A support bot that reduces workload without annoying customers
A bad support bot stands between the customer and a person. A good one answers faster than a person and steps aside in time.
Start from real conversations
Export the last 300–500 support messages and sort them by topic. Almost always 10–15 phrasings cover most requests: order status, hours, returns, warranty, address, pricing.
Automate that top of the list rather than trying to script every case. A bot that answers fifteen common questions and instantly hands over the rest beats one with a hundred branches.
Handover to a human
The rule: no more than two failed attempts. If the bot misunderstands twice, it passes the conversation to a person and says so plainly instead of asking to rephrase.
The operator must receive the full dialogue history, not just the last message, or the customer repeats everything and frustration doubles.
State working hours. Outside them the bot should say honestly when someone will reply and log the request.
Two languages done properly
Language is chosen at first launch and stored. Do not rely on the Telegram interface language — many users in Tashkent run it in English while writing Russian or Uzbek.
Uzbek answers are written separately, not machine-translated: stilted translation in support is obvious and damages trust more than replying in Russian would.
What to measure
Share of conversations closed without an operator, average first response time, and the list of questions the bot failed on. That last list drives monthly script updates.
Track bailouts separately: if more than a third tap 'talk to a human' on the first screen, the script does not match why people write to you.