During the first years of adoption, most companies used AI tools on an occasional, experimental basis. But when the technology starts to be integrated into real processes, corporate assistants or internal automation, behind every interaction there is consumption directly linked to the organisation's operational capacity. And the main providers structure their costs according to the volume of information processed and generated, so token consumption starts to behave like a business metric related to productivity, processes and efficiency.
Why are tokens the new unit of work for AI?
Every query to an assistant, every document analysed, every answer generated and every automation executed involves a specific amount of consumption. The more context the model needs, the more complex the task or the more autonomy the agents have, the greater the volume of tokens. The consultancy Deloitte describes the token as the new currency of the AI economy, the unit that translates infrastructure decisions into tangible economic terms, and notes that many companies already generate more than 10 billion tokens a month, with the proportion exceeding 100 billion expected to triple between 2025 and 2028 (Deloitte, AI token economics for CFOs). To give a sense of the scale of the phenomenon, Deloitte states that Google already processes around 1.3 trillion tokens a month, some 130 times more than a year earlier (Deloitte, Navigate the economics of AI).
Until recently many companies barely looked at this metric because usage was limited. That changes when AI stops being an isolated tool and becomes part of the whole organisation's daily operations. At that point, token consumption starts to resemble that of any other critical infrastructure: just as electricity consumption, cloud usage or server capacity are monitored, AI introduces a new variable linked to digital work capacity.
The figures show how far it can grow. A few weeks ago, the OpenClaw project made public a consumption of more than 603 billion tokens in just 30 days using autonomous agents for software development, at an associated cost of around 1.3 million dollars (Tom's Hardware). A caveat is worth adding: the bill was covered by OpenAI, where the project lead works, and the figure reflects the most expensive execution mode; without it, the gross cost would have been around 300,000 dollars. Even though it is an extreme scenario, it illustrates a reality that is starting to spread: the more autonomous and contextualised AI is, the greater the associated consumption.
Why is AI starting to behave like industrial infrastructure?
One of the most interesting changes is that companies will start to link AI consumption directly to their productive activity. An industrial company will be able to associate tokens consumed with production orders processed, technical analyses or validated documentation. A law firm, with contracts reviewed or case files analysed. A sales team, with complex proposals or customer analyses.
In other words, AI becomes measurable operational infrastructure, and that changes the way it is managed. It stops being just software and starts to behave like a productive resource that needs monitoring, optimisation and consumption control. Various industry analyses warn that the growth of autonomous systems is driving up computational consumption and forcing new governance and financial control practices specific to AI (ITPro).
What is the real risk: using AI or using it without control?
As assistants multiply within companies, many discover a phenomenon similar to shadow IT, but applied to artificial intelligence (shadow AI). Each department uses different tools, employees bring in platforms on their own initiative and automations appear that consume resources constantly without an overall view of their real impact.
This creates two problems at once. Costs can grow rapidly without clear control mechanisms, and the organisation loses visibility over which processes consume the most, which teams really use AI and what operational return lies behind it. Hence one of the next major business needs: governing consumption. Because the question will not just be how much it costs to use AI, but what real value each unit of consumption generates. This is not a theoretical concern: Gartner analysts are already warning that the cost of AI-assisted development could exceed developers' salaries by 2028 if consumption is not optimised (ITPro on Gartner).
Why do not all tokens generate the same value?
One of the most common mistakes is measuring only the volume of consumption without relating it to the impact it produces. Spending tokens on repetitive, low-value queries is not the same as using them to speed up critical processes, reduce analysis times or amplify the capacity of specialist teams.
That is why many organisations will need models that relate AI use to real business indicators: which departments generate the most value per token, which assistants deliver productivity improvements, which automations justify their cost and which processes would be better run on lighter, private or specialised models. The conversation will stop revolving around which model is most powerful and focus instead on which architecture is most efficient for each need. Along these lines, a recent academic study proposes standardised economic metrics to measure the real cost of deploying AI, incorporating both infrastructure and ongoing operational consumption (Curcio, LCOAI, arXiv 2025).
Why will AI efficiency be a strategic advantage?
In the coming years, many companies will find that not all tasks require the most advanced or expensive models. There will be processes that need great reasoning capacity and hundreds of tasks that can be run on lighter or private models at a much lower cost. Smart management of consumption thus becomes a competitive advantage: the organisations that make the best use of AI will not be those that adopt the most technology, but those that best balance power, cost, privacy and efficiency.
This is where architectures capable of dynamically routing each task to the most suitable model and monitoring consumption centrally become important, such as our AI Model Hub, or the use of private AI models for the workloads that justify it. The industry itself is moving in that direction: the cost of intensive AI use is driving interest in hybrid infrastructure and in systems that optimise token use (ITPro).
Why will companies have to think like technology operators?
Many organisations still see AI as just another software tool. But it is starting to behave like a cross-cutting technological infrastructure that affects the whole company, and that brings new responsibilities: controlling consumption, managing costs, defining usage policies, overseeing security, setting priorities and deciding which models to use for each task. Just as no company today leaves its energy spending or cloud infrastructure uncontrolled, it will soon no longer be able to afford to use AI without clear governance mechanisms. The difference is that productive capacity will now also depend on how much AI it is able to consume, manage and optimise sustainably. This is precisely the domain of AI consultancy and intelligent automation.
The future of business AI will also be measured in operational efficiency
Artificial intelligence will not only transform how companies work: it will also change how they measure their productive capacity. Tokens will end up being an operational unit linked to processes, productivity and value creation, and companies will have to manage that consumption like any other strategic resource. The challenge will no longer be adopting AI, but doing so in a way that is efficient, controlled and aligned with the real production, knowledge and growth capacity of each organisation.
Co-founder and CTO of LurNova, with more than 20 years' experience in developing digital technologies. Specialised in designing secure, private and governed artificial intelligence architectures and solutions for organisations.
