Scoping a Trustworthy Support Agent

This is a teardown of an AI support agent shipped on Base44 for a B2B SaaS client. The goal: let customers ask natural-language questions about their account and usage, and get answers grounded in real data — without the agent ever inventing a record, deleting something, or escalating a hallucinated ticket.

The brief

The client wanted a support agent that could answer questions like “why did my invoice jump last month?” and “how many seats am I using versus my plan?”. The data lived in three places: the billing entity, the subscription entity, and the usage logs. A human support rep could answer in 5–10 minutes. The agent had to answer in seconds — and be wrong zero times on the facts.