Westpac Tightens Control Over AI Spending and Shifts Routine Tasks to Cheaper Models

Australian bank Westpac Banking Corp has tightened control over spending on artificial intelligence: the company tracks AI token consumption by employees across the organization and directs simple tasks to cheaper models, Bloomberg reports.

What Exactly Westpac Does

The bank has implemented stricter monitoring of how employees use generative AI models in their daily work. This is about tracking token consumption — units in which the volume of text processed by a language model is measured and, accordingly, the cost of a request. Tasks that do not require the capabilities of flagship models are redirected by the bank to cheaper and lighter models, leaving expensive ones for complex scenarios.

• Westpac Banking Corp — one of Australia's largest banks
• The bank tracks AI token consumption by employees across the company
• Simple tasks are transferred to cheaper models instead of flagship models

Why Companies Have Started Counting Tokens

The cost of using large language models grows along with the scale of their deployment: the more employees connected to AI tools and the more frequently they use them, the higher the bill from the model provider. For banks and other large corporations where the number of AI service users is measured in thousands, these expenses become a significant budget item. Hence the growing interest in so-called FinOps for AI — the practice of distributing load between different models on the principle of "complex task — expensive model, routine — cheap one," instead of routing all queries through the most powerful and expensive system.

Westpac's approach is typical of large financial organizations that simultaneously want to take advantage of generative AI and avoid uncontrolled growth in cloud AI service expenses, especially in a regulated industry where every expense item passes through additional scrutiny.

What It Means

As generative AI becomes part of daily operations at large companies, managing expenses on models becomes a separate discipline: business learns not just to implement AI, but to count how much each request costs, and optimize it just like any other operating expense.