Notes from Serochat
Honest, practical writing on building AI customer support that answers from your own docs — and knows when to hand off to a human.
Why AI chatbots make things up
Hallucination is not a glitch. It is what a language model does when it has no source. Why it happens, what it costs, and what actually reduces it.
What ticket deflection actually measures
Ticket deflection counts the conversations that avoided a human. It cannot tell a customer helped from one abandoned — so here is what to check instead.
Signs you don’t need a chatbot yet
We sell one, so read this sceptically. Four situations where buying an AI chatbot now would be a waste, and what to do instead until they change.
Your customers don’t read the FAQ
They never read it, and they never will. The FAQ is usually right; the delivery is broken — and a doc-grounded bot is the documentation answering back.
What “grounded in your docs” really means
“Grounded” is not a buzzword. It is a concrete promise about where every answer comes from — and a way to check it.
Why we built Serochat to say “I don’t know”
Most support bots bluff when they are unsure. We think a good answer includes admitting when there is not one — and handing the conversation to a human.
When should support hand off to a human?
A confident wrong answer costs more than a polite “let me get someone.” Here is how we think about the handoff line.
Why source links beat confidence scores
A percentage tells your customer nothing they can act on. A link to the source tells them everything.
Measuring an AI support agent without vanity metrics
Deflection rate looks great on a slide and tells you almost nothing. Here is what we watch instead.