intended to be used to evaluate the creditworthiness of natural persons or establish their credit score, with the exception of AI systems used for the purpose of detecting financial fraud”. Text quoted: consolidated version of 27 July 2026, read on 12 November 2026 in the Cellar repository. This quotation classifies no system: classification follows Article 6, including the derogation in paragraph 3, which our article on the high-risk AI system reads. Qualifying a given tool, a credit decision tool for example, is for the institution and its lawyers.
For high-risk systems placed on the market, put into service or used by financial institutions, Article 74, paragraph 6, designates as market surveillance authority the national authority responsible for their financial supervision, insofar as the use is directly connected with the provision of financial services; paragraph 7 allows the Member State, in appropriate circumstances, to designate another. Article 27, placed in Section 3 of Chapter III, requires deployers of a high-risk system referred to in point 5(b) to carry out a fundamental rights impact assessment before deployment. Under Article 113, third paragraph, point (c)(i), Sections 1 to 3 of that chapter, except Article 6, paragraph 5, apply from 2 December 2027 to systems classified as high-risk under Article 6, paragraph 2, and Annex III.
The GDPR, Regulation (EU) 2016/679, applies as soon as personal data is processed; its Articles 22 and 35 are read in our article GDPR and artificial intelligence. The DORA regulation raises two questions to examine with the lawyers, based on our reading of DORA for a banking project: is putting an AI system into service a major change, and is the model provider an ICT third-party service provider? The specific anti-money laundering framework, which we do not read here, is examined with the compliance function.
Keeping these answers up to date for the whole estate requires a register of systems and a classification grid, described on our AI governance page.
A proposed order, and the situations in which to refrain
Three criteria underpin this order: review by a person, the effect of an error on a customer, and the legal examination the case calls for.
- Internal operations first, on tasks whose results can be checked and corrected without exposing a customer.
- Adviser assistance next, as long as the adviser signs off what goes out.
- Fraud and compliance through the triage of alerts in front of an analyst, before any automation of a block.
- Credit last, or in parallel on its own timetable: errors there directly affect a customer, and the qualification of the tool is first examined with the lawyers.
It is better to refrain, or to reduce the scope, when no one is named to set aside an output, when the right to use the data has not been established, or when no business owner commits to how the expected gain will be used. This last point falls under AI strategy, which arbitrates between cases and sets their stopping criteria.
These use cases plug into an information system that has its own backlogs, which our article on the challenges of a bank’s digital transformation helps to identify; our banking and insurance page describes what supervision changes in the running of a project.
The order is revised with each result. An internal use case that meets its measures authorises the next one; a case whose reviewers set aside a large share of outputs points first to its data.