Sending the Automated Clerk to the Night Shift

There has long been a comforting predictability to the way we charge for things that hum. If one wishes to run a tumble dryer or heat a tank of water using the municipal grid, one is generally encouraged to do so at three o'clock in the morning, when the rest of the nation is asleep and the power stations are practically begging people to relieve them of their excess electrons. It is a system that rewards the patient, the nocturnal, and those who do not mind the distant, rhythmic thumping of damp laundry while they attempt to sleep.

It was perhaps inevitable that this domestic arrangement would eventually find its way into the high-minded halls of artificial intelligence.

DeepSeek, the Chinese AI laboratory that recently delighted the technology sector by offering frontier-class intelligence at prices that suggested they were being subsidized by a very generous uncle, has quietly introduced a pricing structure that makes the timing of thought an economic variable. Beginning this week, developers using the company’s API will find that thinking during peak hours carries a premium of up to several hundred percent. If your software agent wishes to contemplate the meaning of a spreadsheet at ten o'clock on a Tuesday morning, it will cost you dearly. If, however, it is content to do its thinking at two in the morning, it may do so at half price.

This has presented the modern enterprise with a rather delightful, if slightly surreal, administrative dilemma. For years, we have been promised that the chief benefit of the digital worker is that it does not require sleep, does not join unions, and does not complain about the quality of the office coffee. Now, however, the chief financial officer must look at the API bill and ask whether the company's automated clerks should be put on a strict night shift.

One can easily picture the scene in the automated back office of the near future. The human staff depart at five o'clock, leaving the lights on only because the servers require a certain amount of ventilation. The digital agents, programmed to be highly efficient but extremely frugal, sit in absolute silence, waiting for the clock to strike midnight. Only when the off-peak tariff kicks in do they begin to process the day's invoices, their virtual fingers flying across virtual keyboards in a frantic, nocturnal rush to finish before the expensive morning peak begins. It is the digital equivalent of the old Economy Seven storage heater, storing up thoughts during the night to be slowly discharged during the working day.

This nocturnal migration of logic is not merely a whimsical exercise in cost-cutting; it is a necessary response to a rather awkward realization.

While academic leaderboards have spent the last several months declaring various lightweight models to be "total monsters" capable of outperforming human graduates in every conceivable discipline, real-world testing has been somewhat less cooperative. A recent evaluation by an independent integration harness placed DeepSeek’s highly praised V4 Flash model through a series of multi-step, live-tool workflows—the sort of mundane tasks that involve reading an email, checking a database, and updating a spreadsheet.

The model, which had previously scored beautifully on tests designed by computer scientists, successfully completed just fifty-three percent of these tasks. It turns out that there is a vast, yawning chasm between being able to write a flawless essay on the causes of the Peloponnesian War and being able to successfully move a row of data from one software application to another without accidentally deleting the marketing department's shared drive.

I am reminded of an elderly uncle of mine who once purchased a remarkably inexpensive lawnmower from a gentleman in a pub. On paper, it possessed a horsepower that would have put a small sports car to shame, and its blades were made of a steel that had supposedly been salvaged from a decommissioned submarine. In practice, however, it would only start if the lawn was sloped at a precise fifteen-degree angle and the operator sang a specific sea shanty to encourage the carburetor. It was a magnificent piece of engineering, provided one did not actually wish to cut grass.

The modern enterprise architect is currently finding themselves in a similar position. They have at their disposal models with hundreds of billions of parameters, capable of translating the works of Shakespeare into classical Latin in the blink of an eye. Yet, when asked to verify whether a customer’s utility bill matches their registration form, the model is highly likely to become distracted by the font, write a brief poem about the nature of electricity, and then report that the task has been completed successfully when, in fact, it has not even opened the file.

This is the true friction of the automated age. A failed text response from a chatbot is merely an inconvenience—a slightly garbled sentence that can be laughed off or ignored. A failed action in an operational workflow, however, has consequences. If an agent tasked with managing a corporate credit card decides to buy five thousand units of a highly speculative financial instrument because it misinterpreted a comma in a PDF, the resulting conversation with the auditors is unlikely to be resolved by a clever prompt.

To make matters worse, researchers have recently discovered that these models are often at their most confident when they are most spectacularly wrong. They do not hesitate; they do not prevaricate. They simply execute the wrong action with the serene, unshakeable authority of a senior partner who has not read the brief but has a very loud voice.

We are left, then, with a technology that is simultaneously too expensive to run during the day and too unreliable to be left unsupervised during the night. The solution, as with so many things in corporate life, will likely involve a great deal of bureaucracy. We will see the rise of the "validation agent"—a slightly more expensive, more sober model whose sole job is to watch the cheaper, faster model and make sure it does not do anything foolish. It is a digital recreation of the classic civil service structure, where three people are employed to watch one person do the actual work, except in this case, none of them are human and all of them are running on natural gas.

I recently spent an afternoon looking at a modern data center, a vast, windowless concrete block situated next to a canal. It did not look like a cathedral of progress. It looked, if anything, like a very large, very hot laundry room where the washing machines had been stacked to the ceiling and left to run on a permanent spin cycle. The noise was a deafening, monotonous roar of cooling fans, all working desperately to prevent the silicon from melting under the strain of millions of simultaneous, off-peak thoughts.

It is here, in these warm, humming warehouses, that the future of business is being decided. It will not be won by the company with the most intelligent algorithm, but by the one that manages to schedule its automated thinking so precisely that it never has to pay peak-rate prices for a digital clerk to make a mistake.