The AI metric that matters isn't token cost

The cheapest AI programme in the business can still be a waste of money.

As AI spending rises, it is understandable that leaders are watching model, compute and inference costs closely. But token consumption is an input, not a verdict. Reducing it may improve a budget line; it does not, by itself, improve the business. The more useful question is 'what changed because AI was introduced?'

A workflow may consume substantial compute and still represent excellent economics if it reduces fraud, resolves customer issues first time, shortens a critical decision or unlocks meaningful revenue. Equally, a low-cost assistant that makes one isolated task a little faster – but leaves the customer journey, margin or cycle time unchanged – is not a compelling investment.

That is the distinction many organisations still miss. AI does not have an innovation problem. It has an execution problem.

Too often, enterprise AI is assessed through the numbers that are easiest to collect: licences purchased, users activated, agents launched or tokens consumed. Those are adoption metrics. They tell us that AI is present. They do not tell us whether the organisation works differently or performs better. The real unit of AI economics is not the token. It is the outcome of a redesigned workflow.

Before asking how cheaply AI can run an existing process, leaders should ask whether the process deserves to survive in its current form. If a customer-service team still has to search across five systems, wait for approvals and re-enter information, adding an AI layer may simply make an outdated workflow run faster in circles.

Australia's execution gap

Australia illustrates the gap between AI adoption and business transformation. Publicis Sapient's 2026 research found that 53% of Australian enterprise decision-makers say AI is highly or fully embedded in core workflows. Yet only 38% say AI is fundamentally changing how their business operates. AI is entering the workflow, in other words, faster than organisations are redesigning the systems, decisions and operating models around it.

The obstacle is rarely access to models. It is the enterprise underneath them: fragmented data, legacy systems, functional hand-offs and unclear accountability. In those conditions, AI can increase activity without reducing effort. It can search repeatedly for information the business already has, generate work that people must check, and add cost before it creates value.

Many organisations are running 21st-century technology inside 20th-century operating models. That is the execution gap: teams have access to increasingly capable tools, but the systems, processes and decision rights around them were built for a slower, more fragmented business.

For CIOs and technology leaders, this means integration and operating-model design are not technical clean-up tasks to address after an AI pilot. They are part of the business case. The question is not simply whether a use case works. It is whether the organisation can absorb it, govern it and turn it into a better decision or customer outcome.

Governance as performance

Governance matters for the same reason. Done well, it is not a brake on performance; it is how an organisation determines where AI can act autonomously, where human judgement adds value, how exceptions are handled and who owns the result.

The right model is not always the most capable, nor the least expensive. It is the model, workflow and control structure that delivers the required outcome reliably, at an acceptable cost and with appropriate safeguards around data, decisions and risk. Australian businesses should absolutely manage token spend. But they should not mistake efficiency in consumption for value creation.

The companies that pull ahead will be those that can show, workflow by workflow, what AI has changed: faster decisions, fewer failures, better customer outcomes, stronger margins or new sources of growth. That is a harder standard than counting tokens. It is also the one that matters.