Using a design science approach, I test whether machine learning can replace the first-stage allocation of activity-based costing (ABC). I call this combination machine learning activity-based costing (MLABC). I conduct three numerical experiments using simulated datasets and find evidence that MLABC can produce relatively accurate overhead allocations like ABC if (1) the data include longitudinal correlations between cost drivers and cost resources, (2) correlations between cost drivers and cost resources include interactions, and (3) avoiding ABC’s cost study does not leave the firm ignorant of a cost driver that accounts for a substantial amount of variance between cost drivers and cost resources. I find limited evidence that MLABC can facilitate active experimentation with the firm’s cost function to learn more about it. I also conduct two supplemental mini-cases with data from practice. These mini-cases help test assumptions from my numerical experiments.