The agent case studies circulating in Indian marketing WhatsApp groups are almost all American. They assume clean English data, email-first funnels and a well-maintained CRM. Three of those assumptions do not hold here, and each breaks agents in a specific way.
Language
Indian customer data is multilingual and often transliterated. A single support queue holds English, Hinglish typed in Latin script, Devanagari, Tamil and more. Agents handle formal Hindi reasonably and Hinglish poorly, because Hinglish has no standard spelling. Kitna, kitnaa and how much all appear as the same intent.
The practical consequence is that intent classification accuracy drops sharply on real Indian inbound compared to the benchmark numbers vendors quote. Test on your own message history before you trust anyone else figure.
WhatsApp-first funnels
Most agent tooling assumes email, where a thread is a clean unit with a subject line and clear boundaries. WhatsApp conversations run for months, mix service and sales, include voice notes and images, and have no boundaries at all.
Feeding that to an agent as a single context produces confident nonsense, because the agent cannot tell which part of a six-month thread is current. Teams that get this working segment conversations by time and topic before the agent sees them. That preprocessing is most of the work.
CRM hygiene
An agent making decisions from your CRM inherits every problem in it. Duplicate contacts, stale opportunity stages, free-text fields holding six different conventions. Automation tolerates this because rules only read the fields you pointed them at. An agent reads everything and reasons from all of it, including the wrong parts.
The fix is unglamorous. Deduplicate, standardise the fields the agent will read, and delete what you do not trust. Teams that skip this and blame the model are usually looking at a data problem.
What to do before a pilot
Run three checks. Take 200 real inbound messages and measure intent classification yourself. Take twenty of your longest WhatsApp threads and see whether the agent can identify the current request. Pull 100 CRM records and count how many have a field an agent would misread. If any check comes back poor, fix that before buying anything.