A Software Developer with a Very Expensive Habit

There is a peculiar sort of optimism that descends upon a modern corporate office when it is handed a tool of infinite capability and no visible price tag. It is the same optimism that leads a toddler to believe that the garden hose, once turned on, will eventually fill the entire parish with water, or that a small dog believes it can successfully digest a sofa. For several months, the corporate world has been operating under the pleasant assumption that artificial intelligence is a sort of magical, weightless ether—a substance that exists purely to turn three-word bullet points into three-page memos, entirely free of charge.

This illusion was recently shattered in rather spectacular fashion at Rippling, an enterprise software firm that, like many of its peers, had spent the early part of the year encouraging its staff to "tokenmaxx." This is a term that sounds like a minor character in a science fiction novel but actually refers to the practice of stuffing as many words as humanly possible into a large language model to see what happens. What happened, as it turned out, was a bill.

It was during a routine executive meeting in March that the company’s Chief Financial Officer presented a spreadsheet that caused several senior managers to look closely at their tea. The company was on track to spend approximately forty percent of its entire research and development headcount budget not on the salaries of the engineers themselves, but on the digital tokens those engineers were consuming. Had the trend continued, the company would soon have been spending ninety percent of its payroll budget on the privilege of having its employees talk to machines.

To put this in perspective, forty percent of an R&D budget is a very large sum of money. It is the sort of sum that could comfortably purchase a small island, a fleet of moderately reliable helicopters, or several thousand very fine grandfather clocks. Instead, it had been spent on inference.

Upon closer inspection, the management discovered that the distribution of this expenditure was not what one might call democratic. A mere ten to fifteen percent of the workforce was responsible for nearly two-thirds of the total bill. Most remarkably, a single software engineer had managed to run up a personal tab of fifty thousand dollars in a single month.

One must pause to appreciate the sheer, quiet majesty of this achievement. To spend fifty thousand dollars on digital tokens in thirty days requires a level of dedication that borders on the heroic. It is the computational equivalent of leaving every hot water tap in a mansion running while one goes on a fortnight’s holiday to Spain. One imagines this engineer sitting in a dimly lit room, repeatedly asking a frontier model to rewrite a basic sorting algorithm in the style of an eighteenth-century bishop, or perhaps demanding that it explain the concept of a semicolon using only metaphors involving badgers.

I once knew a junior clerk in Gloucestershire who spent an entire fortnight attempting to automate the filing of his own expenses, only to discover he had spent more on custom leather-bound ledger books than the value of the expenses themselves. There is a comforting, timeless human quality to this: the desire to build a machine so complex that its operation entirely eclipses the task it was meant to perform.

The response from Rippling’s executive team was swift and took the form of a product launch. Rather than simply telling their employees to stop asking the machine to write sonnets about database migrations, they did what any self-respecting software company does: they built a dashboard. The "AI Spend Console" was born, complete with a promotional video featuring the CFO sitting on a stool while employees cheerfully feed wads of actual cash into a paper shredder.

The tool is designed to do something that the major AI providers have absolutely no interest in doing: telling you how much money you are actually spending. As it turns out, the companies that sell tokens are not particularly keen on helping you buy fewer of them. They are much like a publican who, when asked if you have had enough to drink, merely smiles and offers you a double gin.

The new console functions as a sort of digital chaperone. It tracks which engineers are running up massive bills while their peers are quietly rewriting their code anyway. It also acts as an automated router, directing simple queries away from the incredibly expensive frontier models and toward cheaper, more modest alternatives. It turns out that if you merely need a machine to check if a comma is in the right place, you do not need to consult a multi-billion-dollar digital oracle that requires the electrical output of a small hydroelectric dam to function. A much smaller, cheaper model—perhaps one of Chinese origin like GLM 5.2, which is currently the darling of the cost-conscious developer—will do the job perfectly well for a fraction of a penny.

Through these methods, Rippling managed to reduce its token expenditure from forty percent of its budget to a much more civilized fifteen percent, all while consuming the same volume of tokens. They simply stopped letting the sales team use the most advanced models in existence to perform basic grammar checks on their emails.

Yet, there is a deeper, slightly more melancholy irony at play here. For years, the great fear of the modern office worker was the arrival of the automated supervisor—a cold, unblinking algorithm that would monitor their keystrokes and measure their bathroom breaks. Instead, we have arrived at a situation where human managers must now spend their afternoons monitoring the keystrokes of the algorithms to ensure they are not being too extravagant with their adjectives.

We have built a world where the employee is no longer watched to see if they are working, but to see if they are letting the machine work too hard on their behalf. The corporate panopticon has turned its gaze inward, not toward the human soul, but toward the API key.

There is a certain quiet dignity in watching a machine attempt to explain why it has spent several thousand dollars of someone else's money on a series of very polite, but entirely unnecessary, paragraphs. One suspects that as these spend consoles become standard office equipment, we will see the rise of a new corporate figure: the Token Captain, a person whose sole job is to walk around the office reminding people to turn off their chatbots when they leave the room, much like our parents used to do with the landing light.

In the end, we are reminded that no matter how advanced our technology becomes, the oldest laws of the universe remain undefeated. If you give a human being an infinite supply of something, they will eventually find a way to use fifty thousand dollars of it to do something that could have been accomplished with a pencil and a very brief moment of quiet reflection.