Businesses Finding It Harder to Predict AI Spending

A recent Wall Street Journal report published on October 5 suggests that businesses are struggling to predict their AI expenses. According to the study, only 11% of 400 surveyed companies could accurately forecast their AI costs. Unlike regular software, AI works more like a human worker, taking actions, making decisions, and occasionally making mistakes.

More advanced AI models usually cost more per token, but they can also perform tasks more efficiently, potentially lowering costs. However, using a cheaper model for tasks it's not designed for could result in unexpected costs due to an increase in token usage. This was demonstrated by a study conducted by researchers from Stanford University, Carnegie Mellon University, the University of California, Berkeley, and Microsoft Research, who had AI models complete over 6,800 tasks in various areas like math, programming, and science.

In 32% of these scenarios, cheaper models ended up costing more than their more expensive counterparts. Lingjiao Chen, one of the researchers, emphasized that price should not be the sole factor in determining which AI model is actually cheaper. Meanwhile, a separate report from September by PYMNTS Intelligence found that 60 U.S. companies surveyed were using AI in 7 out of 8 business functions, including payments, finance, product development, and customer experience.

However, only 20% of the 437 deployments examined were integrated into a function's regular work. The study also revealed that 62% of these companies spent more than $10 million on new AI tools in the previous year. Companies with deeper AI deployment faced more challenges such as skill gaps and unclear responsibility for the technology.

The report noted that firms that had already incorporated older AI forms in at least two functions before 2022 were more likely to have embedded newer AI in three or more functions.