← All articlesAI · 8 min read
Where LLM assistants pay off in customer support, and where they do not
Vikram Nair, Founder & CEO ·
We have shipped four support assistants in the last eighteen months and declined two more. The pattern that separates the wins from the write-offs is simple: the assistant must have something to look up and something to do.
Where it worked
- A listings portal where 61% of enquiries were answerable from structured data and project documents.
- A SaaS help desk with a mature knowledge base and clear escalation rules.
- An e-commerce brand where order status, returns and exchanges could be read from the OMS and actioned through APIs.
Where we advised against it
A lender whose support questions were mostly about individual loan decisions. The answers lived in underwriters' heads, not in documents, and the regulatory risk of a confident wrong answer outweighed the savings. A chatbot that says "let me connect you to an agent" 80% of the time is a slower phone tree.
Measure deflection and customer satisfaction together. Deflection alone rewards an assistant that gets rid of people.
What we do on every deployment
- Retrieval over your own content with citations shown to the agent on handoff.
- A weekly evaluation set that grows with every flagged conversation.
- Hard limits on what the assistant may promise: refunds, pricing and legal statements always route to a human.