Why Australian Banks Need to Fix Their Processes Before Deploying AI Agents

By Pankesh Upadhyay, Head of Banking, Financial Services and Insurance ANZ at Cognizant

Australian banks and insurers are increasingly discovering the more complex capabilities of AI agents, such as following a workflow, drawing information from multiple applications and recommending, or, within limits, taking, the next step. With less manual handling, faster service and more time for staff to deal with complex customer needs, the appeal is clear, but an agent is only as strong as the foundation underneath it. 

And this is where banks get stuck. 

Pankesh Upadhyay

While financial institutions have spent years trying to address technology debt, often they overlook the informal and undocumented processes which experienced employees have been using to keep things moving around legacy tech. This becomes a problem when an AI agent is fed a workflow without considering the human judgment and workarounds that are taking place every day. And ultimately, by implementing AI onto outdated processes, the organisation will simply scale the old, inefficient process in a new form.

We call this ‘people debt’.

Why ‘people debt’ can undermine AI ROI in Australian financial services

‘People debt’ is the practical knowledge employees have gained through experience, which is essential to completing work, but has never actually been clearly captured in a procedure, system or business rule. For example, a bank employee may know that a missing field is a routine documentation issue in one case, but a reason for further due diligence in another. An insurance claims specialist may recognise when an unusual detail needs review, or when an apparent delay is caused by an internal hand-off rather than the customer. This kind of context and nuance often lives in people’s brains, not in the formal workflow.

Whilst a pause to check a system, ask a colleague or confirm a detail may seem like a hindrance, it may just be the thing that stops an incorrect result reaching a customer, creating financial exposure or requiring costly rework. If we remove it without understanding its purpose, banks risk scaling the very weaknesses they were meant to eliminate.

This is particularly significant as we’re seeing 68% of banks piloting generative AI, whilst only 36% report measurable productivity gains from AI investments. It highlights the gap between AI activity and the organisational foundations and readiness needed to turn investment into results.

Map financial services workflows before deploying AI agents

To reach its full potential, AI agents need to know which source is authoritative, when an exception changes the path and who can resolve a case the rules do not cover. Before automating, banks and insurers should identify the occasions when employees have created workarounds that fall outside the core process, such as repeated hand-offs, manual spreadsheets, calls to other teams, duplicate checks and recurring exceptions. These signals should be investigated rather than automatically being removed.

Once identified, the steps employees are taking can be categorised into three groups: controls that protect customers or manage risk; workarounds that exist because systems or processes are fragmented; and activity that adds little value and can be simplified. And employees who actually do the work should help to verify the process. 

While AI can summarise approved case information and spot patters, employees should be there to help check its findings against company policies, rules and the experience of the people responsible for the process.

For example, when a claim comes in, an AI agent could gather the customer’s details and documents, identify missing information and suggest what to do next. If the records conflict, the situation is unusual or the decision could significant impact the customer, then the case should be re-routed to a claim’s professional for further checks,

Build AI governance and human oversight into operations

In Australia, APRA-regulated entities are required to manage operational risks, maintain critical operations through disruptions and manage risks associated with service providers. This reinforces that technology, data and third-party dependencies are part of the operating environment institutions must control, and of course AI agents fall into this as well.

Before an agent can act, an institution should define the data it can use, the systems and actions it is allowed to have access to, what requires human review, and the person held accountable for its outcomes. It also needs a clear stop-and-escalate path when information conflicts, inputs are incomplete or a case falls outside approved rules.

Those safeguards must sit in the operating layer, including permissions, business rules, approval steps, logging and ongoing monitoring. A controlled execution path prevents agents from roaming across systems and gives teams a record of what happened and why. In a high-impact process, an AI recommendation should not be the decision just because the workflow makes it easy to accept.

Governance also depends on culture. Staff need to feel empowered to question an agent’s output and raise a flawed process, without being blamed for the workaround they developed to keep work moving. The goal is not to document every informal habit and make it permanent, but rather to understand what the habit was compensating for, then redesign the process.

Turning institutional knowledge into operational resilience

Experienced employees are the key to making processes better, faster and more resilient with AI. They have a wealth of knowledge leadership may not be aware of, and which can help define what should be automated, what needs review and what should remain under human control.

In this way, success should also be measured in quality, positive customer outcomes and reduced errors, rather than how many cases an agent can get through or how quickly it completes a task. After all, a process done quickly doesn’t necessarily mean a process done better.

Organisations must capture the judgement behind that task, remove unnecessary workarounds and define who remains accountable. If you give an agent a broken workflow and it can reproduce the problem at machine speed. But by redesigning the work before applying AI with clear controls, that institutional knowledge can become a source of resilience, not a legacy problem carried into the next generation of systems.

Author Bio:

Pankesh Upadhyay is Head of Banking, Financial Services and Insurance ANZ at Cognizant, the leading AI Builder and technology services provider. He is a banking and finance expert and has worked at Cognizant for over 15 years.

Related Articles

LEAVE A REPLY

Please enter your comment!
Please enter your name here

- Advertisement -spot_img

Latest Articles