AI Decision Models and Confidence Scores

In the world of artificial intelligence, unlike long-term reasoning models designed to solve complex problems, new decision models that make fast and routine choices are coming to the fore. Quickly selecting from a set of fixed probable answers to a specific question, these models append a confidence score indicating the accuracy of the response. When the confidence score falls below a predetermined threshold, the decision is routed to a human.

Last month, TypeSafe AI introduced the "System One" decision model category with its Jev model, developed to make rapid decisions enabling software to act directly. Companies such as OpenAI and Microsoft quickly followed suit with developments in this field. IBM Research Scientist Kaoutar El Maghraoui noted that traditional large language models (LLMs) generate responses slowly by processing tokens one by one, whereas decision models read the input once and arrive directly at a result. IBM Data Platforms CTO Volkmar Uhlig emphasized that these models have gained popularity thanks to the rise of AI agents operating in the background and completing tasks on behalf of humans. According to a survey of 2,000 CEOs worldwide by the IBM Institute, 25 percent of operational decisions are currently made by artificial intelligence without human intervention, and this rate is expected to rise to 48 percent by 2030.