--- title: "Fraud Detection with Machine Learning: Scoring, Agents, Review - S-PRO" url: https://s-pro.io/blog/fraud-detection-machine-learning lang: en description: "How machine learning fraud detection works in banks, where AI agents add value in real time, and which Swiss controls apply to transaction monitoring models." updated: 2026-10-31T05:00:15+00:00 source: S-PRO (https://s-pro.io) --- [Back to insights](https://s-pro.io/blog) Article # Fraud Detection with Machine Learning: Scoring, Agents, Review Author [Igor Izraylevych](https://s-pro.io/author/sproadmin) CEO and Co-founder of S-PRO Published October 31, 2026 Time to read 12 min Table of content - Key Takeaways - Table of Contents - How machine learning differs from rules-based detection - Which models do what - Real-time scoring: where the agent fits - Anomaly detection in banking - Transaction monitoring with machine learning in Switzerland - Controls to put in place before go-live - Perspective - How S-PRO helps - FAQ - Sources Share AI agents can support real-time fraud detection without disrupting the user experience if each part of the pipeline does one job: the machine-learning model scores the payment in milliseconds, the agent assembles the evidence and drafts the case note, and a human decides on blocking or customer contact. Most customers never see this work. Only the small share of payments that the score flags enters a review, and that review is faster because the analyst opens a prepared case instead of a raw alert. This guide is for CIOs, COOs, heads of compliance and heads of digital at Swiss and European banks. It explains how machine learning detects fraud, how it differs from rules, where anomaly detection fits, what Swiss rules say about transaction monitoring, and where an agent layer belongs. S-PRO Bring AI Into Fraud Review S-PRO builds AI agents for financial institutions, including solutions for AML and transaction monitoring that support human-led review. [Explore S-PRO’s AI solutions](https://s-pro.io/) ## Key Takeaways - Machine learning scores each transaction against learned patterns. Rules still handle known typologies and deterministic obligations, so mature programs run both. - An AI agent does not need to block anything. Its value is evidence gathering, summarizing and case-note drafting, with a person deciding on blocking or contacting the customer. - Anomaly detection finds behavior nobody labeled in advance. It works best as a second layer behind a supervised model. - In Switzerland, banks and securities firms must generally run a computer-assisted transaction monitoring system under the FINMA Anti-Money Laundering Ordinance. Models sit on top of that obligation, not instead of it. - Data quality, explainability and governance decide whether a model works in production more than the choice of algorithm does. ## Table of Contents - Which models do what - Anomaly detection in banking - Perspective - How S-PRO helps - FAQ - Sources ## How machine learning differs from rules-based detection A rules engine applies explicit if/then conditions: a payment above a threshold to a new beneficiary in a high-risk country triggers an alert. Rules are transparent and easy to explain to a regulator. They are also static. Someone has to write each new rule, and fraudsters change tactics faster than rule sets are updated. A machine-learning model instead learns from historical transactions and customer behavior. It produces a probability-style score for each event and can be retrained as patterns change. Context is where it adds most. A card payment abroad looks unusual in isolation, but against a customer’s history of travel it may be perfectly ordinary. The practical conclusion is not to choose one. Rules remain the right tool for obligations that must be deterministic, such as sanctions hits or hard limits. Models are the right tool for probabilistic signals that no one could write down as a rule. A bank’s design question is how the two feed one decision flow. ## Which models do what Banks rarely rely on a single algorithm. Different model families carry different roles in the same pipeline. | Model family | Typical role | What to watch | | --- | --- | --- | | Logistic regression | Transparent baseline score on labeled fraud history | Misses complex interactions between features | | Gradient-boosted trees | Main supervised scorer on tabular payment and customer data | Needs explainability tooling for analysts and audit | | Neural networks, including sequence models | Behavior over time, device and session signals | Harder to explain, more data and monitoring effort | | Graph methods | Links between accounts, beneficiaries and devices, such as mule networks | Requires joined-up data across systems | | Unsupervised anomaly detection | Flags behavior that matches no known pattern | Alert volume depends on the threshold you set | The table describes common roles, not a ranking. The right mix depends on how many confirmed fraud labels you have, how fast a decision is needed and how much explanation your analysts and supervisors expect. ## Real-time scoring: where the agent fits In a payment flow the latency budget is tight. The scoring model has to answer while the customer is still on the payment screen. That is why the model, not an agent, sits in the real-time path. An agent that reasons over several systems takes longer than a payment screen can wait, and it should not be the component that decides. The agent works next to the score: 1. **The model scores** the payment and, where a threshold is crossed, raises an alert. 2. **The agent assembles evidence**: recent transactions, device and login changes, beneficiary history, earlier alerts on the same customer, and the model’s main contributing factors. 3. **The agent drafts the case note** in the bank’s own template, so the analyst starts from a structured summary and checks it against the source data. 4. **A human decides** whether to hold the payment, contact the customer or release it. The decision and its reasoning are logged. This split keeps friction away from legitimate customers. The customer experience depends on the model’s threshold and on how quickly the analyst can act, not on an autonomous system making irreversible calls. We design this as human-in-the-loop by default, and we do not present autonomous blocking as a goal. For more on how we build AI agents for banks, see the end of this article. ### Why authentication and scam patterns shape the design The joint EBA and ECB report on payment fraud, published on 15 December 2025, puts fraud at about 0.002% of the value of transactions in 2024. It also found that users bore about 85% of credit-transfer fraud losses, mostly in scams where the payer was manipulated, and that strong customer authentication remains effective. We cite these figures as the report states them, for the European payment area as a whole, not for Switzerland. The design lesson is that fraud is rare by value, and that manipulated payers authorize the payment themselves. A model that only looks for stolen credentials will miss them. Signals such as a first-time payee, an unusual amount for the customer or a payment made during a live phone call matter more here, and a prepared case note helps the analyst decide whether to call the customer before the money leaves. ## Anomaly detection in banking Supervised models need labeled fraud, and labels are scarce and late. Anomaly detection fills the gap by asking a different question: is this behavior unlike anything seen before? Two Swiss and international examples show the approaches: - A [University of Zurich risk-management report](https://www.df.uzh.ch/dam/jcr:106d4950-949e-4434-9b15-4de803f3a587/RiskON_Report_Innovation%20in%20Risk%20Management_2024.pdf) describes an Isolation Forest model that identifies anomalous transaction behavior, particularly among relationship managers. It needs no labeled data, and compliance teams can set an alert threshold to prioritize investigations. - A [BIS working paper from 15 May 2024](https://www.bis.org/publications/working-paper-1188-finding-needle-haystack-machine-learning-framework-anomaly-detection-payment-systems) proposes layers for real-time payments: a supervised model separates typical from unusual payments, then unsupervised learning looks for anomalies among the unusual ones, including types of anomaly not known in advance. The layered idea is useful in any bank. The supervised model keeps the volume manageable, and the anomaly layer catches what the first model was never trained on. The threshold is a business decision: lower it and analysts see more alerts, raise it and you risk missing unusual cases. Set it with the team that will work the queue. ## Transaction monitoring with machine learning in Switzerland Fraud detection and transaction monitoring overlap in data and techniques but answer to different obligations. For Swiss banks and securities firms, [Article 20 of the FINMA Anti-Money Laundering Ordinance](https://www.fedlex.admin.ch/eli/cc/2015/390/de) generally requires a computer-assisted transaction monitoring system that helps identify transactions with increased risk. The same ordinance, in Article 26, requires internal directives that set out the criteria for identifying increased-risk transactions and the fundamentals of the monitoring system. Two consequences follow for model projects. - **Models extend the system, they do not replace its documentation.** A [Swiss HSLU Economic Crime Blog article from August 2025](https://hub.hslu.ch/economiccrime/transaktionsmonitoring-mittels-kuenstlicher-intelligenz/) describes AI as a way to support a rule engine through additional scoring and review. Your directives should describe how scores are used. - **Complement before you replace.** Innosuisse [reported in March 2025](https://www.innosuisse.admin.ch/en/artificial-intelligence-at-the-service-of-finance-professionals) on FINVA AML, developed in Ticino with SUPSI and QBT, which evaluates transactions in context and is intended to complement existing risk-scoring and monitoring systems while reducing false positives. That is the pattern we recommend: add a layer on top of the existing AML stack. On the national level, a preliminary study coordinated by the Swiss Bankers Association, published in April 2025, recommends further evaluating a [centralized real-time risk-score service for account-to-account payments](https://www.swissbanking.ch/_Resources/Persistent/2/7/7/4/2774d3317252c8478c63ef3d8fe63d115b96ee47/Results%20report%20of%20the%20preliminary%20study%20-%20Collaborative%20Fraud%20Prevention.pdf) using network-level machine learning, and reports that the Swiss National Bank supports the approach. It is an evaluation recommendation, not a live service, so plan your own capability in the meantime. The BIS General Manager [said in April 2024](https://www.bis.org/speeches/sp240417.pdf) that AI and machine learning can improve pattern recognition, predictive power and anomaly detection in AML monitoring, including on very large and less structured datasets, with graph neural networks as one example. The same speech treats deployment as a matter of risk management, not a replacement for governance and oversight. ## Controls to put in place before go-live A model that works in a notebook still has to survive audit, drift and a busy Monday queue. A checklist we use when we scope projects: - **Data foundation.** Research on AML in banks points to old and complex IT systems as a source of data quality and availability problems, and says the data foundation needs investment before AI succeeds (Aalto University thesis, 2021). Map the sources, owners and lineage first. - **Model risk management.** FINMA Guidance 08/2024 addresses model and data risks when institutions use AI. Read it with your model validation team and document training data, testing and ownership. - **Explainability for analysts.** Show the main factors behind each score in the case view, so an analyst can challenge it. - **Drift monitoring.** Track score distributions, alert rates and confirmed outcomes over time, and define when retraining or rollback happens. - **Privacy and hosting.** Decide where data is processed and who can see customer data in prompts and logs. Swiss data hosting is a common requirement for private banks. - **Human oversight and audit trail.** Log inputs, outputs, the agent’s draft and the human decision. - **EU AI Act check.** If you serve EU customers or operate in the EU, check Annex III of the EU AI Act against the use case before deciding how to treat it. Use the Commission’s AI Act page or EUR-Lex for current application dates. Industry analysts also stress that detection has to connect to decisioning. A [RegTech Analyst article from 2026](https://regtechanalyst.com/why-financial-crime-stacks-must-connect-detection-to-decisioning/) says fragmented financial-crime stacks with siloed data stop senior leaders from seeing risk exposure consistently. A model that raises alerts into a disconnected queue creates work without improving decisions. The agent layer is where that connection is made. ## Perspective > We see the same pattern in most fraud and transaction monitoring programs: the model is rarely the bottleneck. The bottleneck is the work around the alert, where analysts pull data from several systems, rebuild the customer’s story and write it up before they can decide anything. > > > > That is the work we build AI agents to take on. The model keeps scoring, the agent prepares the evidence and the note, and your team keeps the decision. We think this is the responsible route for a regulated bank: it speeds up the slowest step without moving accountability away from people. > > > > We also advise starting with one use case and one queue, measuring it against your own baseline, and extending only when your compliance and model-risk teams are comfortable. Our solutions are designed with EU AI Act, FINMA and DORA requirements in mind, and every agent action is logged for audit. > > > > *The S-PRO team* ## How S-PRO helps S-PRO is a Swiss international firm headquartered in Zurich, founded in 2014, with 250+ engineers and ISO 27001 and ISO 27701 certification. We have delivered 100+ AI projects, 30+ of them agentic. For fraud and transaction monitoring we work on AI agents for banks and on AML & transaction monitoring: evidence assembly, case-note drafting and human-in-the-loop review that sit on top of your existing scoring models and AML systems. We build your solution on our own agentic platform, S-PRO Agents, and run it as a service. S-PRO is a Certified partner in the Claude Partner Network, and our team includes Claude-certified engineers and architects. If you want to test one fraud or alert-handling queue, talk to us through [our website](https://s-pro.io). ## FAQ ### Can machine learning detect fraud in real time without slowing payments? Yes, when the scoring model is the only component in the payment’s critical path. It returns a score while the customer waits, and anything slower, such as evidence gathering by an agent, runs afterward on flagged payments only. The latency target is set by your payment system, so test it in your own environment. ### Is machine learning better than rules for fraud detection? They do different jobs. Rules are transparent and suit known typologies and regulatory obligations. Models learn patterns and context that nobody wrote down. Most large programs use both, with rules and model scores feeding the same decision flow. ### Does Switzerland require machine learning for transaction monitoring? No. Article 20 of the FINMA Anti-Money Laundering Ordinance generally requires banks and securities firms to operate a computer-assisted transaction monitoring system. It does not prescribe machine learning. If you add models, your internal directives under Article 26 should describe how they are used. ### What is the difference between anomaly detection and supervised fraud models? A supervised model learns from labeled examples of fraud and legitimate activity. Anomaly detection needs no labels and flags behavior that differs from the norm, so it can surface new schemes. Many banks run the supervised model first and anomaly detection on the unusual remainder. ### Should an AI agent block suspicious payments on its own? We do not recommend it. The agent should assemble evidence and draft the note, and a person should decide on blocking or contacting the customer. That keeps accountability clear and gives you an audit trail. ## Sources - [University of Zurich: Innovation in Risk Management report](https://www.df.uzh.ch/dam/jcr:106d4950-949e-4434-9b15-4de803f3a587/RiskON_Report_Innovation%20in%20Risk%20Management_2024.pdf) - [BIS working paper 1188: Finding a needle in a haystack](https://www.bis.org/publications/working-paper-1188-finding-needle-haystack-machine-learning-framework-anomaly-detection-payment-systems) - [FINMA Anti-Money Laundering Ordinance (SR 955.033.0), Fedlex](https://www.fedlex.admin.ch/eli/cc/2015/390/de) - [HSLU Economic Crime Blog: Transaktionsmonitoring mittels künstlicher Intelligenz](https://hub.hslu.ch/economiccrime/transaktionsmonitoring-mittels-kuenstlicher-intelligenz/) - [Innosuisse: Artificial intelligence at the service of finance professionals](https://www.innosuisse.admin.ch/en/artificial-intelligence-at-the-service-of-finance-professionals) - [Swiss Bankers Association: Collaborative Fraud Prevention, preliminary study results](https://www.swissbanking.ch/_Resources/Persistent/2/7/7/4/2774d3317252c8478c63ef3d8fe63d115b96ee47/Results%20report%20of%20the%20preliminary%20study%20-%20Collaborative%20Fraud%20Prevention.pdf) - [BIS: Managing AI in banking, speech of 17 April 2024](https://www.bis.org/speeches/sp240417.pdf) - [RegTech Analyst: Why financial crime stacks must connect detection to decisioning](https://regtechanalyst.com/why-financial-crime-stacks-must-connect-detection-to-decisioning/) Subscribe Get technical insights, and industry news Thank you! Get ready! You will receive handpicked content right to your inbox. --- Talk to an engineer who builds AI agents for banks. ISO 27001 and ISO 27701 certified. S-PRO is a Certified partner in the Claude Partner Network. Zurich, since 2014, 250+ engineers. Clients include AMINA Bank and Earlybird. [Book a discovery call (45 min)](https://s-pro1.pipedrive.com/scheduler/9AOA0Ix/s-pro-introductory-meeting)[or send a message](https://s-pro.io/get-in-touch) Link copied to clipboard [Get in touch](https://s-pro.io/get-in-touch)