How should ASX-listed companies control AI token costs before bill shock hits?
AI adoption is moving from enthusiasm to operating discipline.
For the past two years, most leadership conversations about generative AI have started with possibility. What can we do with ChatGPT? How do we get staff using Copilot? Which teams should experiment with Claude, Gemini or other models? Where are the best use cases?
Those questions still matter. But for CTOs at ASX-listed companies, a new question is becoming just as important: how do we stop AI usage from turning into a token bill shock?
Short answer: ASX-listed companies should control AI token costs by tracking usage by use case, matching model capability to task complexity, setting sensible caps, reporting actuals against forecast, and training teams to think about AI tokens the same way they already think about cloud consumption. The goal is not to constrain AI adoption. It is to make sure the best tokens go towards the work that creates the best outcomes.
Why are AI token costs becoming a CTO problem?
Generative AI is no longer confined to a few innovation pilots. Across ASX-listed companies, it is now appearing in coding, customer service, fraud detection, internal knowledge tools, analytics, search, workflow automation and product features.
That shift changes the economics. In a small pilot, token use can feel invisible. At enterprise scale, it becomes a run cost. Every long prompt, large document upload, agentic workflow, code-generation session, retrieval call and lengthy output consumes tokens. The more useful the tools become, the easier it is for consumption to grow quietly in the background.
The Australian Financial Review recently captured the issue well in its article on corporate AI bill shock, describing Westpac chief executive Anthony Miller using an AI token tracker to keep the cost of each interaction front of mind. The same article quotes Mindfields managing director Mohit Sharma on token transparency: “That silence is expensive.”
That is the right framing. Tokens are not crypto tokens in this context. They are units of language model usage, primarily from providers such as OpenAI and Anthropic, and they determine how much it costs to send prompts, process context, generate outputs and run increasingly capable AI workflows.
What is “AI token cost” in plain English?
An AI token is a small unit of text processed by a large language model. When a user asks a question, uploads a document, runs a coding assistant, asks an agent to research a topic or generates a long answer, the model processes input tokens and produces output tokens.
For CTOs, the important point is not the technical definition. The important point is that different tasks consume different amounts of tokens, different models charge different rates, and more advanced workflows can multiply usage quickly.
A short email edit might use very few tokens. A legal or strategy document review might use far more. An agentic workflow that searches, reads, writes, checks and rewrites can consume tokens across multiple steps. Coding tools can be especially hungry because they read files, reason through changes and generate large outputs.
This is why a simple licence view is not enough. Seat costs matter, but for many enterprise AI use cases, the variable model usage cost becomes the thing that needs active management.
Why not just use the most powerful model for everything?
Because not every task needs the most powerful model.
This is the most practical lesson in the current token-cost debate. The AFR article points to a growing market split between frontier models and cheaper alternatives, including smaller models and open-weight models. The article also notes that some open-weight models can be materially cheaper than closed models from OpenAI and Anthropic, depending on the task and hosting model.
For an ASX-listed company, the answer is not to chase the cheapest model blindly. Nor is it to default every task to the most expensive frontier model. The discipline is matching the complexity of the task to the right tool.
A board paper review, a high-risk customer workflow or a complex engineering task may justify a frontier model. A simple classification task, first-pass summary, internal draft or repeatable extraction step may not. The CTO’s job is to help the organisation make those trade-offs deliberately, not accidentally.
What are ASX-listed companies already proving with AI?
The context from recent ASX-listed company disclosures is that AI is no longer theoretical. Banking and enterprise software are furthest ahead, with some of the clearest numbers.
Commonwealth Bank has connected AI to fraud-loss reduction, customer service improvements, faster credit reviews and internal engineering use cases.
Macquarie has reported 130,000 productivity hours returned to staff in seven months from Gemini Enterprise.
Westpac has reported a 46 per cent coding-productivity gain in an AI pair-programming experiment and has extended Microsoft Copilot to employees.
Xero has taken a deliberate multi-model path, with OpenAI and Anthropic/Claude integrations and AI agents embedded into product workflows.
Telstra has reported hundreds of AI use cases deployed at scale, while also warning that software licensing, cloud and AI-provider costs can offset the benefits if not managed.
Coles, Suncorp, NIB, Wesfarmers, Woolworths and others are using AI across automation, customer service, shopping assistants and productivity use cases.
The message is clear enough: broad tool access alone underdelivers. The payoff comes from redesigning specific workflows around AI. But the more workflows move into production, the more CTOs need a proper cost model behind them.
How should CTOs control AI token costs?
The practical answer is to treat AI usage like cloud usage. Not as a one-off software purchase, but as a measurable, variable operating cost that belongs to a business owner and a use case.
That does not require a heavyweight governance theatre. It does require a few habits.
Track token usage by product, model, business unit and use case.
Forecast usage before production, then report actuals against forecast.
Set caps or alerts for early-stage use cases until the economics are proven.
Consider a model-routing approach so simple tasks do not default to expensive models.
Make high-consumption tools visible to finance, procurement and technology leaders.
Train staff to understand that prompts, large context windows and agentic workflows have a real cost.
The point is not to make employees anxious about every prompt. It is to give the business enough visibility to know whether AI spend is going towards useful work, or just disappearing into a long tail of unmeasured experimentation.
What should an AI cost dashboard show?
A useful AI cost dashboard does not need to be complicated. It should answer a few practical questions that executives can understand.
Which teams are using AI the most?
Which models are driving the highest cost?
Which use cases are consuming the most tokens?
Which workflows are showing measurable benefit against cost?
Where are token volumes growing faster than expected?
Which tasks could be moved to a smaller or cheaper model without reducing quality?
This is where CTOs can create a more mature conversation with the CEO, CFO and business sponsors. Instead of asking whether AI is good or bad, the organisation can ask a more useful question: where are we getting a return on our AI spend?
How do you move from experimentation to production without blowing the budget?
The companies getting results from AI are not simply giving people tools and hoping for the best. They are choosing specific workflows, measuring the benefit and building capability around them.
For ASX CTOs, that means moving from access to redesign. A pilot can be loose. Production needs a clearer owner, a cost assumption, a model choice, a quality check, a support path and a benefit measure.
The best approach is often staged. Start with a high-friction workflow. Map how it works today. Estimate the cost of the current process. Test what AI can change. Measure the token and platform cost. Then decide whether the workflow deserves more investment.
That gives the CTO a practical way to say yes to AI while still protecting the economics.
How Time Under Tension helps ASX-listed companies
Time Under Tension helps organisations move from AI curiosity to AI capability, and from scattered experimentation to practical implementation.
For ASX-listed companies like Officeworks, Adairs, Baby Bunting and Amotiv, that has included AI inform sessions for leadership teams, hands-on training for staff, AI Adoption Sprints to identify and prioritise use cases, AI Roadmaps, AI Accelerator support, model and platform selection, workflow redesign, and custom generative AI product development.
We are Microsoft partners and an OpenAI Services Partner, so we can help organisations get value from Microsoft Copilot, ChatGPT Enterprise and custom AI applications while making sensible decisions about model choice, token usage and enterprise-scale adoption.
If your organisation is trying to work out where AI should move from pilot to production, or how to manage token costs before they become a board-level surprise, contact Time Under Tension.
The new discipline is AI unit economics
The first wave of enterprise AI was about access. The second wave is about adoption. The next wave is about unit economics.
For CTOs, that does not mean becoming the department of no. It means helping the organisation use the right model for the right task, spend tokens where they matter, and prove that AI is creating more value than it consumes.
That is how AI moves from exciting experiment to reliable operating capability.