Every major wave of business technology arrives wrapped in a promise that initially feels straightforward. Computers were meant to make offices more productive, enterprise systems to integrate organisations and cloud computing to convert fixed IT costs into flexible spending.
Artificial intelligence now carries its own version of that promise: faster work, lower costs and fewer routine tasks for people to perform.
The assumption that usage equates to value creation does not always hold true
Yet the pattern that follows is equally familiar. Technologies are adopted because their strategic logic appears compelling, but the real challenge emerges later, when organisations attempt to demonstrate whether the promised value has actually materialised.
Usage does not mean value
It is in this context that a curious new term has entered the conversation: ‘tokenmaxxing’. The word refers to the practice of maximising the use of AI tools, often treating the volume of tokens consumed as evidence that employees are adopting AI successfully. A token is simply a small unit of text processed by an AI model; every instruction consumes tokens, and every response generates more.
In some organisations, heavy usage has been tracked, encouraged and even celebrated as a sign of transformation. At first glance, this behaviour is entirely understandable. If a company has invested in AI tools, it naturally wants people to use them. However, this logic rests on a critical assumption that usage equates to value creation, which does not always hold true.
Cost implications
In a simple chatbot exchange, the cost may be negligible. However, once AI becomes embedded in everyday workflows, the picture changes significantly. What appears to be a single, straightforward request can trigger a cascade of activity, including database queries, model interactions, validation checks and iterative revisions.
Consider a typical example. An employee might ask an AI agent to investigate a customer issue, but the system may then retrieve account records, analyse previous correspondence, compare policies, draft a response, review it for risk and revise it before presenting an answer.
Finance teams need a more structured approach to evaluating AI
What looks like a single interaction is, in reality, a chain of computational processes, each contributing to the overall cost. Viewed in this way, tokenmaxxing begins to look less like a sign of progress and more like a warning; organisations may be measuring AI activity before they have developed a clear understanding of how to measure AI value.
Research from Deloitte highlights the dangers. In one large healthcare enterprise, token usage growth of 8%–10% per month (one trillion tokens over six months) translated into more than US$6m in annualised, previously unplanned cost increases before the finance team had visibility into the driver behind it.
Earlier this year Uber reported it had spent its entire annual AI budget in just four months due to massive token consumption on tools such as Anthropic’s Claude Code, with one executive running up a US$1,200 bill during a single two-hour coding session. To control costs, the company has now capped employee usage at US$1,500 per month.
In response to these challenges, finance teams need a more structured approach to evaluating AI, one that connects cost to outcomes and risk rather than focusing solely on usage.
A useful starting point is to ask five interrelated questions:
Where is AI being used? This requires mapping AI tools to specific processes rather than treating them as general-purpose technology. Without this level of clarity, it becomes difficult to understand what work is being affected or to compare performance before and after adoption.
What is the full cost of AI? Visible expenses, such as subscription fees, token charges, API usage and licences, are only part of the picture. AI systems also rely on integration work, cloud infrastructure, data preparation, cybersecurity controls, legal oversight, model monitoring, staff training and ongoing human supervision. Outputs must be reviewed, errors corrected and exceptions managed, meaning that the true cost of AI extends well beyond the technology itself.
What outcome is being improved? This question shifts the focus from activity to performance by examining whether AI is reducing costs per task, improving accuracy, shortening cycle times or lowering risk. Metrics such as cost per customer query resolved or cost per invoice processed accurately provide a far more meaningful basis for evaluation than raw token consumption.
A team using fewer tokens but achieving better results may be creating more value
What human knowledge is being displaced? When AI reduces or replaces human involvement, it may also remove tacit knowledge that is not captured in formal systems. Experienced employees often understand exceptions, edge cases and practical constraints that are difficult to encode, and their absence can weaken an organisation’s ability to detect when automated systems are failing.
What happens under different future conditions? AI costs can fluctuate depending on usage, model selection and provider pricing, while systems may become harder to replace as they are embedded more deeply into workflows. Organisations therefore need to consider how costs and risks will evolve over time.
Taken together, these questions encourage a shift towards evaluating AI in terms of cost per useful outcome rather than cost per token. This perspective highlights that a team using fewer tokens but achieving better results may be creating more value than one generating large volumes of activity without improving performance.
The accounting challenge is a familiar one: distinguishing between activity and value
It also allows organisations to recognise when a seemingly expensive system is justified because it reduces errors, accelerates processes or enhances risk detection. Without this level of visibility, businesses risk managing AI through averages. The overall bill may increase, but it becomes difficult to identify where value is being created and where resources are being wasted.
Rigorous approach
In this sense, tokenmaxxing is more than just a buzzword; it is an early signal that some organisations may be confusing consumption with progress. For accountants, this presents an opportunity to bring greater discipline to the conversation by linking cost, performance and risk in a more rigorous way.
The question is no longer whether AI is being used; the more pressing issue is whether organisations can account for what it costs, what it changes and what value it ultimately delivers.
In that respect, AI may be a new technology, but the accounting challenge it presents is a familiar one: distinguishing between activity and value, and ensuring that the business case is grounded in outcomes rather than in the size of the bill.
More information
See ACCA’s AI guidance and further resources
Read Ian Guider’s Comment on how AI will affect careers
Deloitte has a guide on CFOs and AI token economics
