Practical AI in Banking: the series
This series looks at how to use AI in Australian banking and payments. In this sector, transactions clear in seconds and every change passes risk and compliance review. A wrong answer can put real money in the wrong account. We’ve written it from engineering experience. PicNet spent several years building and supporting a real-time banking platform for an Australian payments provider. That work shaped our view of where AI can help in this environment and where it has no business being.
Who it’s for
We wrote it for IT managers, heads of technology and operations leaders at banks, mutuals, payments providers and fintechs. Most of them have had an AI proposal land on their desk. They need to work out whether it’s worth funding, how it will get past their risk team, and who will support it once it goes live. If that’s your job, these posts are meant to help you ask better questions of vendors, consultants and your own team.
What it covers
We focus on areas where banking operations have real, recurring pain. Real-time platforms produce a constant stream of alerts and logs. AI can group and summarise them so an on-call engineer gets to the likely cause sooner. In reconciliation, a model can suggest matches for the exceptions that rule-based matching leaves behind, and a person confirms or rejects each one. In AML work, AI can pull together the context for an alert and draft a summary, which leaves analysts more time for judgement. For legacy systems, models are useful for reading and documenting old code and interfaces before anyone changes them. Governance runs through all of it: how to run AI so that it holds up under APRA’s prudential standards, including CPS 230 on operational risk and CPS 234 on information security.
How we approach it
People own every decision in every system we describe. The model makes a suggestion, a person decides, and the system records both, so an auditor can later see what happened and why. Where AI would remove that human checkpoint on anything that moves money or affects a customer, we’ll say it shouldn’t be used there.
We’ll also be straight about cost. That means build effort and also the less visible costs: model running costs, the work of testing outputs against real data, integration with core systems, and the ongoing support a production system needs. We’ll cover the limitations too. Models make confident mistakes. Sometimes a well-written rule or a SQL query does the job better and more cheaply, and when that’s the case, we’ll recommend it.
You won’t find market sizing or maturity curves here. Each post deals with a working system, or a design for one, along with the trade-offs we’d expect a technical leader to weigh before committing budget.
PicNet builds production AI systems for Australian organisations. If you’re working out where AI fits in your banking or payments operations, talk to us about what a first project could look like.