PKSHA Technology's 森下賢志: at times, a 時給1400円 part-timer beats generative AI

PKSHA Technology was founded in 2012 and worked on putting AI into practical business use before generative AI appeared, so Morishita's view comes from that longer history. He points back to the earlier deep learning boom, when GPU-based processing was applied to tasks that rule-based AI or machine learning could already solve, raising costs. He says the current generative AI boom is not so different, and that the first step in deployment is deciding whether generative AI is really needed from a cost standpoint.

The examples in the article show how that check plays out. In credit card fraud detection, millisecond-level processing and monthly data in the tens of millions of records per company make machine learning the better fit. For Summit Store's employee work-allocation tables, PKSHA used mathematical optimization rather than generative AI, because shift creation has strictly defined outcomes while generative AI output fluctuates. At the same time, Morishita treats natural language processing as generative AI's strength, noting that manufacturing has seen more use cases for handwritten documents and layout-heavy materials.

Morishita's own framing for choosing a system is how much it must match the particularity of the work, alongside the choice between using SaaS, building in-house, or waiting. The stakes appear to hinge on whether companies run that cost and fit comparison before committing budget, since the article suggests the money is hard to justify when a cheaper human option already does the task.