AI in payments operations: reconciliation matching, exception prediction and governance

Artificial intelligence in payments operations is most valuable where the work is repetitive and the data is plentiful, and reconciliation is the best example. A nostro reconciliation matches thousands of statement entries to ledger entries every day, and most of them match exactly. The ones that do not match are the breaks, and the analyst must investigate each one. A machine learning model can match the entries that differ only in timing, in value date, or in charges, and can suggest the match with a confidence score. The analyst reviews the suggestion, confirms or rejects it, and the model learns from the feedback. The result is that the analyst spends time on the real breaks, not on the routine matches. Exception prediction is the next use. A model trained on past exceptions can predict which payments are likely to fail, before they fail. A payment to a new beneficiary, in a new corridor, for an unusual amount, at an unusual time, is more likely to be returned or rejected. The model scores the payment at initiation, and if the score is high, the payment is held for review, or the customer is asked to confirm the details. The prediction must be explainable, because the operations team must know why the payment was held, and the customer must be given a reason that makes sense. Fraud detection is the most mature AI use in payments, and it is also the most regulated. A fraud model scores each payment in real time, using the amount, the beneficiary, the device, the location, and the customer's history. The model must be accurate, because a false positive blocks a legitimate payment and angers the customer, and a false negative lets a fraud through and loses money. The model must also be fair, because a model that discriminates by a protected characteristic is unlawful. The requirements must state the accuracy target, the false positive rate, and the fairness test. AI governance is the discipline that makes AI safe to use. It starts with the data: the training data must be representative, must be labelled correctly, and must not contain bias. It continues with the model: the model must be documented, must be versioned, and must be tested on data it has not seen. It ends with the operation: the model must be monitored in production, must be retrained when its performance degrades, and must be explainable when it makes a decision. The EU AI Act classifies some payment AI as high risk, and requires a conformity assessment, a risk management system, and human oversight. Model risk management is the bank's internal control for AI. The model must be validated by a team that did not build it, must be documented in a model inventory, and must be reviewed annually. The validation must test the accuracy, the stability, the fairness, and the explainability. A model that cannot be validated cannot be used, however accurate it appears to be. The business analyst must write the requirements for the validation, including the test data, the metrics, and the sign-off. Human oversight is the non-negotiable requirement. An AI model must never make a final decision about a customer's money without a human being able to review it. A payment blocked by a fraud model must be reviewable by a human, and the human must be able to overturn the decision. A reconciliation match suggested by a model must be confirmed by a human before the ledger is updated. The requirements must state the human oversight for each AI decision, and the acceptance criteria must include a test where the human overturns the model and the system respects the decision. For a business analyst, AI in operations is a set of requirements with a use case, a data source, a model, a metric, a governance control, and a human oversight. The acceptance criteria must include a test where the model is evaluated on unseen data, the metrics are reported, and the human oversight is demonstrated. Last reviewed 7 November 2026. Further reading: Regulation (EU) 2024/1689, the EU AI Act; the Bank of England and PRA guidance on AI and machine learning; the Basel Committee work on model risk; the Financial Stability Board work on AI in finance.