Stores lose trust in AI the first time it offers 20% off or sounds like a call centre. Brand voice and guardrails are how you keep the speed of AI without giving away the shop.
Brand voice
In AI Training you set tone (friendly, professional, apologetic), a human sign-off, and a short sample of how you already write. Taskra uses that on email and chat so a midnight WISMO reply still sounds like you.
Walk through the settings in setting your brand voice.
Guardrails
- Never invent discount codes or shipping promises
- Never name competitors or recommend another store
- Never give medical, legal or investment advice
- Always escalate refunds and angry threads
Write the list down. Models follow explicit bans better than vibes. See setting guardrails.
Sources still come first
Voice without catalogue data is a well-spoken guess. Keep Shopify sync on and add FAQs for anything that is not on a store page. Voice shapes the sentence; sources decide the facts.
Frequently asked questions
What is a guardrail in Taskra?
A hard rule: competitor names, unauthorised promises, or phrases the model must never use, even if a shopper asks.
How do I make replies sound like my store?
Pick a tone, add a sign-off, and paste two or three real replies you have already sent. That sample beats a long style guide.