Hardening an enterprise WhatsApp chatbot: lessons from banking

Building a chatbot is easy. Making it reliable enough for a bank is engineering. Sub-workflows, a QA supervisor agent, and logs you can actually read.

A WhatsApp chatbot demo takes a week. The same chatbot surviving real customers, real edge cases and a banking compliance review is a different project entirely. We took over one of those after user-acceptance testing had surfaced the gap.

Three moves that changed everything

→Split the monolith: one giant n8n workflow became a set of small sub-workflows, each testable and deployable alone. A bug in payments no longer breaks onboarding.
→Put an agent in charge of QA: a supervisor agent replays critical conversations in production and flags any drift in tone, accuracy or latency — before customers do.
→Make logs diagnostic: every conversation leaves a trace a human can read in seconds, not a JSON archaeology dig.

The uncomfortable truth

Most chatbot failures are not AI failures — they are software engineering failures around the AI: no tests, no monitoring, no isolation. If your bot matters, treat it like the production system it is. The AI is the easy part.