Article by Patrick Neeman (UX Collective). Argues that slow-feeling AI products are usually a front-end problem, not a model problem. Doherty’s 400 ms response-time threshold still applies, but generative AI splits the interaction into two clocks: the acknowledgment (interface, ~100 ms target, 400 ms ceiling, no inference required) and the answer (model latency, acceptable if honestly narrated). Teams that only measure final response latency leave the acknowledgment unowned, which erodes the human-in-the-loop supervision AI features depend on. Includes action items: echo input before the request leaves the client, show real agent state instead of thinking animations, measure acknowledgment latency in the field, give every wait an exit, and audit agent runs hop by hop. https://uxdesign.cc/stop-blaming-the-model-for-slow-ai-heres-how-to-design-for-it-doherty-s-threshold-as-a-guideline-5fa6d52e23fc?source=rss----138adf9c44c---4