The usual way to build an FAQ bot is for someone to sit down and imagine what customers ask. The better way is to read what they actually asked, because you already have several thousand examples and the answers your team gave.
Start with the export
Pull the last three to six months of conversations. Not a sample chosen by anyone — everything, so the frequency data is real. You are looking for volume, not for interesting cases.
Cluster by hand first
Read 200 of them and write down the intent of each in a few words. It is tedious and it is the most valuable hour in the project. Patterns become obvious quickly, and you will find categories nobody would have guessed — often something specific to your logistics or your checkout that generates a surprising share of contacts.
Then count. Usually a small number of intents cover most of the volume, and the tail is long and thin. Automate the head and leave the tail to people.
Sort by volume times simplicity
The best automation candidates are frequent AND have one stable correct answer. A common question with a different answer per customer belongs in a lookup or a human queue, not in an FAQ bot.
Take the answers from your best agents
For each intent, find several conversations your team resolved cleanly and read how they phrased it. Use that language. Answers written by a marketing team read as marketing; answers taken from support read as support, and customers can tell.
Where agents disagree on an answer, you have found a policy gap. Resolve it before automating it — otherwise you will encode one agent's improvisation as company policy.
Write the answers properly
- Answer in the first line. Detail afterwards, if at all.
- Be specific: "within 7 days of delivery", not "within our returns window".
- Include the next action — a link, a reply keyword, a form.
- End with a route to a person for the cases the answer does not cover.
Launch narrow
Automate the three highest-volume intents and nothing else. Watch for two weeks. You will find phrasings you missed and edge cases the answer handles badly, and fixing three intents properly is far easier than fixing fifteen at once.
Measure whether it actually helped
- Containment — conversations resolved without a human. The headline number, and the most flattering one.
- Escalation after an answer — the honest counterweight. If people read the automated reply and immediately ask for a person, the answer is not working.
- Repeat contacts — the same customer returning with the same question within days means it was deflected, not resolved.
- Handling time on escalated cases — good automation should shorten these by arriving with context.
Containment alone can be gamed by making it hard to reach a person. Watching it alongside escalation and repeat contacts keeps you honest about whether customers were helped or merely deflected.
Keep re-reading the tickets
Repeat the export quarterly. Products change, policies change, and the question mix changes with them. An FAQ bot built once and left alone becomes wrong slowly and invisibly, which is the worst way for it to fail.