NextFin News - Westpac is not walking away from artificial intelligence. It is trying to make sure the bank does not turn a productivity tool into an uncontrolled expense line. The clearest sign of that shift is the way the company is now framing model use: staff can still use AI, but managers are being pushed to think harder about when a premium model is worth the cost and when a cheaper or simpler approach is enough.
The reason for that discipline is straightforward. Westpac has already seen that generative AI can materially lift output in the right workflow. In an internal coding experiment, the bank said 60 engineers were split into four groups, with three groups given generative AI tools from Microsoft, Amazon and OpenAI and one control group left to hand code. Westpac said the AI-assisted teams delivered a 46% productivity gain versus the control group, with no noticeable drop in code quality, and the hand-coding team took 3.5 times longer on average to finish the same tasks.
That result matters because it shows the bank is dealing with a genuine operating advantage, not a speculative promise. When AI shortens software delivery without degrading quality, the technology can be worth its keep even before the broader organisational gains are counted. But that same result also explains why Westpac is tightening the conversation around model choice and usage discipline. If the upside is measurable, the question becomes whether the bank can capture that gain without letting token consumption rise faster than the benefit.
Westpac’s own digital estate makes the question more urgent. The bank has said roughly 40% of its development runs through its internal mesh environment, where engineers build user interfaces, APIs, microservices, websites, mobile banking apps and other applications. That makes software teams a natural first target for AI deployment, because the workflows are structured and the output can be tested. It also makes them a natural place for cost leakage if staff default to expensive models for every task.
The market implication is that enterprise AI is moving out of the “try it everywhere” phase and into the “prove it pays” phase. Westpac is a useful example because it has already shown the productivity case, but it is now acting as if the business case still depends on cost control. That combination is increasingly likely to define how large companies roll out AI across the rest of 2026.
Westpac’s AI Push Is Real, but So Is the Bill
Westpac’s AI stance is best understood as a productivity strategy with guardrails. The bank has said it is using AI to improve customer service, productivity and decision-making, while also stressing oversight and responsible deployment on its AI governance page. That is not the language of a company retreating from the technology. It is the language of a company trying to institutionalise it.
The coding trial is the best evidence that the institution sees AI as more than a buzzword. Westpac chief technology officer David Walker said the bank tested the tools across a group of engineers and found a clear performance gain. The headline numbers are important because they are concrete: 60 engineers, four groups, a 46% productivity lift, and no noticeable deterioration in code quality. The bank also said the hand-coding team needed 3.5 times more time to complete the tasks. Those are the kinds of figures that can justify a larger rollout, because they connect AI directly to throughput.
But the same figures also explain why the bank must be selective. The best-case scenario is not that every employee uses the most advanced model for every job. The best-case scenario is that the bank reserves the expensive tools for tasks where the extra capability actually changes the outcome. In a bank, that might include software development, data transformation, internal analysis and complex drafting. It would make less sense for low-value, repetitive work where a cheaper model or a standard workflow gets nearly the same result.
“We found that compared to the control team that just had coded it the normal way, we've got a 46 percent productivity gain across the board in terms of generative AI tools supporting the coding of these three teams that were given these tools, which is quite amazing.”
Walker’s description is significant because it reinforces the idea that Westpac’s AI adoption is being judged on measurable output. The productivity gain is not a theoretical estimate. It is a result the bank says it observed internally. That makes the next stage of adoption less about hype and more about workflow design, governance and unit economics.
Why Cost Discipline Matters More Once AI Works
The moment AI starts to work in a meaningful way, the debate shifts. Before that point, the main risk is missing the opportunity. After that point, the main risk is using the technology too broadly, too expensively or in the wrong places. Westpac appears to be entering that second stage.
That transition is visible in the bank’s broader technology setup. Westpac said about 40% of development runs through its mesh platform, which is the environment where the bank builds much of its in-house code and customer-facing applications. That is exactly the kind of setting where AI can produce repeatable gains, because the work is structured and the output can be measured against clear standards. It is also exactly the kind of setting where management can see whether model usage is efficient.
The issue is that AI costs are not always intuitive to employees. A prompt can feel cheap at the point of use even when it draws on a larger budget behind the scenes. That is why chief executives and technology leaders are increasingly treating token consumption as a business metric rather than a technical curiosity. When they do that, the conversation changes from “How much can we use?” to “How much value are we getting per unit of use?”
For a bank, that is a healthy change. Financial institutions live with narrow operating margins, intense competition and constant pressure to defend efficiency. They are also large employers of knowledge workers, which makes them fertile ground for AI. If the bank can direct the technology at tasks that are both repetitive and valuable, it can get a double benefit: faster delivery and less wasted effort. If it cannot, AI becomes another technology cost that management has to trim or justify.
There is also a governance dimension. Westpac’s AI page signals that the bank wants oversight around how the technology is used, not just more usage. That matters because banks cannot afford untracked experimentation in sensitive workflows. A “sensible” model-use policy is therefore not simply about cost cutting. It is also about control, consistency and making sure the bank’s AI behaviour matches its risk appetite.
What Westpac’s Example Says About Enterprise AI in 2026
Westpac’s approach is likely to become the template for many large companies. The early phase of enterprise AI was about access and enthusiasm: give staff the tools and see what happens. The next phase is about allocation: decide which teams use which models, what work merits the more capable systems and who is accountable for the bill.
That is a more mature phase of adoption. It acknowledges that AI can be both a productivity engine and a cost trap. The technology can reduce development time, improve draft quality and help junior staff work beyond their usual skill set. But it can also become wasteful if organisations treat premium model access as a default rather than a choice. Westpac’s internal guidance appears aimed at preventing exactly that outcome.
The bank’s coding experiment helps explain why this matters so much in practice. A 46% productivity gain is large enough to matter operationally, especially if it scales across teams and projects. But a gain like that does not automatically justify unlimited use. The return depends on what the model costs, how often it is used and whether the output is actually valuable enough to support the expense. That is the calculation Westpac seems to be forcing its staff and managers to make.
In that sense, the bank is not taking a softer view of AI. It is taking a more realistic one. The technology can be powerful, but power alone does not create value. Value comes from choosing the right task, the right tool and the right level of spending. Westpac’s message is that the third part of that equation now matters as much as the first two.
The next phase of the story will be whether Westpac can scale AI without letting usage discipline slip. If the bank does, the productivity case strengthens and the cost debate fades into the background. If it does not, the technology’s promise will remain visible while its economics start to dominate the conversation. Either way, the lesson is clear: the cheapest prompt is not always the smartest one.
Explore more exclusive insights at nextfin.ai.
