Rippling's Real-World AI Model Test
Rippling, a company that provides HR, IT, and finance software for businesses, has published the results of an internal test involving 15 different AI models. The objective was to evaluate their performance on real-world payroll tasks using actual company data. This is a significant departure from typical AI benchmarks, which often use curated datasets or simplified tasks. Rippling's President and CPO, Matt MacInnis, shared these findings, aiming to provide a transparent look at how AI models perform in a production environment. The test involved approximately 2,100 scored agent runs per model. These agents were tasked with performing various functions critical to payroll processing. The surprising outcome was that the least expensive model tested achieved performance scores equivalent to the most expensive ones. This suggests that cost is not a direct indicator of capability when it comes to AI models handling complex, real-world business processes like payroll.
