AI Token Costs Are in the News. Here's What It Actually Means for Small Business.

Australia's big banks now track AI token usage at CEO level, and Uber burned its annual AI budget in three months. The headlines are real — but the risk profile for a 20-person business is not Westpac's. For most SMBs the discipline isn't token counting; it's identifying real workflow pain points and matching the model tier to the task.

TL;DR

• Australia's biggest banks are now tracking AI token usage at CEO level — Westpac and CBA both have active cost controls in place
• Uber burned through its entire annual AI budget in three months with no usage policy
• For small and medium businesses, the risk is significantly lower right now — lower headcount means lower volume and more wiggle room than the headlines suggest
• The real discipline isn't token counting — it's identifying your actual workflow pain points first and applying AI specifically to those
• Not every task needs an expensive frontier model — matching model to problem is where cost efficiency lives
• Hypajump builds cost tracking and model-appropriate selection into every tool we deliver

Why This Is Suddenly Everywhere

On 30 June 2026 the AFR reported that Westpac's CEO Anthony Miller has an AI token tracker on his home screen. Commonwealth Bank has usage caps in place. The story is picking up because AI cost management is becoming a boardroom issue for large Australian organisations.

It's a real issue at enterprise scale. When you tell tens of thousands of employees to use AI freely with no policy around it, costs compound fast. Uber's CTO confirmed in April that uncontrolled AI use burned through the company's annual AI budget in three months. No policy, no visibility, significant bill.

But the coverage is landing on small business owners in a way that deserves some pushback.

The Honest Read for SMBs

Nothing is free. AI tools cost money to run, and that cost is token usage — charged by volume of input and output on every call. This is real and worth understanding.

But the risk profile for a 20-person engineering or manufacturing business is not the same as Westpac's.

Lower headcount means lower volume of AI calls. Lower volume means the cost compounds more slowly and the numbers stay manageable for longer. The scenarios making headlines — annual budgets gone in months — require the kind of scale that comes with thousands of employees using AI tools simultaneously across an organisation.

For most SMBs there is still significant room to move. The token cost conversation is getting louder, but for small business it is not the emergency the coverage implies.

What Actually Matters at SMB Scale

The more important question isn't "how do we control token costs" — it's "where does AI actually move the needle in our workflows."

Companies that get the most out of AI don't automate everything. They identify their specific pain points — the tasks that are slow, error-prone, or eating hours that could go elsewhere — and start there. A focused AI implementation on two or three real workflow problems delivers measurable ROI. A broad rollout with no clear focus delivers a bill and confusion.

Start with the pain point. Apply AI to that specifically. Measure what it saves. Then expand.

That sequencing also naturally controls costs — because you're not running AI across every function simultaneously before you've seen what it returns.

Where Model Selection Comes In

Once you've identified where AI belongs in your workflow, the next decision is which model to use.

This is where most businesses — and some AI vendors — leave money on the table.

Frontier models from Anthropic, OpenAI, and Google are genuinely powerful. For complex tasks — nuanced document analysis, confidence-banded extraction from technical drawings, multi-step reasoning — that capability is worth the cost.

But not every task is complex. Summarising a meeting, classifying an inquiry, reformatting a document, generating a standard response — these tasks don't require the most expensive model tier. There are models that handle 90% of what an SMB needs at a fraction of frontier pricing. Using a high-end model on a basic workflow is the build equivalent of hiring a specialist for admin.

The discipline is matching the model to the problem. That requires knowing what each task actually needs, which comes from building with cost visibility from the start — not retrofitting it after the bill becomes an issue.

Hypajump's Approach

We build cost tracking into every tool we deliver. Every API call is logged with cost and task type so clients have visibility over where spend is going — not just a monthly total.

We're also deliberate about model selection at the build stage. We use Anthropic's Claude as our primary AI layer and we're not switching providers to chase marginal cost savings. But we scope each task and select the appropriate model tier for what that task actually requires. No frontier model on a basic workflow.

Token costs pass through to clients as part of the retainer structure — we don't absorb them. What we do is actively manage model selection so those costs stay proportionate to the work being done, and give clients the reporting to see it clearly.

Common Queries

Should a small business be worried about AI bill shock? Not at current scale for most SMBs. The bill shock scenarios making news require enterprise-level headcounts using AI tools simultaneously without policy. For a business under 50 people, the volume isn't there to create that problem — provided you're applying AI to specific workflows rather than rolling it out broadly with no focus.

What is a token and why does it cost money? A token is the basic unit AI providers use to measure and charge for usage. Input tokens cover what you send to the AI — the length of your prompt or document. Output tokens cover what the AI sends back — the length and detail of its response. Every interaction costs tokens. More complex tasks, longer documents, and multi-step agent workflows use more tokens. Costs are charged per million tokens and vary significantly between model tiers.

How do I know if my business is using the right AI model? Ask whoever built or manages your AI tool: which model is this running on, and is the same model used for all tasks? If the answer is yes to the second question, there's likely an optimisation opportunity. Simpler tasks within most tools can run on cheaper model tiers without affecting quality. If you don't have visibility into per-task costs, that's the starting point.

What's the right way for a small business to start with AI without overcooking the costs? Identify one or two genuine workflow pain points first — tasks that are slow, repetitive, or error-prone. Scope an AI solution specifically for those. Measure what it saves before expanding. That approach naturally keeps costs contained because you're not running AI across everything simultaneously before you know what it returns.