When you press Enter in an AI chatbot, roughly a dozen processing stages occur before your answer appears. The typed message isn't sent directly to the model—it's assembled into a document containing system prompts, tool definitions, memory, knowledge base documents, conversation history, and new input. This process, called context engineering, determines what information reaches the model and in what order.

Models are stateless by design, meaning they rebuild conversation history from scratch each turn and share resources with other users. The initial pause before text appears involves prefill and decode steps, caching, and streaming processes. Different providers structure this document differently—some retrieve everything upfront, others let the model fetch information during processing.

The assembled document impacts accuracy because models have finite attention budgets. Every token added gradually reduces precision, even on simple tasks. This explains why two products using identical models can produce different answers—the document wrapping the question differs. The same underlying model architecture, but different context engineering approaches, leads to varying outputs. Understanding this journey—from input assembly through safety checks to streaming responses—reveals why AI responses can vary and what happens during those seemingly idle seconds between your press of Enter and the first word appearing.