Across 6,200 workers and 1.5 million work messages, @OpenAI has four months of data on which AI tasks people keep coming back to. Some cross-occupation tasks were far stickier than others. Workers returned the following month to discussing products with customers 54% of the time, promotional writing 44%, and creating marketing materials 37%. Explaining financial information came back just 15% of the time, below the 18.5% average. OpenAI reads the difference as workplace norms and the perceived consequences of getting something wrong, and we would put it more bluntly. People keep using AI where a wrong answer is cheap and immediately visible, and quietly stop where checking costs more than doing the work. Lower the price of verification and the 15% tasks start to look like the 54% ones.