AI Agents in Banking: What They Are Actually Being Used For in Australia

When banking and AI appear in the same sentence, the assumption is usually credit decisioning or fraud detection. Those are real, and they are mostly not what is being built right now.

The agents going into Australian financial services are aimed at something less interesting and more immediately useful: the operational middle, where work is repetitive, rule-bound, and currently done by people who would rather be doing something else.

The incomplete application

In lending, a large share of applications reaching assessment come back because a document is missing or a number does not reconcile. Each one is a rework loop — days lost, an assessor interrupted, a broker irritated, and a customer wondering what is happening.

An agent built for an ASX-listed banking group checks the file at submission rather than at assessment. The gap surfaces while the broker is still in the process, not after it has been queued and opened.

Worth noting how it was built: demonstrated on a proxy dataset, so no customer data was involved in development. In regulated environments that is not a nicety. It is often the difference between a project that proceeds and one that does not.

The entirely predictable phone call

A broker channel runs two queues of phone calls. Brokers ring to ask where an application is. The bank rings brokers to chase missing documents. Both conversations are almost completely scripted, and together they consume a channel team.

An agent now handles the predictable half in both directions. What reaches a person is the call that genuinely needs one.

The point is not that phone calls are bad. It is that a relationship manager whose day is consumed by status updates is not doing the part of the job the bank actually values.

The follow-up that arrives three days late

A relationship manager has a good meeting, then six more. The follow-up goes out days later, written from memory, and the CRM quietly drifts out of date.

An agent drafts the follow-up and the CRM update within minutes of the meeting ending, ready for review. The second-order effect matters more than the first: the pipeline starts reflecting reality, because keeping it current is no longer a chore anyone has to remember.

The governance queue

This one is about AI itself. Every team in a large bank wants to build something with AI, and the governance process cannot review proposals fast enough. So teams either wait months or find a way around the process — and routing around governance is the outcome a bank can least afford.

An agent performing a first-line review against the organisation's own policies turns a months-long queue into a same-day first pass. The committee then spends its time on genuinely difficult cases rather than on the straightforward majority.

What banking adds to the problem

Three constraints shape every one of these, and they are why generic advice about agents translates poorly into financial services.

Permissions are not negotiable. An agent that surfaces the wrong customer's file to the wrong banker is not an embarrassment, it is a breach. The index has to enforce existing entitlements from the first document rather than filtering afterwards.

Development often cannot touch production data. Hence proxy datasets. It makes the build slower and it is not optional.

Every output must be explainable. Which is precisely why these agents produce something a person reviews — a flag, a draft, a populated file. None of them decides. That is what makes them defensible to a regulator and, frankly, what makes them safe.

Why the unglamorous work first

There is a reasonable question here: if a bank is going to invest in AI, why start with document checking and status calls?

Because these jobs are well understood, the rules already exist, the output is checkable, and the people doing the work today can tell you exactly where the edge cases are. They ship, they get used, and they build the institutional confidence that makes harder work possible.

The credit and fraud applications will come. They will come faster in the organisations that have already learned how to put an agent into production without incident.