What a support chatbot needs before it goes live
The model is the easy part. Here is the checklist we run through before an assistant talks to a real customer.
Most chatbot projects fail quietly: the bot launches, answers a few questions badly, and the team switches it off within a month. The cause is rarely the language model. It is everything around it.
Clean knowledge
An assistant is only as accurate as its sources. Before launch we remove outdated pages, merge contradicting policies and write the twenty most common answers explicitly. Garbage in, confident garbage out.
Clear boundaries
Decide what the assistant must never do: quote prices it cannot verify, promise delivery dates, discuss competitors. Write these rules down and test them with adversarial questions.
A handover that works
The best assistants know when to stop. Handover should carry the full conversation and a one-line summary to a human, in the tool the human already uses.
Measurement
Track resolution rate, handover rate and customer satisfaction from day one. Without numbers, every complaint sounds like a failure and every compliment like a fluke.
A pilot, not a launch
Run the assistant on a share of traffic for two weeks. Review transcripts daily. Fix the top three problems. Then scale.