Three generations of an AI support assistant

Every chat reached a person, including the many questions already answered in the help center. Agents spent their time on repeats, and customers waited for answers that existed.

1. An assistant with its own knowledge baseA chat bot backed by an OpenAI assistant and a knowledge base I wrote and structured for it. It proved customers would accept answers from a bot.
2. A co-pilot through n8nThe bot moved onto RepairDesk's own knowledge service. Its responses streamed in a format the chat platform couldn't read, so n8n sat in the middle and translated.
3. An async answer engineThe current version plugs into the platform's own bring-your-own-AI slot and answers with sources. It needed one more trick, below.

The chat platform waits five seconds for an answer. The knowledge service takes about seven. A faster model wasn't an option, so I changed the shape of the conversation instead:

• The documented endpoint was the wrong oneThe platform has two similar callback endpoints. Only one accepts answers for this bot type, and finding it took testing, not docs.
• No conversation ID on the first messageThe first question arrives before the conversation has an ID. The workflow now handles that case instead of failing it.
• A builder card failed with undocumented errorsI tried rebuilding the flow in the platform's no-code builder and hit unexplained error codes. I kept the working path rather than fight it.
• Quality checks need the full conversationA related workflow that reviews finished chats was first wired to a step that fires mid-conversation. It now listens for the conversation-ended event.

With routine questions handled, Support was restructured into a leaner Technical Success Manager model focused on problems that need a person.