Task Completion Rate

The task completion rate is a metric. It answers a simple question: how many of the assigned tasks were actually finished in the end? You give a system or a person one hundred tasks and count how many of them were fully solved. If it’s 72, the rate is 72 percent. The English name literally means “rate of task completion”; in German it’s also called Erfolgsquote (success rate). The word “fully” is important: half-finished solutions don’t count.

Why half-finished solutions are worthless

Many older metrics measure individual steps. A language model, for instance, might correctly predict the next word 95 percent of the time. That sounds good, but it says little about whether a task actually gets completed in the end. That’s because a real task often consists of twenty or fifty steps. If even a single one goes wrong, the result is unusable.

Do the math. With a 95 percent hit rate per step and twenty steps, only around 36 percent of runs succeed in the end. Small error rates thus multiply into large failure rates. This is exactly the effect that the task completion rate makes visible, while metrics for individual steps conceal it.

For businesses, this is the decisive figure. Software that fails to complete three out of ten bookings doesn’t save any work. A human has to check and fix the failed attempts. Only once a high success rate is reached does deployment pay off economically.

How such a rate is measured

First you need a collection of test tasks, often called a benchmark. This is a fixed list of assignments that all systems work through equally. Each task comes with a clear success criterion. Example: the file is in the correct location, or the program runs through without an error message.

Then you count and divide: successful tasks divided by the total number of tasks. Checking is done either automatically by a test script or by hand by humans. Automatic checking is cheaper but only works for tasks with an unambiguous outcome. For a text like “Write a polite rejection,” a human has to make the judgment call.

The most common mistake is comparing rates from different tests. An 80 percent rate is meaningless without knowing the task list. Anyone who chooses easy tasks gets high numbers. That’s why credible reports always state which benchmark was used and how many attempts per task were allowed.

Where this figure shows up in news and products

You most often encounter it in connection with AI agents. These are programs that carry out multiple steps autonomously, for example buying a ticket in a browser. When a manufacturer unveils a new model, it almost always cites such a success rate. Typical figures for software tasks currently range between 30 and 80 percent, depending on the test set.

The metric is also old and widespread outside of AI. In website design, it’s used to check whether test users can fill out a form. In customer service, it’s used to count what proportion of inquiries were resolved without a follow-up question. The logic is always the same: count what was fully completed.

When reading announcements, some skepticism is warranted. A high rate can mean the model is good, or that the test was easy. The number only becomes useful in comparison: the same task set, different systems, identical conditions. Other metrics round out the picture, such as the time required and the cost per completed task.