Daikin builds field AI on 3000時間 of repair video

Daikin's efforts sit at the intersection of two pressures: a shortage of engineers who repair and inspect air-conditioning equipment, and the limits of general-purpose AI models. 比戸将平 notes that models like ChatGPT and Gemini do not currently cover air-conditioning expertise or Daikin's own product operations, which is why the company is embedding that knowledge into a custom LLM.

The maintenance video data — more than 3000時間 — is the kind of material that cannot be easily collected from the internet, as the dialogue notes, because it comes from actual field work. Daikin is pairing that visual data with a vision-language model (VLM, an AI that handles images and text together) to support engineers and training. Alongside this, it is using LLMs to speed up design and development, aiming to differentiate its products.