When AI Should Not Approve a Refund (And Other Human Review Rules)

AI is great at answering "where is my order?" It is terrible at deciding whether a customer deserves a full refund for a damaged package they may or may not have received.

The difference between good AI support and bad AI support is not the AI. It is the rules that decide when AI acts alone and when it stops and hands the ticket to a real person.

Here are suggested guardrails to agree with your support operator before letting any automation touch customer tickets.

AI resolves automatically (no human review needed)

These ticket types are predictable, have a known answer from store data, and carry low risk:

• Order status and tracking — the answer is in the shipping data
• Shipping availability — the answer is in the shipping policy
• Stock status — the answer is in the inventory
• Store hours and location — the answer is on the website
• Product specs — the answer is in the product description
• FAQ questions — the answer is already documented

These can be candidates for automation when the source data is reliable and you have tested the answers. Measure your own ticket mix rather than assuming a fixed share.

AI drafts, human approves

These ticket types need a human to click approve before the response goes out. AI prepares the draft, a person reviews and sends.

• Address changes — the AI collects the new address, a human confirms it is correct before updating the order
• Return requests — the AI checks the return window and starts the process, a human confirms the reason and condition
• Exchange requests — the AI identifies the item and exchange policy, a human confirms the replacement SKU

This layer catches the 15% of tickets that are mostly routine but have a small chance of going wrong.

AI escalates immediately (human takes over)

These ticket types never get an AI response. The AI detects the intent and routes the ticket directly to a human.

• Complaints and angry customers — the AI sends an acknowledgment ("I understand why this is frustrating. I am handing this to a person now.") and stops
• Refund disputes — the AI never approves or denies a refund on its own. A human reviews the order history, the customer's claim, and the store policy
• Compensation requests — the AI never offers discounts, store credit, or compensation. A human decides
• Repeated issues — if the same customer contacts support 3+ times about the same issue, a human takes over
• Anything involving payment changes — the AI never touches billing, payment methods, or charge disputes

This layer protects the 15% of tickets where judgment, empathy, or margin protection matters.

Why these rules matter

A chatbot that auto-approves refunds will lose you money. A chatbot that tries to handle angry customers will make them angrier. A chatbot that guesses on policy questions will give wrong answers.

The guardrails are not about limiting AI. They are about knowing where AI is great and where it is dangerous.

AI is great at:

• Response time (measure your actual baseline and post-launch results)
• Consistency (same answer every time)
• Availability (24/7)
• Volume (unlimited tickets at no extra cost)

AI is dangerous at:

• Judgment (should this customer get a refund?)
• Empathy (an angry customer needs to feel heard, not processed)
• Exceptions (your return policy says 30 days, but this customer has a valid reason for day 35)
• Margin protection (the AI does not know that this product costs you $40 to replace)

The 7-day test

The only way to know if the guardrails work for your store is to test them on your actual tickets. A 7-day free trial shows you:

• Which of your tickets AI resolves automatically
• Which ones AI drafts and a human approves
• Which ones AI correctly escalates
• Where the rules need adjustment for your store

Days 1-2: every AI draft gets human review. You see every response before it goes out.

Days 3-7: you adjust the rules based on what you see.

If the rules do not work for your store, you walk away. No cost, no contract.

Reply 'trial' and we set up the guardrails for your store this week.