A change committee at a mid-sized bank. Three requests for machine learning budget are on the table. Credit risk wants to replace the logistic regression in its personal loan scorecard with a gradient boosting model, because on historical data it separates good and bad borrowers better. Fraud wants a model that scores every instant payment before it leaves the bank. Retail wants an assistant in the mobile app to take pressure off the contact centre. Each request promises impact in the millions. The budget and the data people will cover one of them this year.

The CIO usually picks the one with the biggest promised benefit. That tends to be credit scoring, because every point of discriminatory power turns into lower credit losses or more sales. A year later the model is still waiting for validation, compliance is asking how rejected applicants will be given reasons, and internal audit wants to know whether a change in how customers are ranked counts as a material model change.

This article is for bank CIOs and the people who prepare these decisions for them. It shows which machine learning use cases in banking have measurable impact and what supervisory requirements each of them triggers, in the EU and in the US. The thesis may sound counterintuitive: the order in which a bank delivers ML should follow the cost of oversight and how quickly impact can be measured, not the size of the promised benefit. Credit scoring often pays the most, but it is the wrong first project. Start with fraud detection, build model risk management on it, and only then move into decisions about customers.

Why is “where should we use ML in the bank?” usually asked in the wrong order?

Because use cases are compared on benefits, while cost is counted as the cost of building the model. In a bank, building the model is the smaller part of the work. The larger part is everything that has to exist around it: validation by an independent function, documentation, monitoring, a way to explain decisions to customers, a model change procedure and answers for the supervisor. Those costs vary between use cases far more than the cost of training.

Take two models of similar technical complexity. The first scores fraud risk on a payment and raises an alert for an analyst. When it is wrong, the analyst loses a few minutes or the customer gets an extra authentication step. The second assesses creditworthiness and declines an application. When it is wrong, the bank refuses credit to a specific person who has a right to ask why, and the supervisor has a right to check that the model does not discriminate. Same algorithm, entirely different cost of surroundings.

The second dimension is time to measurement. A fraud model gets feedback in days: the payment turned out to be fraud or it did not. A credit model gets feedback in months, because defaults emerge with a lag. If a bank starts with a project whose impact it will see in eighteen months, it spends eighteen months not knowing whether its ML process works at all.

So I suggest a simple matrix. One axis is the cost of oversight, from low (the model supports a person in an operational task) to high (the model decides about a customer or about capital). The other is time to a reliable measurement of impact. Pick the first project from the “low oversight cost, short time to measurement” quadrant, even if it promises less. Its real product is not only a model but the machinery built around it: a model inventory, a validation process, monitoring. Later projects reuse that machinery and get cheaper.

This is not an argument against ML credit scoring. It is an argument for introducing it in a bank that already knows how to oversee models, rather than making it the project on which the bank learns.

Where do banks actually use machine learning, according to supervisors?

Before choosing a use case, look at what supervisory data says rather than vendor decks. The European Banking Authority, in its follow-up report on machine learning for IRB models (EBA/REP/2023/28, August 2023), summarised, citing its autumn 2022 risk assessment questionnaire, that the main areas of ML use in banking are fraud detection, AML/CFT, client and transaction profiling and clustering, and real-time monitoring of payments. In credit risk, EU institutions using ML mainly apply it to creditworthiness assessment and credit scoring for credit granting.

The same report shows what banks are not yet doing at scale. For internal models used to calculate capital requirements, most institutions still use variants of regression, and decision trees, random forests or other ML techniques are used by fewer institutions. The EBA attributes this partly to doubts about whether such models can comply with IRB requirements. Bear the sample in mind: the EBA received 14 consultation responses and itself says the conclusions should be taken with caution.

A broader picture comes from the Bank of England and FCA survey published on 21 November 2024, with 118 responding firms. 75% were already using AI, up from 58% in 2022, and a further 10% planned to within three years. 55% of use cases involved some degree of automated decision-making, but only 2% were fully autonomous. A third of all use cases were third-party implementations, up from 17% in the previous survey. And the number that matters most to a CIO: 46% of firms reported only a partial understanding of the AI technologies they use, against 34% reporting a complete understanding.

Three conclusions follow. First, ML in banks is not an experiment but a daily reality in operational areas. Second, fully autonomous decisions are rare: the model usually advises, and a person or a rule decides. Third, the share of models bought from vendors is growing, and with it a problem supervisors have raised for years: the bank is accountable for a model whose code and training data it often cannot see. I will come back to that under model risk management.

Why is fraud detection the best first use case?

Because it meets three conditions at once: the problem is large and countable, feedback arrives quickly, and the model can work as a ranking for a person instead of making a decision about a customer.

The scale is documented. The joint EBA and European Central Bank report on payment fraud from August 2024 found that payment fraud reported across the European Economic Area amounted to EUR 4.3 billion in 2022 and EUR 2.0 billion in the first half of 2023. In that half-year, fraudulent credit transfers sent by EU/EEA payment service providers totalled EUR 1.131 billion and fraud on EU/EEA-issued cards EUR 633 million. Card payments had the highest fraud rate, at 0.031% of value.

Two details in that report matter more than the totals. More than half of the value of fraudulent credit transfers came from manipulation of the payer, where the customer authorises a transfer to the fraudster’s account. Strong customer authentication does not stop that, because the genuine customer is the one authenticating. And for credit transfers, payment service users bore more than 80% of fraud losses. This is exactly the gap where a behavioural model has an edge over a rule: it recognises that this customer has never sent this amount at this time to a new payee, and it does so while the instant payment has not yet left.

The time pressure is rising. The EU Instant Payments Regulation required payment service providers in euro area member states to offer sending of instant euro credit transfers and verification of payee from 9 October 2025. When a payment settles in seconds, there is no time for manual review, so the risk decision has to be automated, and the model has milliseconds, not hours.

There is also a regulatory argument that is rarely made in committee. Annex III, point 5(b) of the EU AI Act, which classifies creditworthiness assessment as high-risk, explicitly excludes AI systems used to detect financial fraud. A fraud model is still subject to GDPR, supervisory guidance and internal model risk management, but it does not trigger the full set of high-risk obligations. That is no excuse for carelessness: a model that freezes a customer’s account affects them in a very real way. But the oversight cost is clearly lower than for scoring.

An example from the US public sector shows the order of magnitude. On 17 October 2024 the US Department of the Treasury announced that its enhanced fraud detection processes had prevented and recovered more than USD 4 billion in fiscal year 2024, including USD 1 billion recovered by expediting the identification of Treasury check fraud with machine learning. Treasury is not a bank, but the mechanism is the same: the model does not replace the investigation, it points it to where the probability of fraud is highest.

One note on the boundary of this article. Payment fraud detection and AML transaction monitoring are related but separate systems, with different rules and different alert handling. Here I deal with payment fraud. The first step of the AML process, customer identification, is covered in the article on building a compliant KYC process.

How do you measure a fraud model’s impact without fooling yourself with a detection rate?

The most common mistake is reporting one number to the board: “the model catches 92% of fraud”. That number means nothing without the other one, namely how many legitimate payments the model stopped along the way. With fraud rates in hundredths of a percent, a model that flags 5% of all payments will almost certainly catch most fraud and paralyse customer service while doing it.

An honest set of metrics has four layers. Value of fraud prevented in euros or dollars, not in transaction counts, because a few large authorised push payment frauds outweigh hundreds of small card frauds. Alert precision, the share of alerts that turn out to be fraud, which determines how many analysts you need. Customer friction, measured as legitimate payments declined or held, abandoned transactions and complaints. Decision latency, because for instant payments a decision after the fact is worthless.

Convert all of this into one currency. For each decision threshold, calculate the value of fraud prevented minus the cost of handling alerts minus the estimated cost of lost and abandoned transactions. Plot that result against the threshold and you will see where the model makes money and where it stops. The business chooses that threshold, not the data scientist.

The second condition of honest measurement is comparison with what the bank has today. The model does not compete with zero; it competes with the existing rules engine. The standard is a few weeks in shadow mode (the model scores payments but does not affect decisions), followed by a champion–challenger test on part of the traffic. Keep a small control group in which some low-risk alerts are not blocked but reviewed afterwards. Without it, the bank does not know how much fraud the model lets through, because it only sees what it stopped.

The third condition is awareness of delayed and incomplete labels. Card fraud surfaces when a customer disputes a transaction, sometimes weeks later. Some fraud is never reported. A model trained on “confirmed fraud from last month” learns from an incomplete picture, and its results for the most recent weeks always look better than they are. Report impact for periods whose labels have had time to settle, and tell the board how long that window is.

If the bank builds this measurement for its first model, it has a template for every later one. The same logic (value in currency, comparison with the current process, control group, label settlement window) carries over to scoring, forecasting and customer service. Translating a model metric into a business metric and planning an online test are covered in more depth in the article on when to hire machine learning consulting.

Why does credit scoring deliver the most value and carry the most expensive oversight?

The value is real. The EBA’s 2023 report confirms that ML techniques can improve a model’s discriminatory power and help identify relevant risk drivers, especially where risk differentiation has historically been poor and relationships are non-linear. Respondents described using random forests and gradient boosting to select risk drivers and clustering techniques to set PD and LGD score ranges. Better ranking of customers by risk means that at the same loss rate a bank can approve more applications, or at the same approval rate it can lose less.

That last framing is the right business metric for scoring: approval rate at a constant loss rate, or loss rate at a constant approval rate. A few points of Gini improvement are a success for a data scientist and meaningless to a board until someone converts them into additional loans or avoided impairments. Doing that requires data the bank often lacks: it does not know how applicants rejected by the old model would have repaid. This is the classic reject inference problem, and every honest scoring business case should say how it was handled.

Now the cost. For EU banks, consumer credit scoring triggers several regimes at once. The EBA Guidelines on loan origination and monitoring (EBA/GL/2020/06, applicable since 30 June 2021), in section 4.3.4, require institutions using automated models for creditworthiness assessment to understand their methodology, input data, assumptions, limitations and outputs, and to have procedures for detecting and preventing bias and ensuring input data quality, measures for traceability, auditability and robustness, regular assessment of output quality including backtesting, and control mechanisms, model overrides and escalation procedures. Add GDPR Article 22, the AI Act’s high-risk requirements and national supervisory expectations on model risk.

For US lenders, the core obligation is older and very concrete. Under the Equal Credit Opportunity Act and Regulation B (12 CFR 1002.9), a creditor must notify an applicant of adverse action, generally within 30 days of receiving a completed application, and the statement of reasons must be specific and indicate the principal reasons for the action. The official interpretation adds that disclosing more than four reasons is not likely to be helpful, and that for credit scoring systems the reasons must relate to factors actually scored. In Circular 2022-03 of 26 May 2022, the CFPB stated that creditors cannot use complex algorithms if doing so prevents them from giving specific and accurate reasons, because the requirements apply to all credit decisions regardless of the technology used. The CFPB withdrew a large batch of guidance documents in May 2025, so check the circular’s current status with counsel; the regulation it interprets has not changed.

Each of these regimes demands something a fraud model needs far less: an explanation at the level of an individual decision that can be given to a customer, an analysis of impact on protected groups, documentation readable by someone outside the data team, and independent validation before first use. If a bank builds those capabilities from scratch on its scoring model, it builds them under the most pressure: on a revenue-generating product, under supervisory scrutiny, with sales waiting for results.

None of this means ML scoring is a bad idea. It means its cost is dominated by the model’s surroundings, not by the model. A bank that already has a model inventory, a working validation process and explanation tooling from a fraud project will deliver ML scoring faster and more cheaply than a bank that starts with it.

What does the EU AI Act require of credit scoring, and where do the deadlines stand in November 2026?

The classification is unambiguous. Annex III, point 5(b) of the AI Act covers AI systems intended to evaluate the creditworthiness of natural persons or establish their credit score, with the exception of systems used to detect financial fraud. Point 5(c) covers risk assessment and pricing for natural persons in life and health insurance, which matters for bancassurance groups.

The deadlines have moved. The high-risk requirements for Annex III systems were originally due to apply from 2 August 2026. Following the Digital Omnibus amendment, Chapter III, Sections 1 to 3 (risk management, data and data governance, technical documentation, record-keeping, transparency, human oversight, accuracy and robustness, among others) apply to Annex III systems from 2 December 2027, and to AI embedded in Annex I products from 2 August 2028. As of November 2026, a bank operating in the EU therefore has just over a year to prepare its ML scoring for these requirements.

There is an important nuance for models already in production. According to the implementation timeline, obligations for Annex III systems placed on the market before 2 December 2027 apply to them only if they undergo significant changes in their design from that date. It sounds like a loophole, but in practice it is a trap. An ML scoring model that has to be retrained on new data will sooner or later go through a change that is hard to call insignificant. A bank counting on an old model to “wait it out” is only postponing the problem to the first serious recalibration.

Article 86 also matters. It gives a person affected by a decision that a deployer takes on the basis of output from an Annex III high-risk system, where the decision has a significant adverse effect on their rights, the right to clear and meaningful explanations of the role of the AI system in the decision-making procedure and the main elements of the decision. Combined with GDPR, this means explainability for scoring is a requirement to design for, not an option to consider.

Not everything waits until 2027. The transparency obligations in Article 50, including telling people they are interacting with an AI system, apply from 2 August 2026. That affects digital channel assistants, discussed below. What Article 14 requires of human oversight, and how to design it, is covered in the article on where a person must stay in an AI process.

One question remains open. In its 2023 report the EBA argued that it should be clarified that the high-risk category covers creditworthiness assessment of natural persons at loan origination, not, for example, IRB models used to calculate capital requirements. If a bank uses the same model for both purposes, it should take the more cautious reading and ask counsel about current Commission guidance before assuming either use is out of scope.

What does explainability mean in practice: GDPR Article 22, SCHUFA and Dun & Bradstreet?

For years, EU banks treated GDPR Article 22 as something that could be designed around: a credit officer formally makes the decision, so it is not “based solely on automated processing”. Two judgments of the Court of Justice of the EU have weakened that approach considerably.

In its judgment of 7 December 2023 in Case C-634/21 (SCHUFA Holding), the Court held that the establishment of a credit score by a credit information agency must be regarded as an automated individual decision, prohibited in principle by the GDPR, in so far as that agency’s clients, such as banks, attribute a determining role to it in granting credit. The lesson for a CIO: if a credit officer in practice approves what the model shows, a human signature does not change the nature of the decision. What counts is the model’s real role, not the org chart.

In its judgment of 27 February 2025 in Case C-203/22 (Dun & Bradstreet Austria), the Court described what an explanation must look like. The controller must describe the procedure and principles actually applied in such a way that the person can understand which of their personal data were used and how. The Court said it could be appropriate to inform the person of the extent to which a variation in the data taken into account would have led to a different result. Merely communicating an algorithm is not a sufficiently concise and intelligible explanation. Where the controller considers that the information includes trade secrets or third-party data, it must provide it to the supervisory authority or court, which will balance the interests involved.

This moves explainability from data science good practice into product design, on both sides of the Atlantic. A counterfactual explanation (“had your monthly debt payments been EUR 200 lower, the application would have been approved”) is exactly the type of information the Court pointed to, and it maps well onto the specific principal reasons that Regulation B requires. But it has to be designed: choose variables the customer can influence, limit the output to a few principal reasons, and check that explanations are stable, so that two similar applications do not receive completely different reasons for decline.

What tools do banks use? The EBA’s 2023 report found that among respondents the most common interpretability tool was Shapley values (40%), followed by enhanced reporting and documentation of methodology (28%), graphical tools (20%) and sensitivity analysis (8%). Shapley values are a good tool for a validator, but they are not yet an explanation for a customer. A list reading “contribution of variable X: −0.17” meets neither the C-203/22 standard nor the Regulation B standard. It has to be translated into the language of the decision, and that translation has to be checked for fidelity to the model.

The practical conclusion: before choosing a scoring algorithm, a bank should decide on the format of the customer explanation and verify that the chosen model can deliver it reliably. The EBA loan origination guidelines say the same thing more plainly: if explanations are required during the model’s use, consideration should be given to developing an interpretable model. Sometimes a well-engineered regression beats gradient boosting not because it predicts better but because it can be defended.

What does the EBA expect from ML in IRB models, and why shouldn’t a bank start there?

Internal models used to calculate capital requirements are the most regulated place a bank can use ML. The EBA’s 2023 report is very specific here and worth reading before any decision.

First, banks use ML selectively in IRB: mainly for PD estimation in the risk differentiation phase, less for LGD, EAD and CCF. Risk quantification, which needs long-run averages, does not require ML, and new data sources often lack the history anyway. Second, ML appears in validation as a challenger model: for robustness analysis, challenging the variables of the final model and benchmarking. That is the cheapest and safest entry point for ML into IRB, because a challenger does not calculate capital.

Third, the main difficulties are overfitting (especially in low-default portfolios), skills and explainability. The EBA expects comparison of performance on the development sample with out-of-sample and out-of-time results, and validation of hyperparameters by appropriately trained validation staff. Institutions that responded generally did not outsource the development of ML models, and where they did, only for the development phase, partly because of concerns about intellectual property.

Fourth, and most important for system architecture: model changes. The EBA states that any update of the rating system used for own funds requirements must be assessed under the prudential model change framework, and that a change in the algorithm assigning obligors to grades may require supervisory approval. For models updated frequently or automatically, two requirements stand out: a significant change in rank ordering or in the distribution of exposures is a material model change, and no change may be split into several changes of lower materiality. At the same time, the EBA clarifies that as long as the rating system used for capital is not updated, the framework does not constrain self-training challenger models that incorporate new data.

For a CIO the conclusion is architectural. An IRB model cannot retrain itself in production. Every retraining is a version that goes through a materiality assessment and potentially an approval process. The bank’s ML platform must therefore distinguish models that may be updated automatically (challengers, some fraud models) from models where every version is a regulatory decision. If that distinction lives in the team’s heads rather than in the tooling, sooner or later someone will ship a new capital model through the ordinary CI/CD pipeline.

So IRB is a poor place for a first ML project. It is a good place for the third or fourth, once the bank has proven versioning, validation and monitoring. Until then, ML can work in IRB as a validator’s tool, not as the production model.

When does liquidity forecasting benefit from ML, and when does the model fail exactly when it is needed?

Cash flow forecasting rarely makes it into AI presentations, yet it often delivers measurable impact. Every day a bank forecasts ATM and branch withdrawals, deposit inflows and outflows, and intraday liquidity needs in payment systems. Every forecasting error costs money: too much cash in ATMs is idle money and extra cash-in-transit runs, too little is unhappy customers. An oversized liquidity buffer is foregone interest income.

ML has good conditions here. Data is plentiful and high-frequency, and patterns depend on the calendar, paydays, holidays and local events, the kind of non-linearities that gradient boosting or recurrent networks capture better than simple moving averages. Impact can be measured directly: forecast error against the current method, converted into buffer size, logistics cost or funding cost. Treasury or cash logistics makes the decision, not the customer, so the oversight cost is moderate.

There is, however, a line that must not be crossed. A model trained on history forecasts well in conditions that resemble history. A liquidity crisis is by definition a situation that was not in the data. The Federal Reserve’s review of Silicon Valley Bank, published on 28 April 2023, describes how SVB lost over USD 40 billion in deposits on 9 March 2023, and management expected to lose over USD 100 billion more the next day, roughly 85% of the bank’s deposit base. For comparison, the review cites estimates that Wachovia’s failure in 2008 involved about USD 10 billion in outflows over 8 days, and Washington Mutual’s USD 19 billion over 16 days. No model trained on calm years would have predicted that dynamic.

The conclusion: ML is a good tool for operational forecasting in normal conditions and a bad substitute for stress testing. Stress scenarios must remain expert constructs, consistent with supervisory expectations for liquidity risk, with assumptions that can be debated at ALCO. An ML model can feed those scenarios, for example by segmenting deposits by behaviour, but it should not replace the question “what if customers behave differently from anything we have seen before?”

For model risk management this illustrates an important principle in SR 26-2, the new US guidance discussed below: model risk depends not only on complexity but on purpose and exposure. The same algorithm forecasting ATM withdrawals has low materiality; used to manage the regulatory liquidity buffer, it has high materiality. The model inventory should record the use, not just the algorithm.

What can AI in customer service measure, and what should it not promise?

Assistants in mobile apps and on contact centre lines are the most visible use of AI in banks today, and since large language models arrived, the most tempting. Budget requests promise relief for agents and round-the-clock service. Some of those promises can be measured; some cannot.

Intent recognition and routing (did the customer reach the right path first time), the share of conversations resolved without an agent, handling time and cost per contact are all measurable. The trap is in the second metric. The share of conversations “resolved by the assistant” also rises when customers give up. Always pair it with the repeat contact rate on the same issue within a few days, with complaints, and with post-conversation satisfaction. An assistant that closes 60% of conversations while one in three of those customers calls back has not saved what the dashboard shows.

Do not promise what the bank cannot control. An assistant built on a language model should not independently answer questions with legal or financial consequences for the customer: complaints, loan terms, debt restructuring, suspected fraud. Those intents should go to a person; the assistant can at most collect data and prepare the case. In the EU, Article 50 of the AI Act has required since 2 August 2026 that people be told they are interacting with an AI system, unless that is obvious from the circumstances.

An interesting signal is how supervisors treat generative AI. In SR 26-2 of 17 April 2026, US banking supervisors explicitly placed generative and agentic AI models outside the scope of the guidance, describing them as novel and rapidly evolving, and left banks to determine appropriate governance and controls for them. Traditional statistical and quantitative models and non-generative, non-agentic AI remain in scope. That does not mean generative AI is free of oversight. It means there is no settled method for validating it yet, and the bank has to design its own.

Here the oversight cost shifts from model validation to testing system behaviour. A system that answers slightly differently each time needs test suites, answer evaluation and production monitoring. How to organise that is covered in the article on testing LLM applications. Customer service AI has a moderate oversight cost only when the assistant’s scope is narrow and clearly bounded.

How do you build model risk management that handles the next use case without a new project?

This is the real product of the first ML project. For US banks, and as a reference point far beyond the US, the central document changed this year. On 17 April 2026 the Federal Reserve, FDIC and OCC issued SR 26-2, Revised Guidance on Model Risk Management, which supersedes SR 11-7 of April 2011, the global reference for model risk management for fifteen years, along with SR 21-8 on BSA/AML models. The guidance is expected to be most relevant to banking organisations with over USD 30 billion in total assets.

Four features of SR 26-2 are worth knowing. It is explicitly risk-based: model materiality comes from exposure (how large the portfolio or business impact) and purpose (models used to meet regulatory requirements or manage financial risk are generally higher risk). Its definition of a model excludes simple arithmetic and deterministic rule-based processes. It states that it does not set enforceable standards and that non-compliance with the guidance itself will not result in supervisory criticism, while supervisory action may still follow unsafe or unsound practices stemming from poor model risk management. And, as noted, generative and agentic AI are out of scope. For EU banks, the equivalent expectations come from the EBA guidelines, CRR requirements for internal models and national supervisors’ model risk expectations, but the structure is very similar.

From these sources you can derive a minimum set worth building with the first model:

  • A model inventory recording use, business owner, materiality, version, inputs and validation status. Without it you cannot assess aggregate risk, where several models rely on the same data or assumptions and fail together.
  • Independent validation proportionate to materiality, covering conceptual soundness (design, assumptions, data selection and, for ML, interpretability measures or benchmarking against other models), outcomes analysis against real-world results and an ongoing monitoring plan.
  • Effective challenge, meaning critical analysis by people with the expertise, independence and organisational standing to force change. The last element is the one most often missing: a validator who cannot stop a release is a reviewer, not a control.
  • Production monitoring with thresholds that send a model back for review, recalibration or redevelopment.
  • A path for vendor models. The bank often does not receive code, data or methodology, yet model risk principles still apply. It has to understand the vendor model’s design and development data, then run its own monitoring and outcomes analysis.

On the technical side, this corresponds to an MLOps platform with data and model versioning, a model registry linked to the inventory, and a record of predictions and features for every decision, so that you can reconstruct why the model scored a specific payment or application the way it did. Storing features as of the moment of decision is especially important for customer explanations and adverse action notices: reconstructing an explanation six months later from data that has since changed is close to impossible. If the platform or models are bought, EU banks also have to apply DORA’s ICT third-party rules, covered in the article on what DORA changes in choosing a financial software vendor.

Who builds it? The EBA identifies the shortage of skilled people as one of the main obstacles. A bank needs data engineers for pipelines and feature records, ML engineers for the platform and deployment, data scientists for the models, and a separate group of validators with statistical expertise. The sequence of those hires is covered in data scientist, data engineer or ML engineer: who to hire first. If ML has to fit into the bank’s wider architecture, see also where to build and where to buy in banking software.

What does the use case map look like: impact, time to measurement and oversight cost?

The table below brings together the conclusions of the previous sections. The ratings “low”, “medium” and “high” are my estimate based on the regulations described above and the nature of the decision, not the result of a study. A specific bank might move a use case by one category depending on whether the model decides on its own or supports a person.

Use caseBusiness impact metricTime to reliable measurementMain supervisory requirementsOversight costOrder
Payment fraud detectionFraud value prevented minus alert handling and customer friction costWeeks (with a label settlement window)GDPR, internal model risk management; excluded from AI Act Annex III 5(b)Low–medium1
Cash and operational liquidity forecastingForecast error vs current method, converted into buffer and logistics costWeeks–monthsModel risk management; higher materiality if used for regulatory liquidityLow–medium1–2
Digital channel assistant (narrow scope)Cost per contact at constant repeat contact and complaint ratesWeeksAI Act Art. 50 since 2 Aug 2026; GDPR; behavioural testing; outside SR 26-2 scope if generativeMedium2
ML challenger models in validationWeaknesses found in production modelsMonthsModel risk management, CRR (challenger does not calculate capital)Medium2–3
Consumer credit scoringApproval rate at constant loss rateMonths–quartersEBA LOM 4.3.4, GDPR Art. 22, AI Act high-risk from 2 Dec 2027 and Art. 86; in the US ECOA/Reg B adverse actionHigh3
ML in IRB capital modelsRisk differentiation accuracy, RWA impactQuarters–yearsCRR, model change framework, supervisory approvalVery high4

How should you read it? The “Order” column does not say which use case is most valuable. It says in which order the bank builds capabilities that the next ones depend on. Fraud detection teaches impact measurement, working with delayed labels and monitoring. Forecasting teaches materiality by use. The assistant teaches testing of non-deterministic systems. Challenger models build validators’ ML skills without capital risk. Only with that set does scoring stop being a research project.

Someone will ask whether this is too slow, since competitors are deploying ML scoring now. Perhaps, but the AI Act deadline applies to everyone equally, and a bank that reaches December 2027 with a working inventory, validation and explanations will be better placed than one with a better model and none of those.

What are the key takeaways for a bank CIO?

Machine learning in banking has documented use cases with measurable impact, but the cost of each is dominated by the model’s surroundings: validation, explanations, monitoring and regulatory compliance. That cost varies between use cases far more than the cost of the model itself, and it should decide the order.

Remember six things. First, assess use cases on two dimensions, oversight cost and time to reliable measurement, not only on promised benefit. Second, start with payment fraud detection: the problem is large, labels arrive quickly, the model supports analysts, and the EU AI Act excludes it from the credit scoring high-risk category. Third, measure impact in currency, against the current process and with a control group, and define the label settlement window. Fourth, prepare credit scoring for 2 December 2027 in the EU and for adverse action requirements in the US, but enter it only with working model risk management; design the customer explanation first, then choose the algorithm. Fifth, make your platform distinguish models that may be updated automatically from those where every version is a regulatory decision. Sixth, do not replace liquidity stress scenarios with ML models.

If you are building a team for a first ML project, or for the platform and model validation around it, see how we provide IT specialists for fintech and banking, or book a call. We add data engineers, ML engineers and validators to the bank’s own team through staff augmentation, with senior rates typically EUR 335–485 per day, a two-week specialist replacement guarantee (replacement within five business days) and a 30-day notice period.