Analysis: Rising AI bills reveals the flaw in thinking that businesses could replace expensive employees with almost limitless digital intelligence<\/p>\n
For the past few years, businesses have been encouraged to imagine AI<\/a> as a remarkably cheap worker. It does not sleep, take holidays, ask for a pay rise or leave for a competitor. It can write software, answer customer queries, analyse documents and complete administrative tasks at extraordinary speed. For companies under pressure to reduce costs, the attraction is obvious: remove some of the people<\/a>, automate the work and allow the savings to flow through.<\/p>\n But something unexpected is now happening inside many of the companies that embraced this logic most enthusiastically: their AI bills are rising. For example, Uber reportedly exhausted<\/a> its entire 2026 budget for AI coding tools in just four months after employees rapidly increased their use of the technology. Similarly, according to Deloitte<\/a>, one large healthcare organisation saw token usage grow by 8 to 10% each month. Within six months, this reached roughly one trillion tokens and more than $6 million in annualised costs that had not been anticipated.<\/p>\n We need your consent to load this rte-player contentWe use rte-player to manage extra content that can set cookies on your device and collect data about your activity. Please review their details and accept them to load the content.Manage Preferences<\/a><\/p>\n From RT\u00c9 Radio 1’s Today with David McCullagh, what is tokenmaxxing and why is it causing a problem for businesses?<\/p>\n As a result, businesses have now begun imposing limits, shifting routine work onto cheaper systems and questioning whether the amount of AI being consumed bears any relationship to the value being created. The trend even has a suitably absurd name, namely ‘tokenmaxxing<\/a>‘. The term refers to maximising the use of AI tools, often based on the assumption that more AI consumption must mean more productivity. Some companies have tracked token usage through internal dashboards, encouraged employees to increase it and treated heavy use as evidence that workers were embracing the future.<\/p>\n However, tokens are not productivity and are simply a unit through which AI consumption is measured and billed. A token is a small piece of text processed by an AI model. Every instruction consumes tokens, as does every response. In a simple chatbot exchange, the cost may be negligible.<\/p>\n But businesses are no longer using AI only for occasional queries. They are embedding it into workflows, connecting it to company data and asking it to complete multi-step tasks. An AI agent investigating a customer complaint might search databases, retrieve records, compare policies, draft responses and revise them. What appears to be a single request can involve dozens of interactions behind the scenes.<\/p>\n We need your consent to load this rte-player contentWe use rte-player to manage extra content that can set cookies on your device and collect data about your activity. Please review their details and accept them to load the content.Manage Preferences<\/a><\/p>\n From RT\u00c9 Radio 1’s The Business, what are the jobs that AI can’t do?<\/p>\n Each step consumes tokens. The system is not just answering a question, it is performing a chain of computational work, with the meter running throughout. This is why the AI cost problem is more complex than rising prices from providers like OpenAI<\/a> or Google<\/a>. In fact, the cost per token has generally been falling. Competition is increasing, hardware is improving and more efficient models are becoming available. The paradox is that even as individual tokens become cheaper, total AI spending can rise.<\/p>\n This is a familiar economic pattern. When something becomes cheaper, people tend to use more of it. With AI, lower costs encourage businesses to expand its use across more tasks and processes. A few employees experimenting with a chatbot becomes an entire team using AI coding tools. A handful of automated responses becomes thousands of customer interactions. A system that once generated text is now expected to complete entire workflows.<\/p>\n A company can cancel an AI subscription, but it cannot quickly rebuild lost organisational knowledge.<\/p>\n The issue is not that any single AI request is expensive, but that millions of cheap requests can quietly accumulate into a significant operating cost. This becomes especially important when companies justify AI adoption by reducing staff<\/a>. If a system can perform work previously done by ten people, the financial case appears straightforward: compare the cost of the technology with the salaries removed.<\/p>\n But that calculation is incomplete. The AI bill is only one part of the cost. Systems must be integrated, monitored and maintained. The system may still require engineers to integrate it, employees to supervise it, security teams to protect it, lawyers to assess its compliance and specialists to test whether its outputs are reliable<\/a>. Data must be cleaned and updated. Outputs must be checked for errors. Backup processes are required when systems fail or produce unreliable results. The AI may perform the visible task, but a broader infrastructure is required to make it dependable.<\/p>\n We need your consent to load this rte-player contentWe use rte-player to manage extra content that can set cookies on your device and collect data about your activity. Please review their details and accept them to load the content.Manage Preferences<\/a><\/p>\n From RT\u00c9 Radio 1’s Drivetime, how AI is affecting business<\/p>\n This means that a company may not have replaced a salary with a single technology cost. Instead, it may have replaced it with a collection of less visible expenses spread across software, infrastructure and human oversight. Some of these costs are difficult to track. Time spent fixing AI workflows, reviewing outputs or handling failures may not be counted as part of the system’s cost. Neither are the indirect effects, such as frustrated customers abandoning a process. A company can cancel an AI subscription, but it cannot quickly rebuild lost organisational knowledge.<\/p>\n There is also the question of who controls the price. Employees are expensive, but their costs are relatively visible. Salaries, pensions, taxes and benefits can be forecast. AI costs can change according to usage, model selection, task complexity, data volume and the commercial decisions of external providers. Recent reporting suggests<\/a> businesses are already moving less demanding tasks away from the most advanced American models and toward cheaper alternatives. Instead of using one premium model for everything, companies can route simple requests to smaller systems and reserve the most capable models for difficult work such as advanced coding, analysis or reasoning.<\/p>\n Open-weight models<\/a> offer another possibility because organisations can download, adapt and sometimes operate them within their own infrastructure rather than paying a frontier provider for every interaction. Chinese-developed models are already attracting attention for precisely this reason. Coinbase chief executive Brian Armstrong recently said<\/a> the company was experimenting with Chinese models as defaults because they were significantly cheaper than leading American alternatives.<\/p>\n We need your consent to load this rte-player contentWe use rte-player to manage extra content that can set cookies on your device and collect data about your activity. Please review their details and accept them to load the content.Manage Preferences<\/a><\/p>\n From RT\u00c9 Radio 1’s The Business, is it armageddon for Ireland’s tech industry?<\/p>\n The danger comes when organisations promised large AI savings and must now preserve those savings at all costs. For example, an organisation facing a rapidly increasing bill may be tempted to choose providers on the basis of headline token prices while asking fewer questions about where data is stored, whether customer information is used for training, who developed the model, how it was tested or what happens when it fails. The cheapest AI system may be cheap because somebody else is carrying the risk.<\/p>\n The lesson from tokenmaxxing is therefore not that businesses should stop using AI – nor is it that employees are always cheaper than machines. AI can undoubtedly automate useful work, increase productivity and allow people to complete tasks faster. The lesson is that consumption is not value, and replacing a visible labour cost with a poorly understood technology cost does not automatically create efficiency. Businesses need to measure the cost of completing an outcome, not merely the price of generating a token.<\/p>\n The lesson from tokenmaxxing is that replacing a labour cost with a poorly understood technology cost does not automatically create efficiency<\/p>\n They need to ask whether the system is reducing the time, money and error involved in completing useful work. They also need the ability to move between providers, retain enough human expertise to challenge automated decisions and recognise that the lowest-cost model may not be the lowest-risk choice.<\/p>\n The great promise of workplace AI was that businesses could remove expensive human labour and replace it with almost limitless digital intelligence. Tokenmaxxing reveals the weakness in that promise. Digital intelligence is not limitless. It is metered, supplied by somebody else and surrounded by costs that are only becoming visible as companies attempt to use it at scale.<\/p>\n The companies in the greatest danger may not be those with the largest AI bills. They may be those that removed the people before they understood what the replacement would really cost.<\/p>\n Follow RT\u00c9 Brainstorm on WhatsApp<\/a> and Instagram<\/a> for more stories and updates<\/p>\n The views expressed here are those of the author and do not represent or reflect the views of RT\u00c9<\/p>\n