Token Economy and the Future of Software Costs

As AI moves from assistance to autonomous action, companies are discovering a new economic contradiction: intelligence is becoming cheaper to produce, while the cost of consuming it may become increasingly difficult to control.

For decades, companies have paid for software they could access. Now, increasingly, they are paying for the intelligence they consume, and the distinction is beginning to change how organisations think about technology expenditure.

Traditional software subscriptions provided a degree of financial predictability. Businesses purchased licences, assigned them to employees and paid a relatively stable amount each month or year, regardless of whether an application was used occasionally or throughout the working day. Artificial intelligence introduces a different economic relationship because the resources required to perform a task can vary enormously depending on its complexity.

An AI assistant answering a straightforward question might consume relatively little computing capacity, while the same system analysing thousands of documents, developing complex software or conducting research across multiple sources can require substantially more. As businesses integrate AI into everyday operations, they are encountering an uncomfortable contradiction: the more extensively and successfully they use the technology, the harder its costs may become to predict.

Behind this shift lies something more consequential than another adjustment to software pricing. It signals the emergence of what we might call the token economy, in which the consumption of machine intelligence becomes an increasingly important economic variable.

The end of predictable software costs

The first signs of this transformation are already visible across the software industry, where providers are experimenting with pricing arrangements that reflect the variable cost of delivering AI capabilities.

Cursor, the AI coding company, offers a particularly revealing example. In 2025, it changed its pricing approach, moving away from fixed allowances of model requests toward usage allowances more closely linked to underlying computing costs. The reasoning was understandable: a difficult programming task could consume an order of magnitude more tokens than a simple request, making it increasingly difficult to justify charging the same amount for both.

The transition, however, also revealed how unfamiliar these economics remained to many customers. Some users complained about unexpected charges, prompting Cursor to clarify its pricing model and offer refunds in certain circumstances. What appeared to be a commercial adjustment exposed a broader tension between customers accustomed to predictable software subscriptions and providers whose costs increasingly depend on how much work their systems perform.

Microsoft’s Copilot Studio illustrates a related approach. Alongside conventional licensing arrangements, Microsoft uses Copilot Credits to measure different forms of agent activity, with generating answers, performing actions and applying advanced reasoning consuming different amounts of credits. Certain advanced operations also incur charges linked directly to token usage. Salesforce has developed consumption-based options through Agentforce, including Flex Credits for agent actions, while continuing to offer alternative models based on conversations or user licences.

These developments do not suggest that conventional subscriptions are about to disappear. They indicate that software companies are searching for commercially sustainable ways to reconcile the predictability customers expect with the variable costs of providing increasingly sophisticated intelligence.

The paradox of cheaper intelligence

At first glance, the economics of artificial intelligence should be moving in a more reassuring direction. Competition among model developers, improvements in computing infrastructure and advances in algorithmic efficiency have helped reduce the price of many AI operations, making capabilities that were previously expensive accessible to a much wider range of organisations.

The difficulty is that lower prices can change consumption patterns just as dramatically as they change the cost of individual operations.

Consider a business processing ten million tokens each month at a price of €10 per million, resulting in a monthly expenditure of €100. If technological improvements reduce the price by 70%, bringing it down to €3 per million tokens, the company should enjoy substantial savings, provided its usage remains unchanged.

But suppose the same business begins deploying autonomous AI agents across its operations, increasing monthly consumption to one hundred million tokens. Despite the dramatic reduction in unit price, its expenditure would rise to €300.

The cost of individual intelligence operations has fallen dramatically, while the total bill has tripled.

The calculation is illustrative, but the underlying mechanism is familiar in economics. When a resource becomes more affordable, organisations often discover additional ways to use it, particularly when technological progress makes entirely new activities possible. In the case of AI, lower prices may therefore coincide with a substantial expansion in demand.

When efficiency drives greater consumption

The difference becomes especially apparent when comparing a conventional chatbot with an autonomous agent. An employee might ask a chatbot to summarise a report and receive an answer after a relatively limited exchange. An agent pursuing the same objective could retrieve supporting documents, consult databases, compare different sources, investigate inconsistencies and revise its analysis repeatedly before producing a conclusion. Each additional operation consumes resources, and the accumulated cost can become significant when thousands of agents perform similar activities across an organisation.

There is also a physical dimension to this growth in consumption. Every additional operation depends on computing infrastructure, electricity and cooling, meaning that improvements in the efficiency of individual tasks do not necessarily translate into a smaller overall environmental footprint. The economic and environmental implications of AI therefore depend on the relationship between efficiency gains and the scale of consumption they encourage.

From software licences to a token economy

Understanding this transformation requires a closer look at what organisations are actually purchasing when they consume AI services.

In generative artificial intelligence, a token is a small unit of information processed by a language model. Depending on the system, it may represent a word, part of a word or another element of information. Providers commonly distinguish between input tokens, which the model receives, and output tokens, which it generates, with prices varying according to the model, the type of processing and the computational resources involved.

Tokens have consequently become a practical means of measuring and billing certain forms of AI consumption. They are not currencies in the financial sense, nor should they be confused with the digital tokens associated with cryptocurrencies or the tokenisation of financial assets. Their economic importance comes from their ability to make computational activity measurable and commercially exchangeable.

The development has a precedent in the evolution of cloud computing. Businesses once purchased physical servers to obtain processing capacity, committing capital to infrastructure that might remain underutilised for substantial periods. Cloud services subsequently allowed them to acquire computing resources on demand, with expenditure increasingly linked to actual consumption.

Artificial intelligence extends a comparable principle to activities traditionally associated with human cognitive work. Instead of purchasing only a software application, an organisation can acquire the capacity to interpret information, generate content, analyse scenarios or support decisions, paying according to the resources or services involved.

Beyond traditional software licensing

This changes the commercial unit around which software contracts can be organised. Where traditional licensing often revolved around the number of users, AI pricing may depend on tokens, completed actions, workflows or business outcomes. Fixed subscriptions, prepaid credits, consumption allowances and outcome-based agreements will probably coexist, reflecting different customer needs and different ways of distributing financial risk.

The implications extend beyond procurement. As companies become accustomed to measuring the computational resources required for certain cognitive tasks, they may also reconsider how they compare internal and outsourced capabilities, calculate the economics of knowledge work and evaluate investments in automation.

When AI agents become economic actors

The financial implications become more significant when artificial intelligence moves beyond responding to individual requests and begins executing sequences of activities with greater autonomy.

Many AI applications still operate within workflows that people initiate and supervise. More capable agents, however, can break complex objectives into smaller tasks, select tools, interact with external services and assess intermediate results before deciding how to proceed. While these capabilities can make automation considerably more useful, they also mean that decisions affecting resource consumption may occur without direct human intervention at every step.

Consider a financial institution deploying an AI agent to assess the risk exposure of an investment portfolio. The system could retrieve market information, purchase additional data from external providers, compare historical scenarios, run simulations and prepare recommendations. Depending on its architecture, it might also invoke specialised models or collaborate with other agents, creating an extensive chain of operations from what initially appeared to be a single request.

Each calculation, external query or additional model interaction may carry a cost. Within the permissions established by its operator, the agent can therefore make operational choices that directly influence expenditure.

The limits of delegated financial authority

This introduces a dimension of automation that conventional software governance has rarely needed to address in quite the same way. If an agent can choose between models with different capabilities and prices, should it always use the most powerful one? How many times should it repeat an expensive analysis in pursuit of greater confidence, and what happens when the search for a better answer conflicts with an organisation’s financial limits?

More importantly, who assumes responsibility when an automated process consumes considerably more resources than expected?

Delegating a task to an AI agent also means delegating choices that may carry financial consequences.

These issues reach into financial authority, operational accountability and the allocation of decision-making responsibilities. Organisations adopting autonomous systems will need to specify the activities agents can perform, the resources they may access and the expenditure they are authorised to incur. Without such boundaries, an apparently successful automation initiative could introduce costs that its operators struggle to explain or control.

Who controls the price of intelligence?

The emergence of a token economy also raises questions about the distribution of economic power across the AI industry.

Many organisations develop their applications using models supplied by a relatively small number of technology companies. Application providers may control the customer relationship, develop specialised interfaces and integrate proprietary business knowledge, but they often depend on external suppliers for the intelligence that makes their products work.

That dependency can create an uncomfortable commercial imbalance. A software company may offer its customers a predictable monthly subscription while paying variable fees for every interaction with an underlying AI model. If customers substantially increase their consumption, operating costs may grow faster than revenues, eroding the profitability of a service that initially appeared commercially attractive.

Changes introduced by model providers can add further uncertainty. Adjustments to pricing, usage policies, available capabilities or technical interfaces may force application developers to redesign services or revise their own commercial arrangements, even when they have little influence over the original decision.

For enterprise customers, the implications can become more serious as AI moves deeper into operational processes. Replacing a model provider may require organisations to redesign workflows, validate alternative systems, revise security controls and retrain employees. The dependency therefore extends beyond conventional software lock-in to include the availability, reliability and affordability of externally supplied intelligence.

Governance must follow consumption

Companies have already developed financial controls for cloud computing, including usage monitoring, cost allocation, forecasting and accountability across business units. Artificial intelligence requires comparable discipline, although the variability of its consumption and the growing autonomy of agents introduce additional complications.

Two apparently similar tasks may consume very different quantities of computing resources, depending on the models involved, the information processed and the number of intermediate operations performed. Agents can also initiate activities without continuous human supervision, while alternative systems may produce comparable results at markedly different costs.

Perhaps the most difficult challenge is establishing a meaningful relationship between expenditure and business value. An expensive analysis is not necessarily useful, just as a low-cost answer is not necessarily reliable. A dashboard reporting millions of consumed tokens can provide valuable financial information without explaining whether the organisation has achieved anything worthwhile.

From consumption visibility to operational control

Effective governance therefore requires collaboration across functions that have traditionally approached technology from different perspectives. Procurement teams need visibility into pricing structures and contractual dependencies, finance departments must develop credible consumption forecasts, and technology teams need to consider model selection, system architecture and operational efficiency. Risk and compliance professionals, meanwhile, must understand how autonomous processes receive authority, how their activities are monitored and when human intervention becomes necessary.

Some organisations may establish token budgets for departments, applications or individual agents. Others might impose limits on permitted actions, require approval for expensive operations or direct routine activities toward less costly models.

A financial institution, for example, could allow an agent to conduct preliminary portfolio analysis independently while requiring explicit authorisation before purchasing expensive external datasets or initiating additional simulations. The intention would be to preserve the benefits of automation without allowing operational convenience to bypass established financial responsibilities.

These controls need not obstruct innovation. Properly designed, they can help organisations expand their use of AI with greater confidence because managers retain visibility over expenditure and understand which decisions remain under human authority.

As autonomous systems take on more operational responsibilities, AI governance and financial governance will become increasingly interconnected.

Intelligence has a price. Who defines its value?

Major technological transformations have repeatedly changed the way organisations acquire and allocate resources. Industrialisation altered the economics of physical production, while cloud computing transformed access to computing infrastructure. Artificial intelligence is now beginning to influence how companies obtain capabilities associated with cognitive work, bringing new opportunities alongside unfamiliar economic responsibilities.

The token economy represents one aspect of this transition. By making certain forms of machine intelligence measurable and billable, it allows organisations to purchase computational capabilities with greater flexibility and to allocate resources according to actual usage.

But greater precision in measuring consumption can also create a misleading sense of precision about value.

A million tokens might produce an extensive report that contributes little to a decision, while a much smaller quantity could help identify a critical financial risk, prevent an operational failure or clarify an important strategic choice. The volume of processing provides information about the resources consumed, but tells us considerably less about the relevance, reliability or usefulness of the resulting judgment.

This distinction becomes more important as organisations delegate increasingly complex work to autonomous systems. Leaders will need to assess AI through the decisions it improves, the activities it makes possible and the outcomes it helps achieve, while remaining attentive to the costs and dependencies that accompany those capabilities.

They will also need to preserve the human responsibility behind the use of these systems. An agent may conduct an analysis, select a model or initiate expenditure, but the organisation remains accountable for the objectives it establishes, the authority it delegates and the decisions it ultimately accepts.

The challenge, therefore, extends well beyond containing technology budgets. As intelligence becomes easier to acquire and consume, organisations must ensure that the availability of computational capability does not weaken the discipline of professional judgment.

The token may measure the cost of intelligence. It cannot measure the value of judgment.

The Token Economy: The Paradox of Paying Less and Spending More