Any retail chain has that scale of data in surveillance tapes. Order of few mb/s * 10's of cameras per location * 100's of locations * 1000's of seconds per day * 100's of days per year => order of 10^15 bytes per year. So far it hasn't been economical to collect and analyze these tapes. But it is approaching feasibility, and AI is reaching the point it could give useful output on that massive input. Patterns of how shoppers interact with merchandise. What can you change to sell a few percent more items?