LESSON 4 OF 6
Support: help without annoying people
BY THE END OF THIS LESSON
Handle routine questions faster without degrading the customer experience.
Deflection is the wrong goal
Optimise for deflection and you build something that stops customers reaching you. That is not a support win, it is a churn mechanism with a dashboard.
The goal is resolution: the customer got what they needed, faster than before, and did not have to fight to reach a person when they needed one.
Where it works, and where it must not
Works: answering documented questions, drafting replies for an agent to review, triaging and routing, turning a resolved ticket into a knowledge-base entry, translating.
Must not: anything involving a refund decision, an account action that cannot be undone, a complaint that has escalated, or a customer who is upset. Route those to a person immediately — the escalation path is the most important part of the design, not an edge case.
Ground it in your own answers
Never let it answer from general knowledge. Give it your documentation and your resolved tickets, and instruct it to say "I do not have that documented, let me get someone" rather than improvising.
An invented policy is worse than a slow reply, because you will be held to it.
WORKED EXAMPLE
IMPROVISING
You are a helpful support agent for our software. Answer the customer's question.
GROUNDED
Answer only from the documentation below. [docs] If the answer is not in it, reply exactly: "I want to get this right — let me put you with someone who can confirm." Do not improvise a policy, price, timeline or exception. Route to a human immediately if the message mentions a refund, cancellation, data deletion, legal action, or reads as upset — regardless of whether you could answer it. Match our style: plain English, no apology-padding, say what happens next.
The first will confidently invent a refund window. The second cannot, and the escalation rules fire on the cases where being fast is worth least and being right is worth most.
YOUR TURN
Design the escalation rules before the answers.
RUN THIS
I want to use AI on first-line support. Before we discuss what it should answer, help me define what it must never handle. Ask me about my product, what customers get upset about, and which actions cannot be reversed. Then give me: an escalation rule list, the exact wording it should use when handing over, and the three question types that are safest to automate first. My business: [describe it]
What a good result looks like
You want a short, unambiguous escalation list. Anything fuzzy will be interpreted generously in the wrong direction.
KNOWLEDGE CHECK
Answer all 3 correctly to complete this lesson.
1. Why is deflection the wrong metric?
2. What should it do when the answer is not in the documentation?
3. Which should always route straight to a human?
REMEMBER THIS
Ground every answer in your own docs, and design the escalation path first.