Indian businesses deploying WhatsApp chatbots usually measure containment: the share of conversations the bot resolved without a human. Optimising containment is how you build a bot people hate.
The failure mode
A bot that cannot answer and will not escalate traps the customer in a loop. They try three phrasings, get the same menu, and either abandon or block. The block is the expensive outcome, because it removes them from every future message including order updates.
Escalation triggers worth encoding
- Two consecutive low-confidence intent matches
- Any message containing a complaint, refund or cancellation term
- The customer typing a variant of agent, human or representative, in English or Hinglish
- Any conversation longer than four bot turns without resolution
- Any message expressing frustration, detected on sentiment rather than keywords
The Hinglish problem
Intent classification on Indian inbound is weaker than vendor benchmarks suggest, because Hinglish has no standard spelling. Test on your own message history before trusting a containment number. A bot at high accuracy on formal English can sit far lower on real traffic.
What to tell the customer at handoff
State that a person is taking over and give a realistic wait. Silence during handoff reads as the bot failing again. If it is outside working hours, say so and commit to a time rather than leaving the conversation open.
The number to optimise
Resolution rate, not containment. A bot that escalates half its conversations and resolves the rest cleanly beats one that contains ninety percent and leaves people stuck.
General information, not legal advice. Rules current as of July 2026.