Writing 'big' software requires so many non-technical resources (marketing/legal/training/testing) that any AI productivity gains are minimal (writing the software is only 15% of lifecycle anyway) Writing small software (in house, prototypes) is small ball anyway so will have only small effects.
