{"id":182021,"date":"2026-06-29T18:14:29","date_gmt":"2026-06-29T16:14:29","guid":{"rendered":"https:\/\/www.startupbusiness.it\/ai-in-business-the-allocation-of-intelligence-not-its-consumption\/182021\/"},"modified":"2026-07-03T09:49:17","modified_gmt":"2026-07-03T07:49:17","slug":"ai-in-business-the-allocation-of-intelligence-not-its-consumption","status":"publish","type":"post","link":"https:\/\/www.startupbusiness.it\/en\/ai-in-business-the-allocation-of-intelligence-not-its-consumption\/182021\/","title":{"rendered":"AI in business: the allocation of intelligence, not its consumption"},"content":{"rendered":"\n

For much of the generative AI revolution, organisations have operated on the basis of a seemingly logical assumption: if AI boosts productivity, then greater use of AI should automatically generate more business value. This belief drove the early phase of enterprise adoption of large language models, turning token consumption into a sort of universal metric for innovation. Usage dashboards, the number of prompts, API calls and AI agent activity quickly became indicators of digital progress, whilst executives and teams celebrated the growth in adoption as tangible evidence of the transformation underway. <\/p>\n\n

As time went on, however, a paradox emerged that became increasingly difficult to ignore. Many organisations were consuming unprecedented amounts of artificial intelligence, yet were struggling to demonstrate commensurate financial returns. Budgets were growing faster than measurable benefits, inference costs (the processing of data by the model) were rising, and concerns about governance were beginning to multiply. The problem was not a lack of adoption. On the contrary, AI was now ubiquitous. What was missing was the ability to link the consumption of AI to the creation of economic value. <\/p>\n\n

This tension is fuelling one of the most significant managerial transformations of 2026. The conversation is gradually shifting from what we might call \u2018Tokenmaxxing\u2019 \u2013 that is, the pursuit of the maximum possible use of AI \u2013 towards \u2018ROImaxxing\u2019, the disciplined pursuit of measurable economic results. Behind this shift lies an even deeper realisation: artificial intelligence is becoming a new corporate resource to be allocated, managed and optimised in the same way as financial capital, technological infrastructure and human talent (News18, 2026). <\/p>\n\n

The belief that drove the first wave of enterprise AI<\/h2>\n\n

The first phase of AI adoption in businesses was driven by a relatively simple line of reasoning. If artificial intelligence makes employees more productive, then greater use of AI should automatically lead to better results. In this context, token consumption became a key indicator because it was easily measurable. Whilst the actual impact on the business took months or years to assess, usage metrics were available in real time. <\/p>\n\n

Managers were able to monitor in real time the number of prompts generated, the API calls made, the tokens consumed and the activity levels of the AI agents. It was, however, far more difficult to measure improvements in decision-making quality, customer satisfaction, operational efficiency or revenue growth. As a result, many organisations ended up optimising what they could actually observe. <\/p>\n\n

The scale of the phenomenon became apparent during 2026. According to Insight (2026), Google now processes over 3.2 quadrillion AI tokens per month, a volume approximately seven times higher than in the previous year. At the same time, Meta internally monitors around 60 trillion tokens generated by over 85,000 employees via a platform known as Claudeonomics (Top1Markets, 2026). These figures demonstrate just how deeply artificial intelligence is now integrated into business processes. However, they also show how consumption has become a metric of success in its own right, regardless of the results generated. <\/p>\n\n

The key point is that consumption is an input, not an output. As economic theory points out, increasing the consumption of a resource does not automatically guarantee a proportional increase in the value created. Buying more servers does not necessarily result in better software. Similarly, using more artificial intelligence does not guarantee better decisions, more efficient processes or sustainable competitive advantages. <\/p>\n\n

When consumption grows faster than value<\/h2>\n\n

The fundamental weakness of Tokenmaxxing becomes apparent the moment organisations start weighing up the costs and benefits. As AI implementations have expanded, many companies have found that the cost curve has been rising much faster than the results curve. <\/p>\n\n

Uber is one of the most frequently cited examples. According to India Today (2026), the company had reportedly used up its annual budget for Claude Code by April, after just four months of operation. Monthly expenditure per developer ranged from around 150 to 2,000 dollars, depending on usage intensity. At the same time, some providers have increased the prices associated with token consumption, calling into question the assumption that inference costs would decrease indefinitely over time (Weixin, 2026). <\/p>\n\n

Even more significant are the findings from productivity analyses. Research conducted by Jellyfish on around 12,000 developers from 200 organisations found that the heaviest users consumed around ten times as many tokens as median users, but produced only twice as much measurable output (Jellyfish, 2026; Ghaffary, 2026). In economic terms, this means that productivity was rising, but consumption was rising much more rapidly. <\/p>\n\n

This observation is crucial because it challenges one of the central assumptions of the early stages of AI adoption. If consumption increases tenfold and output only doubles, overall efficiency deteriorates. The organisation is not maximising the value created; it is simply increasing its consumption of computational resources. The lesson is not that artificial intelligence does not work. The lesson is that consumption and value are not synonymous. <\/p>\n\n

Because the problem wasn\u2019t the budget<\/h2>\n\n

Many observers have interpreted these difficulties as a budgeting problem. In reality, the budget was merely a symptom of a deeper issue. <\/p>\n\n

Most organisations have built their spending forecasts using models inherited from traditional enterprise software. Software licences tend to be relatively predictable. AI consumption, on the other hand, is highly variable. A small number of users or autonomous agents in recursive loops can account for a disproportionate share of total expenditure, making it extremely difficult to apply traditional planning models. <\/p>\n\n

The incentive system has further complicated the situation. Many companies have rewarded usage levels rather than the value created. Internal dashboards, usage rankings and adoption targets have often encouraged employees to use as much AI as possible, without considering whether that usage was producing economically significant results. <\/p>\n\n

As is often the case in complex systems, individuals have optimised the metric used to assess them (Goodhart\u2019s Law). The result was predictable: increased activity, higher consumption, rising costs and results that often fell short of expectations. In other words, the budgeting model failed because it had been built on top of an incomplete performance model. The implicit equation of Tokenmaxxing \u2014 more tokens equals more value \u2014 proved to be far less robust than many had imagined. <\/p>\n\n

The emergence of ROImaxxing<\/h2>\n\n

During 2026, numerous organisations began to change course. According to News18 (2026), companies such as Uber, Microsoft, Salesforce, Meta and DoorDash introduced measures to limit or better regulate the use of AI. This shift reflects a growing focus on economic efficiency rather than simply on technological adoption. <\/p>\n\n

Amazon, for example, is reported to have scrapped its internal AI usage rankings after observing opportunistic behaviour on the part of employees, who were generating large volumes of low-value work solely to improve their position in the rankings (Sohu, 2026). Microsoft is reported to have reduced some of its Claude Code licences in an attempt to reassess the efficiency of its spending (Warren, 2026). <\/p>\n\n

GitHub, for its part, has introduced pricing models based on actual Copilot usage through the introduction of dedicated AI credits, thereby increasing financial transparency and user accountability (Rodriguez, 2026). Meta is even reported to have scrapped certain internal rankings based on token consumption, implicitly acknowledging that usage is not necessarily a measure of value (Top1Markets, 2026). <\/p>\n\n

These examples point to a wider shift in managerial philosophy. Organisations are beginning to treat intelligence as an economic resource rather than an unlimited commodity. The question is no longer how much AI employees use. The question is what economic value that use generates. As Yamini Rangan, CEO of HubSpot, observes, companies that focus exclusively on maximising consumption risk maximising waste rather than value (SmarterX, 2026). <\/p>\n\n

The allocation of intelligence as a new managerial discipline<\/h2>\n\n

The significance of this change extends far beyond the budgets allocated to AI. Historically, organisations have learnt to manage financial capital, physical infrastructure and human talent. Today, they face a new challenge: managing intelligence itself. <\/p>\n\n

Artificial intelligence is rapidly becoming an allocable resource. Different models offer different costs, different speeds, different levels of reasoning and different risk profiles. The strategic question is no longer whether to adopt AI, but how to deploy intelligence within workflows, decision-making processes and operational activities to maximise the value created. <\/p>\n\n

This implies that the competitive advantage of the future will not necessarily lie with the organisations that use the most AI. It will lie with the organisations that are able to allocate intelligence more efficiently than their competitors. <\/p>\n\n

The distinction may seem subtle, but it is profoundly transformative. It marks the shift from a mindset focused on technology adoption to one focused on resource allocation. Just as mature companies do not simply maximise capital expenditure but optimise its use, the AI leaders of the future will not maximise token consumption. They will optimise the economic return obtained for every unit of intelligence used. <\/p>\n\n

The strategic question that managers will need to ask themselves in the early years will therefore not be:<\/p>\n\n

“How much AI are we using?”<\/p>\n\n

The question will be:<\/p>\n\n

“How much economic value are we generating for every unit of intelligence consumed?”<\/p>\n\n

This is the question that marks the beginning of a more mature phase for enterprise artificial intelligence. A phase characterised not only by technological enthusiasm, but also by economic discipline, accountability, governance and measurable results. In other words, the era of \u2018Tokenmaxxing\u2019 is gradually giving way to the era of \u2018ROImaxxing\u2019, in which the true competitive advantage will not stem from the quantity of intelligence consumed, but from the ability to allocate it where it generates the greatest economic return. (Photo by Brecht Corbeel<\/a> on Unsplash<\/a>) <\/p>\n\n

References<\/em><\/h2>\n\n