AI Is Cheap Until You Start Counting the Tokens

Artificial Intelligence (AI)

September 02, 2026

AI was supposed to make some work cheaper, and in many cases, it does. A company can use AI to draft content, analyze data, write code, answer customer questions and handle repetitive tasks in seconds. This, in turn, can free people to focus on work that actually requires judgment.

But there is a less comfortable question businesses need to ask: What happens when the cost of running the AI starts approaching, or even exceeding, the cost of the person it replaced?

That’s no longer just a theoretical question. Recent reporting has highlighted companies running into unexpectedly large AI bills as usage grows. For example, Uber reportedly exhausted its full-year AI coding budget within four months, while Microsoft has reportedly taken steps to rein in AI coding tool usage after costs became difficult to justify. Nvidia vice president of applied deep learning, Bryan Catanzaro, has also said that, for his team, compute costs can now exceed the cost of the employees using it.¹

The issue is not that AI is inherently more expensive than people. The issue is that AI costs are easy to underestimate.

The Token Problem

Most people don't think about AI in terms of tokens. They think about the monthly subscription. $20 here, $50 there, maybe a few hundred dollars for a more advanced business plan. This makes AI look remarkably inexpensive compared with hiring another employee.

But businesses using APIs, AI agents or high-volume automated workflows aren't really buying a subscription. They are paying for usage. And usage can multiply quickly.

Tokens are essentially the units used to process information through many AI models. A simple request may use relatively few. An AI agent completing a complicated task may make multiple calls, analyze documents, generate outputs, check its own work and interact with other systems. One task can therefore become many AI calls.

As Asia Financial recently reported, some companies have found token costs exceeding the cost of an employee within a month or two of heavy use.²

The per-token price may be falling, but that doesn't automatically mean the overall bill falls with it. If a business dramatically increases how much AI it uses, consumption can outrun those savings. Fortune recently highlighted this emerging paradox: cheaper tokens can still result in much bigger bills when companies significantly increase their AI usage.³ That's where the economics get interesting.

The Human Cost Hasn't Disappeared

There is another calculation that often gets missed. Replacing part of a person's workload with AI doesn't necessarily eliminate the human role. Someone still needs to decide what should be automated in the first place, and someone needs to check whether the output can actually be trusted. Beyond that, someone has to handle the exceptions, protect sensitive information, and keep an eye on the bill.

In other words, AI can remove work without necessarily removing responsibility. That doesn't make AI a bad investment. Quite the opposite. The value can be significant when AI is applied to the right task, with the right level of oversight and an appropriate model behind it.

The mistake is assuming that more AI automatically means more productivity.

The Better Question isn't "AI or People?"

It’s "Where does AI make economic sense?"

A business might discover that a task costing $40,000 a year in employee time can be reduced to $10,000 using AI. This can be advantageous. Another task might cost $30,000 in employee time but $50,000 in AI usage, infrastructure and human review. That one may not be worth automating. And a third task might be inexpensive to automate but too risky to leave without human oversight.

This is why AI adoption needs to be treated as a business decision, not simply a technology decision. The answer may involve using a smaller model for straightforward work, reserving more expensive models for complex tasks, setting usage limits, redesigning workflows or keeping a person involved at specific decision points.

Companies are already experimenting with these approaches as AI costs become more visible. EY, for example, has reported reducing internal token consumption by routing different tasks to more appropriate AI models rather than sending everything through more expensive models.⁴

AI works best when someone is still minding the shop

The most useful conversation about AI isn't whether it will replace people. It's whether it will make the business better. That means looking at the full cost of an AI workflow, not just the software subscription. It means measuring what the system actually produces, what human oversight it still requires and whether the result is worth the investment.

AI is a powerful tool. But like any tool, it needs to be used deliberately. We help businesses look at where AI can genuinely improve how they work, where it may introduce unnecessary cost or complexity, and where human judgment still matters. Because the goal shouldn't be to use the most AI. The goal should be to use the right amount of AI, in the right places, for the right reasons.

If you're questioning whether your current AI investment is actually delivering the value you expected, book a call with us today. Let's look at the numbers, the workflows and the opportunities together.