chatbot

A chatbot is defined by what it does, not by how it is built: text arrives, text goes back, and no person wrote the reply at that moment. That is the whole term. It says nothing about quality or about the technology underneath: a menu of numbered options and a model trained on billions of sentences both qualify. The nearest neighbour is conversation automation, which is broader: automation covers everything a tool does to an inbox, including sorting and scheduling. A chatbot is only the part that talks.

What are the generations of chatbots?

Three, and they overlap in the wild rather than replacing each other cleanly.

• Pattern matching. The oldest, from the 1960s. The program looks for words it recognises and returns a reply written in advance. Say the wrong word and nothing happens.
• Intent classification. The program is trained to sort what you wrote into one of a fixed set of intents (“wants a refund”, “asking about price”), then plays the answer attached to that intent. Recognition improved; the answers were still written by hand.
• Language models. The program produces the sentence itself, word by word, from what it has read. Nobody wrote the reply anywhere, which is both the whole gain and the whole risk.

What separates a fixed script from a language model?

Predictability, at the cost of coverage. A script does exactly what it was told forever; a model handles what nobody anticipated and never does it identically twice.

	Fixed script	Language model
Where the reply comes from	written in advance by a person	produced at the moment
Unexpected input	falls through to a default	still produces something
Repeatability	identical every time	varies between runs
Cost of new cases	one branch per case	no branch to write
Typical failure	a visible dead end	a confident wrong answer

The second failure mode is the harder one to catch. A dead end announces itself; a fluent, plausible, wrong sentence reads exactly like a correct one.

What does a chatbot not know?

Anything it was not given. It has no access to a subscription status, a purchase history or last week’s exchange unless something hands it those; and it cannot check whether what it just said is true. The word “knows” is misleading. The program is producing likely text, not consulting a record.

That gap is why the interesting question is never “how good is the bot” but “what was it given, and what was it told to do with it”. A program applying a bad rule applies it to every conversation, at speed, without the hesitation a person would have had.

Where does a chatbot sit in a fan inbox?

Somewhere between two other things, and the distinction matters if you work in chatting. A message sequence sends text on a timer and never reads the reply. A chatter reads everything and writes everything. A chatbot reads the reply and answers it, which is why the rules it is given carry weight that a schedule never had.

Two of those rules are measurable. Our data shows that pitching a sale before the sixth message drops conversion by roughly a third, with the optimum after about ten exchanges, and that closing a sales message with an ellipsis is the worst-performing ending measured. A program cannot notice it is doing either. Whoever configures it decides once, for every conversation that follows.