Cost per Completed Task

Cost per completed task is a metric from the world of economics. It answers a simple question: What does it cost to bring a complete assignment to a usable result? What is meant is not the price of a single computing step, but the sum of all steps up to the goal. An example: a program is supposed to read a scanned invoice and enter the amounts into a table. If it takes three attempts to do so and a human still has to briefly check the result at the end, all of that belongs in the calculation. Only at the end does a figure emerge, say four cents per invoice.

Why the price list per word leads you astray

Providers of AI models advertise prices per token. A token is a chunk of text, roughly half a word. These prices are tiny and therefore sound reassuring. But they say nothing about whether the work actually gets done in the end. A very cheap model that botches every second task can end up more expensive than one that costs twice as much but gets it right the first time.

The reason is rework. Every failed attempt costs computing time again. And if a human has to check the result, their working time often costs many times more than the machine. An hour of office work in Germany runs roughly 40 euros. Against that, a few cents of computing costs are barely visible. That is why the hit rate almost always weighs more heavily on total costs than the list price.

For companies, this metric is also the only one that allows comparison. You can place it alongside the costs of the previous solution, whether that involved humans or older software. Only this comparison shows whether switching is worthwhile.

What belongs in the calculation

You start with a clear definition of the task. It must have a distinct endpoint that can be classified as successful or failed. Then you have the system process a larger sample, say a thousand cases. You measure how many of these run through correctly without intervention.

Next, you add up the line items. This includes the computing costs of all attempts, including the failed ones. It includes search queries and database accesses that the system triggers along the way. It includes the time of the people who check or correct the results. Some also allocate a share of development and maintenance costs, because otherwise the system would appear too cheap.

A typical mistake is to count only the successful runs. That makes a shaky system look excellent. The correct approach is to divide the total expenses by the number of tasks actually completed. At a success rate of 50 percent, the amount automatically doubles.

Where the figure shows up in news and products

In quarterly reports from software companies, this metric appears when it comes to customer service. A provider might write, for instance, that a chat interaction now costs them one euro instead of six. The same calculation is used for programming tools, there per closed bug ticket.

More and more providers are even switching their pricing directly to this basis. Instead of paying per user and month, the customer pays per case resolved. This shifts the risk to the provider, because a system with many failed attempts eats into its profit.

The term should be distinguished from the pure operating costs of a model. Those describe only the machine’s share. Cost per completed task describes the economic benefit. When the news says AI has become dramatically cheaper, what is usually meant is the token price. Whether tasks have also become cheaper is an entirely different matter.