Why most banking AI projects are set up to fail — not because the models are wrong, but because the integration infrastructure beneath the AI layer was never designed to support production AI at the speed and data quality it requires.