Banks that put AI models to work on their own data often meet the same problem. The models draft well but fail on the questions that involve money: which clients are exposed to a counterparty, whether two records describe the same company, whether a figure means the same thing in two systems. A common cause is that the bank has never written down what its data means in a form software can follow. Such a description is called an ontology, and without one a bank cannot safely let AI agents act on its behalf. Singapore’s public record shows the gap with no AI involved: the Hin Leong collapse, in which one cargo supported three financings from six banks; MAS’s 2025 finding, after the S$3 billion money-laundering case, that most of nine penalised institutions had anti-money-laundering policies but applied them poorly or inconsistently; and the shared definitions Singapore’s banks have agreed with each other since 2020 to close gaps of this kind. Recent studies find that newer AI models answer more database questions correctly but still give confident wrong answers where definitions are missing, and that AI agents working from a fictional bank’s policy documents completed about a quarter of their tasks. No published study has yet tested an ontology inside a bank. Most of the cost of building one lies in agreeing definitions, and the best public costing of related work suggests that the return comes from reuse rather than from the first project. Ten questions let a board test its own position, starting with one that costs nothing. One development would undo the case: AI able to work out a bank’s ontology from its existing systems.
Author: Dr Joey Tan