Deflection has a good version and a bad version
The bad version hides the contact button, buries the help address, and calls the resulting silence "efficiency." Customers notice, and they leave. The good version answers questions earlier and cheaper than a ticket would — same answer, less friction, for both sides. Everything below is the good version.
Step 1: find the repeats
Export a month of tickets and chat transcripts and tally the questions. Every support queue has a power law: a handful of questions generate a third or more of volume. That short list is your deflection roadmap — nothing else matters until those are handled.
Step 2: fix the cause where you can
Some repeat questions are documentation gaps; others are product gaps wearing a costume. "How do I reset my password?" at volume might mean the reset link is hard to find. "Where is my invoice?" might mean invoices should be emailed automatically. The cheapest ticket is the one the product stops generating — check each top question for a fixable cause before writing an answer for it.
Step 3: put the answer where the question arises
- On the page. If checkout generates shipping questions, the shipping answer belongs on the checkout page — a line of copy, a tooltip, an FAQ block. Zero support interactions.
- In the knowledge base. One public, searchable article per repeat question, written in the words customers actually use. This becomes the backbone of everything downstream.
- Through the AI layer. An assistant trained on that knowledge base answers the long tail instantly, at 3am, without queueing — and hands off to a human, transcript intact, when the question stops being repetitive.
Step 4: let chat absorb what email inflates
A ticket thread is a slow ping-pong: question, clarification, answer, follow-up — days of latency for minutes of content. The same exchange in live chat resolves in one sitting, so it never becomes a ticket at all. Keeping chat honestly staffed (or AI-backed after hours) converts multi-day threads into five-minute conversations.
Step 5: measure deflection honestly
Two numbers, tracked together: volume per customer (raw volume can rise while per-customer volume falls — that is growth, not failure) and resolution without escalation (of the questions your KB and AI absorb, how many stayed absorbed). And one guardrail metric: time-to-reach-a-human. If deflection "works" by making humans unreachable, you have built the bad version — watch that number like a covenant.
What this looks like in MyLiveChat
The knowledge base doubles as your public help center and the AI assistant's training source — one set of answers maintained once. Chat, KB, and AI share the same flat plan, so absorbing more questions never meters your bill.