Training an AI support agent on your past tickets: what to keep and what to throw away
Your ticket history is the best training data you have, and also full of outdated policies, bad days and one-off exceptions. Here's how to clean it up.
When teams first connect Onfre, the most common question is whether it should learn from every ticket they've ever answered. The honest answer: not all of them. Past tickets teach tone, phrasing and the real questions customers ask. They also contain every policy you've since changed and every reply sent at the end of a very long day.
Keep: the questions themselves
The way customers actually phrase things is gold. Nobody writes 'I would like to initiate a return'. They write 'this doesnt fit how to send back'. Learning from real questions is what lets an AI agent understand misspellings, Hinglish and half-sentences.
Keep: your best agents' replies
If you can, mark tickets from your strongest agents or ones with high satisfaction scores. Onfre weighs these more heavily when it learns your tone: how you apologise, how formal you are, whether you use emoji.
Throw away: anything older than your current policy
If your return window changed from 30 days to 15 days in March, every ticket before March teaches the wrong answer. Set a cut-off date, or let Onfre treat your written policy documents as the source of truth and use tickets only for tone and phrasing. That second option is what we recommend for most teams.
Throw away: one-off exceptions
- Goodwill refunds you gave a long-time customer.
- Workarounds for a bug that has since been fixed.
- Replies during an outage or a courier strike.
- Anything that starts with 'just this once'.
These are exactly the cases an AI agent should not learn to repeat. Onfre flags replies that contradict your written policies so you can exclude them in a couple of clicks.
Check before you go live
Before switching Onfre on, it drafts answers to your last 500 real tickets side by side with what your team actually sent. Skim 50 of them. You'll spot gaps in your documentation fast, and fixing a missing policy page is a ten-minute job that improves every future answer.
Your docs decide what Onfre says. Your tickets decide how it says it.