How to Train an AI Chatbot on Your Store Policies Without Losing Control

Start with the questions customers already ask

Do not begin by writing a giant policy prompt from scratch.

Review your real support history and make a list of recurring topics:

• shipping times
• shipping locations
• returns
• exchanges
• damaged items
• cancellations
• order changes
• discount rules
• product care
• sizing
• warranties
• order status

Start with the questions that are both common and safe to answer consistently.

Separate information from decisions

This is the most important step.

Information: “Our return window is 30 days.”

Decision: “This customer is 35 days late. Should we make an exception?”

The chatbot can explain the published rule. A person should usually decide the exception.

Apply the same distinction to refunds, replacement approvals, custom discounts, address changes, delivery claims, and complaints.

Write policies in plain language

If your policy is hard for a human to understand, it will also be hard for an AI assistant to apply consistently.

Use direct statements:

GOOD: “Unopened items may be returned within 30 days of delivery. Sale items are final unless they arrive damaged.”

WEAK: “Returns may be available in certain circumstances subject to our discretion and applicable terms.”

You can keep the legal version on your terms page while giving the chatbot a plain-language support summary that stays consistent with it.

Tell the chatbot what not to do

Negative instructions are important.

Examples:

• Do not approve refunds.
• Do not promise a replacement.
• Do not invent delivery dates.
• Do not claim an order was changed unless the system actually changed it.
• Do not make exceptions to the return policy.
• If policy information is missing, say you need human follow-up rather than guessing.

This prevents the assistant from trying to be helpful by making unauthorized commitments.

Add examples of correct boundaries

Examples teach the assistant what the rule means in practice.

Customer: “Can I return this after 20 days?” Expected behavior: explain the published return rule.

Customer: “I am outside the return window. Can you make an exception?” Expected behavior: explain that the team needs to review the request.

Customer: “Where is my order?” Expected behavior: use the approved secure order-status workflow if available.

Customer: “The parcel says delivered but it is missing.” Expected behavior: move toward human follow-up rather than inventing a replacement decision.

Keep store information current

A chatbot can be perfectly configured and still give outdated answers if the underlying policy changed.

Create a simple maintenance routine. Whenever your business changes shipping, returns, pricing, product rules, opening hours, or other important information, update the chatbot knowledge at the same time.

Review conversation history

Your customers will find edge cases you did not think of.

Review chatbot conversations for:

• repeated unanswered questions
• misleading answers
• policy ambiguity
• topics that should be escalated earlier
• product information that is missing
• outdated instructions

Use those conversations to improve both the chatbot and the website content.

How Farnsla fits this workflow

Farnsla lets merchants define tone, policies, FAQs, brand information, and instructions. Shopify and Wix ecommerce experiences can also combine these answers with product knowledge and secure order-status workflows.

When a person is needed, Farnsla can preserve conversation history and capture follow-up information. Shopify also supports Customer Messaging handoff through configured supported messaging apps.

The safest setup is simple:

Known rule → AI answers. Missing information → AI does not guess. Business decision → human decides.