Why bother
Untrained, an AI writes safe corporate filler. Customers can tell, and a reply that reads like a template does more harm than no reply at all.
Trained on your business, it references what you actually sell, uses the words your team uses, and knows the things you'd never say.
What to give it
Go to /ai and add your business knowledge. The three that move the needle most:
- What you do services, specialities, what makes you different. Stops generic replies.
- Tone warm or brisk, first names or formal, emoji or never. Include phrases you'd never use.
- Common situations the complaints and compliments you get repeatedly, and how you like to answer each.
Approve a few replies by hand
The fastest training signal is your own edits. For the first week, read each drafted reply before it goes out and adjust the wording.
A handful of approved replies teaches tone far better than a long style document, because it shows the model real examples rather than describing them.
Then let autopilot take over
Once drafts consistently read like you, turn on autopilot so routine replies go out without you. Keep a human in the loop for anything negative that's where a wrong tone costs most.
Review the AI's output periodically even after it's dialled in. Your business changes new staff, new services, new complaints and the training should follow.
Good practice
- Never have the AI dispute facts in a negative review acknowledge, then take it offline.
- Keep replies short. Long replies read as defensive.
- Don't promise refunds or outcomes in an automated reply.
- Thank people by name where the platform gives you one.