Why the next phase of AI is about using less of it, in exactly the right places.

Every time an AI model reads a question or writes an answer, it uses tokens: small pieces of text that are counted and billed. For the past two years, success with AI has mostly been measured by how much of it is used. More users, more prompts, more agents, more tokens. The next shift in AI goes the other way. The organisations that get the most value from AI will not be the ones that use the most tokens, but the ones that need the fewest to get the same work done.

What token maxing looks like

Token maxing is rarely a deliberate strategy. It is what happens when AI is added to work without redesigning the work. Every step in a process becomes a prompt. Every document goes through the largest available model. Every routine question triggers a fresh round of reasoning, even when the answer was the same yesterday and will be the same tomorrow.

In a pilot, this hardly matters. Volumes are small, the novelty is high and the bill is easy to absorb. The problem appears when the same design moves into a core process that runs thousands of times a day. What looked like a clever assistant becomes a permanent, variable cost attached to work that never needed reasoning in the first place.

What token minimising means

Token minimising starts from a simple observation: most steps in most processes are not language work. Registering a request, checking data, routing a case, setting a deadline or sending a reminder follows fixed logic. Those steps need a workflow that runs the same way every time, not a model that interprets language.

In a token-minimised design, each type of intelligence does the work it is suited for:

  • Workflows and code handle the predictable steps. They are fast, cheap, repeatable and easy to audit.
  • Small, task-specific models handle the few moments where language matters: understanding what a customer means, classifying a request, summarising a document. They run at a fraction of the cost of the largest models.
  • Frontier models do the work where their intelligence is worth the price: designing the workflow, writing the code, building in tests, finding the cause when something fails and making the system harder to attack.

The most capable models become the builders and repairers of a process rather than the workers inside it. They are called in when a solution is created and when a test fails, not every time the process runs.

Why cheaper AI does not mean a cheaper AI bill

The most common objection is that AI keeps getting cheaper, so efficiency will take care of itself. The price of a given level of capability is indeed falling quickly. Total spending is a different story.

Three forces push consumption up faster than prices come down. Newer models reason longer before they answer, which multiplies the tokens per question. Agents take many steps to complete a single task, each with its own model call. And usage grows the moment something becomes easier, as it always does with useful technology. The same pattern played out with cloud computing: the price per unit fell for years, while total cloud spending kept rising.

Cheaper per answer, in other words, does not mean cheaper in total. The organisations that ignore design will see their AI bill grow with every successful use case.

Not every level of AI works the same way

Token minimising is not a rule for all AI use. It applies differently at three levels:

  • Personal AI, one person working with an assistant to think, write or prepare, benefits from the most capable models. The work is varied, the value lies in the reasoning and the cost usually sits in a fixed licence. Restricting tokens here mostly restricts thinking.
  • Team AI, shared agents and workflows within a team, is where the shift begins. As steps start to repeat, they can be turned into reliable building blocks.
  • Enterprise AI, AI built into core processes that run at scale, is where token minimising stops being an optimisation and becomes the architecture.

The critical moment is the transition between levels. A clever personal prompt becomes a team habit, and a team habit becomes a process the organisation depends on. That is exactly when the design needs to change, and exactly when it most often does not.

What token minimising is not

Token minimising is not an argument against powerful models, and it is not a cost-cutting exercise in disguise. It is an argument for placing intelligence deliberately. Organisations that respond to rising AI costs by limiting access across the board will hold back the very work where AI adds most value. The goal is not less AI. The goal is AI where reasoning is needed, and reliable automation everywhere else.

Where this is heading

Within the next two years, cost per completed task is likely to replace usage as the headline measure of AI success in large organisations. The mechanism is already visible. As agents move from pilots into processes at scale, AI spending shifts from fixed licences to consumption-based billing. The first budget cycles in which that consumption is fully visible will turn AI from an innovation topic into a line item that finance teams scrutinise. At that point, boards will start asking why the most expensive models are handling the most routine work.

Three developments are likely to follow. Organisations will introduce token budgets per process, much like they set cloud budgets today. Procurement will start asking vendors for the cost per completed task, not the price per token. And “workflow first, agent second” will become a default design principle, in the same way “cloud first” did a decade ago.

The organisations that design for token minimising now will be able to scale AI predictably. The others will face a growing bill, and a redesign under pressure.

Questions for every AI initiative

  • Which steps in this process genuinely need language or judgment, and which only need to happen reliably?
  • Is this personal, team or enterprise AI, and is it designed for that level?
  • Do we know what one completed task costs, and how that changes if usage doubles?
  • Where is the largest model doing work that a workflow, a rule or a smaller model could do just as well?

The most valuable AI design is not the one with the most AI in it. It is the one where AI sits exactly where the thinking happens.