The chatbot era gets remembered as a disappointment, which is unfair to what it proved and useless for understanding what came next. Chat taught the industry three durable lessons and ran into two walls no amount of model quality could break. Both halves of that story shaped everything built since.
What chat proved, permanently
Lesson one: customers will talk to software. The industry genuinely didn't know this in 2016; the assumption was that automated help would be tolerated at best. What happened instead was preference in the majority case, for documented questions, instant beat human-with-a-queue almost every time. Lesson two: the knowledge base was the product all along. Chatbots forced companies to discover their docs were the real interface to their support, wrong, stale, and unsearchable, and the discipline of grounding answers in maintained knowledge became table stakes for everything after. Lesson three: deflection has economics. The cost curve of answering a documented question fell to almost nothing, which permanently reset what "support at scale" was allowed to cost.
The two walls
The first wall was blindness. The bot could not see what the user saw, so every state-dependent problem, the misconfiguration, the ambiguous error, the wrong turn mid-workflow, began with the user describing their screen in words they didn't have, and the bot guessing at a reconstruction. LLMs made the guesses smarter and the wall stayed exactly where it was: the input was still testimony, not truth. The second wall was statelessness. Every exchange was a quick hit; the window closed and the context died, so any job spanning steps, sessions, or days, setup, evaluation, learning, was structurally out of reach. The bot could answer a hundred questions about onboarding and could not onboard anyone.
Why "chatbots are dumb" was the wrong diagnosis
By 2023 the answers were often excellent, and the frustration persisted, which should have told everyone the problem wasn't intelligence. Users weren't angry that the bot answered badly; they were angry that answering was all it could do: describe, paste, follow steps, fail, repeat. The ceiling was made of missing capabilities, perception, action, memory, not missing IQ, and the tools that broke through in the years after broke through by adding organs, not by adding smarter answers to the same blind, amnesiac form factor.
What to keep from the era
Respect, mostly. Many teams reading this run chat today, and should: for documented questions at volume, the chat layer's economics remain excellent, and the grounding discipline it forced is the foundation live agents stand on. The era's real bequest is a clean diagnostic: whenever help fails, ask whether it failed on knowledge, chat's solvable problem, or on state, action, or continuity, the walls chat could never cross. (Disclosure: we build what came after at Skippr, and we'd rather honor the foundation than sneer at it, the line is "chat was the right first step, and it hit a ceiling.")
See what a live agent actually does
The category is easier to watch than to define. Fifteen minutes is enough to see where the mechanism differs from everything it gets confused with.