The pitch for AI chatbots is almost always "deflect support volume." That's real, but it only pays off under specific conditions — and building one outside those conditions just adds a maintenance burden with no offsetting benefit.
Where they work
High query volume, low resolution complexity, and a bounded knowledge domain. Booking status lookups, FAQ-style pre-sales questions, and appointment scheduling are strong fits because the answer space is narrow and well-documented.
Where they don't
Low query volume doesn't justify the build and tuning cost. High-stakes, ambiguous queries (a patient describing symptoms, a loan applicant disputing a denial) need a human in the loop regardless of how good the model is — the chatbot's job there is triage, not resolution.
The metric that matters
Track deflection rate against escalation quality, not just deflection rate alone. A bot that deflects 40% of queries but escalates the remaining 60% with no context collected has made your support team's job harder, not easier.
Our AI Chatbots & Assistants team scopes every build against this framework before writing a line of code.