Why build and train your own AI model instead of using a good open-source model run locally?

The vendor explained they built their own model after finding that general models, when they started deploying them about 18 months ago, weren't good enough at complex tasks like K-1 footnote extraction, since footnotes are written differently by preparer and general models struggled with them. They predict general models will catch up on this within about 12 months (citing noticeable improvement in Anthropic's models since December). Building proprietary models let them scale revenue quickly and gave them tighter security control versus relying on general-purpose models, which they view as a bigger security risk. They acknowledged their own solution is 'best-in-class but not perfect' at K-1 footnote allocation and prep, and expect the technology gap between proprietary and general models to close quickly as underlying models and security practices keep improving.