Early chatbots followed scripts. If a user typed something unexpected, the conversation broke. Modern assistants feel different because they model language broadly — and increasingly because they connect to tools, not just text.
Three phases that matter for practitioners
1. Rule-based bots (1990s–2010s): keyword routes and canned answers. Cheap to run, brittle to maintain.
2. LLM chat (circa 2022–2024): flexible language, but limited action unless users copy-paste results elsewhere.
3. Tool-aware systems (today): models that search, call APIs, run code, and hand off to humans with context.
Why prompting became a job
When software behavior is specified in natural language, someone has to define intent, edge cases, and quality standards. That is prompt engineering — whether the title exists on your team or not.
What is next
More specialization: domain-tuned workflows, better evaluation, and tighter integration with company data. General chat will remain useful, but business value shifts toward reliable systems built on top of models.
Understanding this arc helps you choose the right tool for a task — and explain to clients why "just use ChatGPT" is not a strategy on its own.