The immediate implication for Microsoft shops is straightforward: a fixed per-user Copilot license and a usage-metered Azure AI workload belong in the same governance conversation, even though they arrive on different invoices. Seat-based tools can conceal adoption gaps; token-based services can conceal bursty consumption. Both can produce an expanding bill before the organisation has decided which tasks are worth automating.
ABC’s reporting is based on interviews with HiBob, Deloitte, Elastic and AI researcher Jon Whittle, alongside figures cited by Australia’s Assistant Minister for Technology and the Digital Economy, Andrew Charlton. Charlton’s estimate—A$5 billion to A$8 billion spent annually on AI, with most of it flowing offshore—is an official government assessment, not a measured national account. His office also says spending could plausibly reach A$20 billion to A$40 billion a year within a decade. The scale is directional, but it captures a concrete shift: AI is becoming a recurring technology purchase rather than a contained innovation project.
Cheaper capability is not a lower monthly bill
The apparent contradiction begins with the word token. A token is a small unit of text or other model input and output that an AI service processes. API providers often charge by token volume, while tools such as Microsoft Copilot commonly use a per-user licensing model. Both are proxies for consumption, but neither is a direct measure of business value.
ABC cites Epoch AI research showing that the lowest available price for language-model performance at roughly the GPT-4 Turbo level fell from US$15 per million tokens in November 2023 to about US$0.18 per million by February 2025. The underlying Epoch data supports the broad decline. On its MMLU benchmark, GPT-4 Turbo was listed at US$15 per million tokens in November 2023, while Gemini 2.0 Flash reached a slightly higher benchmark score at US$0.18 per million in February 2025.
That figure needs a qualification often lost in vendor pricing discussions: it is a capability-matched comparison, not a promise that an enterprise can run every GPT-4 Turbo-era workload for eighteen cents per million tokens. Epoch compares models that meet a benchmark threshold. A company’s actual cost depends on the model selected, prompt and response length, context windows, retries, tool calls, retrieval data, regional hosting, rate limits, and whether a reasoning-oriented model spends additional compute before producing an answer.
In other words, the price of intelligence at a fixed benchmark can fall rapidly while the amount of intelligence an organisation buys rises even faster. The same pattern has long existed in storage, cloud compute and bandwidth. Lower per-unit cost enables more use, and more ambitious use changes the total budget.
Deloitte’s Stuart Scotis told ABC that companies are asking models to handle increasingly complex tasks. Those requests consume more tokens and may run for longer. A short summarisation prompt and an agent that searches internal documents, invokes systems, checks the result, revises its output and opens a ticket are not comparable workloads, even if both are described internally as “using AI.”
AI is joining the IT budget as a variable expense
For enterprise IT, the practical change is not merely that a new line item has appeared. It is that an unpredictable category is now sitting alongside much more predictable expenses such as endpoint refreshes, Microsoft 365 licensing, storage commitments and network contracts.
ABC reported that Commonwealth Bank’s technology spending rose from A$2.3 billion in fiscal 2025 to A$2.4 billion in its most recent fiscal year, while the bank reported roughly A$200 million in gross benefits from AI use cases in fiscal 2026, including reinvested capacity. That is a useful distinction. Gross benefits are not the same as cash savings, and reinvested capacity is not automatically a reduction in headcount or operating expense. It can still be valuable—faster service, higher throughput or redeployed staff may be the desired outcome—but finance teams should not count the same capacity twice as both productivity and cost reduction.
HiBob’s Anna Volkova told ABC that her company sees speed benefits but expects a clearer bottom-line effect in two to three years. That is candid, and probably closer to the position of many organisations than claims of immediate transformation. The current business case is often a mix of employee time saved, better decision support, pilot-stage process changes and anticipated gains that have not yet appeared in financial reporting.
The missing discipline is an owner for the whole cost chain. Procurement may manage a Copilot license count. Cloud teams may watch Azure consumption. A business unit may buy a separate AI SaaS product with a corporate card. Developers may call model APIs through applications that are difficult to allocate to a department or customer. Security may discover unsanctioned use only after sensitive data has already moved through a third-party service.
Those fragmented decisions make it easy to celebrate lower model prices while missing the aggregate spend. They also make it difficult to decide whether a workload should run through Microsoft 365 Copilot, Copilot Studio, Azure AI Foundry, a third-party hosted model, or an internally governed alternative.
Copilot licensing solves only part of the cost problem
Microsoft Copilot illustrates why “AI cost” should not be reduced to token arithmetic. A per-user product can make budgeting simpler, but it does not establish that every assigned user has adopted the feature, that the work produced is reliable, or that the employee is using it on tasks with material value.
Seat licensing is useful where an organisation wants predictable access and a known set of eligible staff. But a per-seat bill has a different failure mode: licenses accumulate while regular use stays concentrated among a minority of employees. Measuring active use, repeat use and task-level outcomes matters more than counting provisioned accounts.
Usage-based AI carries the opposite risk. Costs can remain tiny during a proof of concept, then rise abruptly when the service is integrated into a high-volume workflow or an autonomous agent repeatedly performs retrieval, reasoning and tool calls. An agent that processes a customer record once is one workload. An agent that checks several knowledge sources, rewrites a response multiple times and calls external systems is a compound workload. The number of end users can remain unchanged while consumption multiplies.
Elastic’s Jeremy Pell told ABC that some companies are responding by placing time-bound token caps on staff and routing work to different models based on task complexity. Those are sensible controls, but a crude hard cap can interrupt legitimate production work while doing little to identify inefficient design. The better control is a budget with visibility: department-level allocation, workload-level tags, alert thresholds, rate controls, and a documented escalation process for exceptions.
The cost model needs observability, not estimates from the model
Whittle’s warning is the most operationally important part of the report. He told ABC that research shows estimates of token use for the same request and technology can differ by a factor of 30. Whether that precise range applies to a particular deployment depends on the architecture and workload, but the conclusion is solid: forecast figures generated by an AI system are not a substitute for telemetry from the system that is actually running.
An enterprise needs to capture, at minimum, the model used, input and output token counts, cached versus uncached content where applicable, number of requests, latency, retries, tool calls, retrieval size, error rate and the business process being served. Without that record, IT can determine only that the bill changed—not whether the change reflects successful adoption, a defective prompt, overlarge retrieval contexts, a runaway agent loop or a user behavior problem.
This is where conventional cloud FinOps practices become relevant. AI chargeback does not need to begin with punitive internal billing. It can start with showback: make each business unit able to see what it consumes, what it receives and which workloads account for the change. The first goal is informed choices, not cost suppression.
A concise operating model should include the following:
- Every production AI workload should have a named business owner, a technical owner, an approved model route and a measurable outcome.
- Teams should set an expected cost per completed task, not merely a monthly token or license budget.
- High-volume services should use the least expensive model that meets a tested quality threshold, with escalation to more capable models for defined cases.
- Agentic workloads should impose limits on steps, retrieval scope, retries and external tool calls before they are offered widely.
- Copilot administrators should compare purchased licenses with active, repeat usage and identify roles where the tool demonstrably reduces completion time or improves output quality.
Australia’s import bill is also a control-plane issue
Charlton’s ANU Crawford School speech puts a national economic framing around the same enterprise issue. He argues that Australia can host data centres and supply energy while much of the higher-value AI stack—chips, models, software, intellectual property and customer relationships—remains owned overseas. His office explicitly cautions that its breakdown of an AI dollar is indicative, not an accounting identity, because companies disclose little about unit economics.
For individual businesses, the more immediate concern is dependency. A company that builds critical workflows around an external model provider has accepted pricing, availability, data handling, product-change and geographic-hosting risks in exchange for fast access to capability. That is often rational. But it requires an exit plan, model portability where feasible, data boundaries, a record-retention policy and a realistic view of what happens if a provider changes a model, retires a feature or revises commercial terms.
The Australian spending story is therefore less about a paradox than about a management failure waiting to happen. Lower unit prices are enabling broader adoption, longer contexts and more automated work. The organisations that benefit will be the ones that can show which AI tasks earn their cost—and stop treating a falling token price as proof that the overall bill is under control.