Chatbot
A program that holds a text or voice conversation with a customer through a chat interface — spanning everything from a scripted decision tree to an LLM-powered agent.
A chatbot is defined by its interface, not its intelligence: it is any program that carries on a conversation through a chat window, and that one word covers a capability range of several orders of magnitude. The lineage runs from ELIZA, the 1966 MIT program that faked a Rogerian therapist with keyword pattern-matching and no understanding at all (History.com), through decision-tree bots that walk you down scripted buttons, to intent-based systems that classify a message and fire a canned reply, to today's LLM-powered agents wired to a knowledge base and allowed to take actions. Calling all of these "a chatbot" tells you almost nothing about whether the thing can actually resolve anything.
They are now everywhere in support, and adoption keeps outrunning competence. The CFPB estimates that over 98 million people — roughly 37% of the US population — used a bank's chatbot in 2022, a figure it projects will pass 110 million by 2026 (CFPB). The same report catalogues the failure modes regulators now see at scale: bots that don't recognise a customer is raising a dispute, that serve inaccurate answers with total confidence, and that trap people in "doom loops" of repetitive, unhelpful prompts with no offramp to a human. A conversational interface is not the same as a resolution.
What it hides: the label flattens a huge quality gap, which is exactly how it gets exploited — a scripted FAQ deflector and an autonomous resolver are sold under the same word, so buyers end up comparing demos instead of confirmed resolution rates. It hides that "conversational" says nothing about "correct": fluent, confident, and wrong is a chatbot's default failure mode, and the company owns whatever its bot tells a customer. And it hides who is doing the choosing — a large share of chatbot sessions are customers trying to escape the bot to reach a person, so a high engagement or "deflection" number can be a measure of obstruction rather than help. Judge a chatbot by confirmed resolution and repeat-contact rate, not by the fact that it answered.