Read the full customer conversation before using AI.
Use only approved SOPs, policies, templates, and scenarios.
Review and edit every AI suggestion before sending.
Escalate when policy, risk, or customer tone requires it.
Standard Prompt
Agents should not paste a conversation and ask, "What should I say?" Use this structured prompt so the AI checks the issue, policy, missing information, escalation need, and suggested reply.
Copy this prompt into the AI notebook
You are my customer service assistant. Use only the SOPs, policies, templates, and useful scenarios in this notebook. Here is the customer conversation: [Paste full conversation here] Please return your answer in this format: Customer Issue: Relevant SOP / Policy: Missing Information: Escalation Needed: Suggested Reply: Agent Checklist Before Sending: Rules: - Do not guess. - Do not promise refunds, discounts, timelines, or exceptions unless the SOP clearly allows it. - If information is missing, ask the customer for it. - Keep the customer reply polite, clear, and professional. - Do not include internal policy details that the customer should not see.
Required AI Output
The AI answer should be easy for agents to review. It should separate analysis from the customer-facing reply.
Customer Issue
A short summary of what the customer needs, without guessing.
Relevant SOP / Policy
The exact SOP, policy, template, or scenario used to support the answer.
Missing Information
Anything the agent must ask the customer for before giving a final answer.
Escalation Needed
Yes or no, with a short reason tied to policy or risk.
Suggested Reply
A ready-to-send draft in a polite, professional, helpful tone.
Agent Checklist
A final check so staff do not blindly copy the AI response.
Daily Workflow
Follow these steps every time. The goal is to understand the case first, use AI as a drafting assistant, and send only a reviewed final response.
Read the customer conversation
Identify what the customer wants, their tone, the support category, and whether enough information is available.
Copy the full conversation
Include the latest customer message, previous team replies, order number if visible, product or service, screenshots, and useful details. Avoid unnecessary private information.
Open the correct AI notebook
Use the notebook that matches the case: Billing and Refund Support, Technical Support, Shipping and Delivery, Complaint Handling, or the general Customer Service SOP Assistant.
Paste the standard prompt
Ask the AI to identify the issue, relevant SOP, escalation need, missing information, and customer-ready response.
Review the AI analysis
Check whether the AI understood the case, used the correct SOP, followed policy, requested missing information, and avoided unauthorized promises.
Edit the suggested reply
Confirm the customer name, order details, tone, and accuracy. Remove internal notes, AI analysis, and private policy details before sending.
Decide whether to escalate
Escalate if the SOP, AI analysis, customer tone, or risk level requires manager review.
Send the final response
Send the edited response through the normal support system. Write as the company representative, not as "the AI."
Save useful new scenarios
If the case is new, difficult, or helpful for training, save the approved scenario in the Useful Scenarios folder.
Review Before Sending
The customer should only receive the final answer, not the AI's reasoning or internal notes.
Agent checklist
- Customer name is correct.
- Order, account, or product details are correct.
- The correct policy was used.
- No unauthorized refund, discount, timeline, or exception was promised.
- No internal notes or AI analysis are included.
- No private company policy details are revealed.
If the AI answer seems wrong
Ask the AI to check again before editing the final reply.
Correction prompts
Check again. Which exact SOP supports this answer? Rewrite the answer more carefully and do not promise a refund unless the SOP allows it.
Escalation Rules
Escalate when the case has high risk, policy exceptions, approval limits, public reputation concerns, or unresolved technical issues.
Common escalation cases
- Customer threatens legal action.
- Customer asks for an exception outside policy.
- Customer is very angry.
- Refund amount is above the approval limit.
- Possible fraud or payment issue.
- Public review or social media complaint.
- Technical issue cannot be solved with normal steps.
- Customer claims the company caused a serious mistake.
Escalation template
Copy for manager review
Hi [Manager Name], I need help reviewing this customer case. Customer issue: [Short summary] Relevant policy: [Policy name] Why I am escalating: [Reason] Suggested next step: [What you think should happen] Customer conversation: [Paste conversation]
Save Useful Scenarios
Approved examples make future AI answers better. Save only helpful, reviewed cases and remove unnecessary personal details.
Scenario format for 05_Useful_Scenarios
Scenario Title: [Example: Customer requested refund after using service for 3 months] Customer Situation: [Summarize without unnecessary personal details] Correct SOP Used: [Policy name] Final Response Sent: [Paste the final approved response] Result: [Customer accepted / escalated / refunded / rejected / pending] Lesson Learned: [What future agents should do in similar cases] Approved By: [Manager name, if needed] Date: YYYY-MM-DD
Quality Control
These rules keep responses accurate, consistent, and safe.
AI suggestions are not final answers
A human must review, edit, and approve every response before it goes to the customer.
Use only approved sources
The AI should use only SOPs, policies, templates, and approved scenarios in the notebook.
Update SOPs regularly
Every week, add useful scenarios and fix unclear SOPs. Every month, review major policies and templates with managers.
Keep one source of truth
Use clear file names like "Refund Policy - Current" and archive old versions instead of keeping many random copies.
Track changes
Each SOP should show the last updated date, owner, approver, and change history.
Risk Management
Most problems can be prevented by using approved sources, reviewing responses, and keeping SOPs current.
| Possible Risk | How to Prevent It |
|---|---|
| AI gives the wrong answer. | Require human review, ask the AI to name the SOP it used, and use the checklist before sending. |
| SOP is outdated. | Review SOPs monthly, assign one owner for each document, and archive old documents. |
| Staff paste sensitive data. | Paste only information needed for support, avoid unnecessary personal details, and follow privacy rules. |
| AI makes promises. | Tell the AI not to promise anything outside the SOP. Agents must check refunds, discounts, deadlines, and exceptions before sending. |
Implementation Timeline
Build the system in phases so the AI notebook has clean knowledge, approved examples, and trained staff.
Build the knowledge base
Create the Google Drive folder structure and add current SOPs, templates, policies, and common scenarios.
Standardize documents
Use the same SOP format: purpose, when to use, policy, required information, steps, template, escalation rule, owner, and last updated date.
Create the AI notebook
Create the NotebookLM notebook and add only the approved Google Drive documents as sources.
Test real cases
Use 10 to 20 past conversations and compare AI suggestions with approved responses.
Train staff
Teach agents how to paste conversations, ask the AI, review answers, edit replies, and save useful scenarios.
Improve weekly
Review difficult cases, AI mistakes, missing SOPs, useful scenarios, and outdated templates.
Simple Version for Staff
Use this short version when agents need a quick reminder.
Quick workflow
- Read the customer conversation.
- Open the correct AI notebook.
- Paste the full conversation using the standard prompt.
- Ask the AI to identify the issue, SOP, escalation need, and suggested response.
- Review the AI answer carefully.
- Edit the response so it is accurate and natural.
- Escalate if required.
- Send the final answer to the customer.
- Save useful new cases into the Useful Scenarios folder.