Role playbook · Customer Support · Free
The support playbook.
Who this is for
The goal is faster replies without lowering the bar. AI drafts, triages, and finds the answer in the knowledge base. A person still decides what the customer hears, at least until the data says otherwise.
Support managers, team leads, senior agents, and the owner who still reads the support inbox at night.
Stop first
Three things to stop before you add AI to anything.
- Stop answering the same ten questions by hand. Write the ten answers down once. That document is the foundation for everything else.
- Stop putting a chatbot in front of customers before it has a knowledge base to read from. It will make things up.
- Stop measuring only speed. Add a quality check, or the fastest wrong answer wins.
First 30 days
Build the knowledge base and measure the queue.
- 01
Step 1
Pull the last 200 tickets. Group them. The top ten topics are usually 60 percent of the volume.
- 02
Step 2
Write or fix the article for each of those ten. Plain language, one question per article.
- 03
Step 3
Measure first-response time and resolution time for two weeks. That is the baseline.
- 04
Step 4
Use AI to draft replies on the ten topics, with an agent editing every one. Note what it gets wrong.
Days 31–60
Triage automatically, draft everything, send nothing unreviewed.
- 01
Step 1
Set up triage: every inbound ticket gets a topic, a priority, and a suggested article before an agent opens it.
- 02
Step 2
Drafted responses for all topics with a knowledge-base source. Agents approve, edit, or reject. Track which.
- 03
Step 3
Define escalation rules in writing: refund requests, legal language, anger, safety. Those go straight to a human, every time.
- 04
Step 4
Review rejected drafts weekly. Each one is either a missing article or a bad instruction.
Days 61–90
Let the numbers decide what goes to self-service.
- 01
Step 1
Compare first-response time, resolution time, and customer satisfaction against the baseline.
- 02
Step 2
For topics where drafts were approved unedited 95 percent of the time, consider self-service answers on the help page. Keep the human path one click away.
- 03
Step 3
Add sentiment flags so frustrated customers skip the queue.
- 04
Step 4
Give new hires the drafts as training wheels: they learn the house voice from approved answers.
Data rules
What goes into a model, and what never does.
- Customer conversations go only into a tool covered by your company agreement. Never paste a ticket into a personal chat account.
- Account changes, refunds, and anything involving payment details are done by a human in the system of record. AI can draft the reply that says it was done.
- If a customer asks whether they are talking to a bot, the answer is yes, and a person takes over.
What to measure
Four numbers. Record them on day one.
First-response time
Median, by topic, weekly.Draft approval rate
Share of AI drafts sent with light or no edits. Under 70 percent means the knowledge base is thin.Escalations caught
Tickets the rules routed to a human that should have been. Audit a sample monthly.Customer satisfaction
Whatever you measured before. Do not let it drop for speed.
What we would build first
Inbound triage with drafted first responses. Every ticket arrives classified, prioritised, and with a draft reply sourced from your own articles, waiting for an agent to approve. Replies get faster in the first week and the bar does not move.
If you want this plan written for your business instead of your role, that is the AI Audit. If your team needs the skills the plan assumes, that is AI for Customer Support.