The rule. When an AI system needs time, design the wait: name the stage the system is in, show the steps it has taken, and set an expectation for how long, instead of parking users behind a bare spinner.
Why. The tolerable wait for simple information retrieval is approximately 2 seconds, and the presence of feedback measurably prolongs what users will tolerate (Nah, 2004). The working guidance built on that finding: give an indicator to anything over about 1 second, use a looped animation only for waits of 2 to 10 seconds, and switch to a percent-done indicator or step count beyond that; in a study NN/g cites, users shown a moving progress bar were willing to wait on average three times longer (Sherwin, 2014). Reasoning models and agents have stretched AI waits from seconds to minutes, which promotes wait design from polish to core interaction structure: OpenAI's own announcement set the expectation for a deep research run at 5 to 30 minutes (OpenAI, 2025).
Seen in the wild. ChatGPT's deep research opens a sidebar summarizing the steps taken and sources used while it works, states the 5 to 30 minute expectation up front, and sends a notification when the report is ready (OpenAI, 2025).