Rule-based chatbots: where they still win

The client ran a support team handling incident escalations and routine questions. We analyzed their conversation logs and found eighty percent of tickets fell into five common intent categories. A well-designed decision tree could resolve them automatically, quickly, and with error rates we could measure and report to stakeholders.

Building the tree meant understanding how support actually worked: what questions led to what conclusions, where did they need to escalate beyond automation, when could the system confidently close a ticket. Rules made sense here because the workflows were stable and the logic was explicit and documented.

We added a statistical classifier as a fallback, for conversations that did not fit neatly into the known patterns. Its job was to identify when a message fell outside the known patterns, not to write responses or make decisions. Classification was narrow enough to be reliable and interpretable.

What the model did not do was generate content or make judgment calls. It tagged ambiguous tickets as uncertain and sent them to a human for review. The split was clear: rules for the predictable and well-defined; learning-based classification only for identifying what you did not know.

That architecture – rules for the known, models for identifying the unknown – turned out to be what we built for most customers repeatedly. It keeps error rates predictable and auditable and keeps humans in control of the failures that matter most to the business and operations.

The reason it works is unsurprising once you think about it carefully. Perfect automation is rare and expensive to build. But partial automation that stays honest about uncertainty is reliable enough to reduce work without creating worse problems downstream for customers or support staff.